pylon_core/sql/

mod.rs

//
// This source file is part of the Pylon open source project.
//
// Copyright (c) 2026 Jaldis B.V.
//
// Licensed under the MIT OR Apache-2.0 license (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//     https://opensource.org/licenses/MIT
//     https://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
//

use crate::ir::{
    IrArraySource, IrCteDef, IrDelete, IrExpr, IrFor, IrForIterator, IrFreeExpr, IrFtsSearch, IrFunctionSelect,
    IrGlobalCte, IrGroup, IrGroupOutput, IrInsert, IrLiteral, IrLockClause, IrLockStrength, IrLockWait,
    IrMultiLinkJoin, IrMultiLinkMutation, IrMultiLinkPointer, IrMultiLinkValueSource, IrMultiLinkValues, IrNulls,
    IrOutput, IrPathJoin, IrPathResult, IrPathSelect, IrPolyFanout, IrPolyImplementor, IrRewrite, IrRowSource,
    IrScalarPointer, IrScalarSetPointer, IrSelect, IrShapePointer, IrSingleLinkCorrelation, IrSingleLinkPointer,
    IrSort, IrSortDir, IrSource, IrStmt, IrUpdate, IrVectorSearch, SearchEnqueueInfo, VectorEnqueueInfo,
};
use crate::parse::ast::{BinOpKind, UnaryOpKind};
use crate::query::{Cardinality, InferencePlan, ShapeDescriptor, ShapeNode};
use std::collections::HashMap;

pub struct SqlOutput {
    pub sql: String,
    pub shape: ShapeDescriptor,
    pub inference_plan: Option<InferencePlan>,
}

/// Build an UPDATE's `SET` clause fragments (`"col" = expr`, no trailing
/// comma) from its explicit assignments plus its rewrites, with a rewrite on
/// a given column overriding — not appending alongside — any explicit
/// assignment to that same column. Mirrors the INSERT-side `rewrite_cols`
/// dedup (`ins.assignments` filtered against `ins.rewrites`'s own columns,
/// present at every INSERT emission site) — every UPDATE emission site
/// independently lacked the equivalent filtering, so an UPDATE that both
/// explicitly assigned and had a rewrite declared on the same property
/// produced two `SET` entries for one column, which Postgres rejects with
/// "multiple assignments to same column" (confirmed live).
fn update_set_fragments(assignments: &[(String, IrExpr)], rewrites: &[IrRewrite], indent: &str) -> Vec<String> {
    let rewrite_cols: std::collections::HashSet<&str> = rewrites.iter().map(|r| r.column.as_str()).collect();
    assignments
        .iter()
        .filter(|(c, _)| !rewrite_cols.contains(c.as_str()))
        .map(|(col, expr)| format!("{indent}{} = {}", qi(col), emit_expr(expr)))
        .chain(
            rewrites
                .iter()
                .map(|rw| format!("{indent}{} = {}", qi(&rw.column), emit_expr(&rw.expr))),
        )
        .collect()
}

/// Emit the SQL body for a computed global CTE — a plain scalar query with a `value` column.
fn emit_for_global_cte(stmt: &IrStmt) -> String {
    match stmt {
        IrStmt::Select(sel) => match sel.rows.as_slice() {
            [IrRowSource::Bound { source, .. }] => {
                let alias = &source.alias;
                let mut sql = format!(
                    "SELECT {}.\"id\" AS \"value\"\nFROM {} AS {}",
                    qi(alias),
                    source_ref(source),
                    qi(alias)
                );
                append_filter(&mut sql, &sel.filter);
                sql
            }
            rows => match rows.first() {
                Some(IrRowSource::Free(IrFreeExpr::Scalar(e))) => format!("SELECT {} AS \"value\"", emit_expr(e)),
                _ => "SELECT NULL AS \"value\"".to_string(),
            },
        },
        IrStmt::PathSelect(sel) => {
            let from_sql = emit_path_joins(&sel.root, &sel.joins);
            let scalar_expr = match &sel.result {
                IrPathResult::Scalar(e, _) => emit_expr(e),
                IrPathResult::Object { alias, .. } => format!("{}.\"id\"", qi(alias)),
            };
            let mut sql = format!("SELECT {} AS \"value\"\nFROM {}", scalar_expr, from_sql);
            append_filter(&mut sql, &sel.filter);
            append_order_by(&mut sql, &sel.order_by);
            append_offset_limit(&mut sql, &sel.offset, &sel.limit);
            sql
        }
        _ => "SELECT NULL AS \"value\"".to_string(),
    }
}

fn emit_global_cte_parts(global_ctes: &[IrGlobalCte]) -> Vec<String> {
    global_ctes
        .iter()
        .map(|g| match g {
            IrGlobalCte::Session(s) => format!(
                "\"{}\" AS (SELECT ${}::{} AS \"value\")",
                s.cte_name,
                s.param_index + 1,
                s.pg_type
            ),
            IrGlobalCte::Computed(c) => {
                let body = emit_for_global_cte(&c.stmt);
                format!("\"{}\" AS (\n{}\n)", c.cte_name, body)
            }
        })
        .collect()
}

pub fn emit(ir: &IrOutput) -> SqlOutput {
    let out = with_subtype_fanouts(&ir.subtype_fanouts, || {
        with_correlated_ctes(&ir.ctes, || emit_output(ir))
    });
    #[cfg(debug_assertions)]
    if let Some(problem) = forward_cte_reference(&out.sql) {
        panic!("{problem}\n{}", out.sql);
    }
    out
}

/// A WITH name is in scope only for the bindings that follow it, so a CTE
/// that reads one defined after it fails at execution time — and does so
/// complaining about a FROM clause, far from what actually put the two in
/// the wrong order. Checked here rather than left to PostgreSQL because the
/// parts are assembled in a dozen places, none of which owns the ordering.
#[cfg(debug_assertions)]
fn forward_cte_reference(sql: &str) -> Option<String> {
    let definitions: Vec<(String, usize)> = sql
        .match_indices("\" AS (")
        .filter_map(|(end, _)| {
            let head = &sql[..end];
            let start = head.rfind('"')?;
            Some((head[start + 1..].to_string(), start))
        })
        .collect();
    for (name, defined_at) in &definitions {
        // A bare quoted name is a CTE; a real relation is always schema
        // qualified (`"marketplace"."Order"`), so these never collide.
        let quoted = qi(name);
        for prefix in ["FROM ", "JOIN ", "CROSS JOIN "] {
            let needle = format!("{prefix}{quoted}");
            if let Some(used_at) = sql.find(&needle)
                && used_at < *defined_at
            {
                return Some(format!(
                    "CTE {quoted} is read at byte {used_at} but not defined until {defined_at}"
                ));
            }
        }
    }
    None
}

fn emit_output(ir: &IrOutput) -> SqlOutput {
    let mut out = match &ir.stmt {
        IrStmt::Update(upd) => emit_update_stmt(upd, &ir.ctes),
        IrStmt::For(f) => emit_for_stmt(f, &ir.ctes),
        stmt => {
            let mut o = match stmt {
                IrStmt::Select(sel) => emit_select_stmt(sel, &ir.ctes),
                IrStmt::PathSelect(sel) => emit_path_select(sel),
                IrStmt::Insert(ins) => emit_insert_stmt(ins),
                IrStmt::Delete(del) => emit_delete_stmt(del),
                IrStmt::Group(grp) => emit_group(grp),
                IrStmt::FunctionSelect(sel) => emit_function_select(sel),
                IrStmt::VectorSearch(vs) => emit_vector_search(vs),
                IrStmt::FtsSearch(fs) => emit_fts_search(fs),
                IrStmt::ScalarUnion(branches) => SqlOutput {
                    sql: emit_scalar_union(branches),
                    shape: ShapeDescriptor {
                        root: ShapeNode::Scalar {
                            name: String::new(),
                            position: 0,
                        },
                    },
                    inference_plan: None,
                },
                IrStmt::Update(_) | IrStmt::For(_) => unreachable!(),
            };
            if !ir.ctes.is_empty() {
                // The statement emitter may already have opened a `WITH` of
                // its own (multi-link/junction assignment, FOR-loop insert);
                // a second one is invalid SQL, so merge into it.
                o.sql = merge_into_existing_with(&o.sql, &emit_user_cte_parts(&ir.ctes))
                    .unwrap_or_else(|| format!("{}{}", emit_cte_prefix(&ir.ctes), o.sql));
            }
            o
        }
    };

    if !ir.global_ctes.is_empty() {
        let global_parts = emit_global_cte_parts(&ir.global_ctes);
        out.sql = merge_into_existing_with(&out.sql, &global_parts)
            .unwrap_or_else(|| format!("WITH {}\n{}", global_parts.join(",\n     "), out.sql));
    }

    out
}

/// Merge CTE definitions into a statement that already opens with a top-level
/// `WITH`, or `None` when it doesn't (the caller then builds a fresh prefix in
/// its own format).
///
/// SQL allows exactly one `WITH` per statement, but CTE parts arrive from
/// several independent places: a user-written `with` block (`ir.ctes`),
/// globals (`ir.global_ctes`), and the CTEs some statement emitters build for
/// themselves — a multi-link or junction-backed link assignment needs
/// `_w__ids`/`_w__ml_add_*`, and the `FOR`-loop insert path needs its VALUES
/// alias. Whichever runs second has to merge; blindly prefixing produces
/// `WITH ... WITH ...`, which PostgreSQL rejects with `syntax error at or
/// near "WITH"` — and only at execution time, since the PyQL compiles fine.
///
/// The merged-in parts go *first*: a non-recursive CTE may only reference
/// siblings declared before it, and the parts already present are the ones
/// that reference what's being added (a statement's own CTEs consume the user
/// binding they were built from).
///
/// Two prefix spellings occur — `emit_cte_prefix`'s `"WITH\n"` and the inline
/// `"WITH "` built by individual emitters. Both are 5 bytes before the parts
/// begin, so one slice covers either.
/// The array operand for `= ANY(...)` / `<> ALL(...)`.
///
/// `x in std::array_unpack(arr)` is the documented way to test membership of
/// an array parameter, but `array_unpack` emits `unnest()`, and nesting that
/// inside `ANY` gives `ANY(unnest(arr))` — a set-returning function in
/// `WHERE`, which PostgreSQL rejects (`set-returning functions are not
/// allowed in WHERE`). `ANY` already takes an array, so the `unnest` is both
/// redundant and invalid: pass its argument straight through.
///
/// Falls back to the pre-rendered form for anything else, including
/// `unnest()` reached by some route other than a direct `array_unpack` call.
fn unwrap_unnest_for_any(right: &IrExpr, rendered: &str) -> String {
    if let Some(array) = unnest_argument(right) {
        return array;
    }
    // `<Enum>array_unpack($1)` — the cast is written over the set, so over the
    // array the same cast is to that type's array.
    if let IrExpr::TypeCast(cast) = right
        && let Some(array) = unnest_argument(&cast.expr)
    {
        return format!("({})::{}[]", array, cast.pg_type);
    }
    rendered.to_string()
}

/// The array a direct `array_unpack` call unpacks, already rendered.
fn unnest_argument(expr: &IrExpr) -> Option<String> {
    let IrExpr::FunctionCall(call) = expr else {
        return None;
    };
    (call.schema.is_none() && call.name == "unnest" && call.sql_template.is_none() && call.args.len() == 1)
        .then(|| emit_expr(&call.args[0]))
}

fn merge_into_existing_with(sql: &str, parts: &[String]) -> Option<String> {
    if parts.is_empty() || !(sql.starts_with("WITH\n") || sql.starts_with("WITH ")) {
        return None;
    }
    Some(format!("WITH {},\n     {}", parts.join(",\n     "), &sql[5..]))
}

// ── Identifier / literal helpers ────────────────────────────────────────────

fn qi(s: &str) -> String {
    format!("\"{}\"", s.replace('"', "\"\""))
}

fn pg_schema(module: &str) -> String {
    if module == "default" {
        "\"public\"".into()
    } else {
        qi(module)
    }
}

pub fn pg_schema_str(module: &str) -> String {
    pg_schema(module)
}

fn qn(module: &str, name: &str) -> String {
    format!("{}.{}", pg_schema(module), qi(name))
}

fn sql_str(s: &str) -> String {
    format!("'{}'", s.replace('\'', "''"))
}

/// `'module::Type'::text` — always position 0 in every non-free-type tuple.
fn type_disc(type_name: &str) -> String {
    format!("{}::text", sql_str(type_name))
}

/// Extract the PostgreSQL schema name (module) from `"module::TypeName"`.
fn module_of(type_name: &str) -> &str {
    type_name.split("::").next().unwrap_or("public")
}

fn source_ref(src: &IrSource) -> String {
    // "@cte:name" sentinel: the source is a WITH-clause CTE, not a real table.
    if let Some(cte_name) = src.table.strip_prefix("@cte:") {
        // A binding that reads a loop variable holds one value per iteration,
        // all of them in the one CTE. A read of it belongs to the iteration
        // in scope here, which is the key it carries.
        return match correlated_cte_iterator(cte_name) {
            Some(iterator) => format!(
                "(SELECT * FROM {} WHERE {}.{} = {}.\"v\")",
                qi(cte_name),
                qi(cte_name),
                qi(OUTER_KEY),
                qi(&iterator),
            ),
            None => qi(cte_name),
        };
    }
    // "@row:NEW" sentinel: a trigger's own row.
    if let Some(row) = src.table.strip_prefix("@row:") {
        return format!("(SELECT ({row}).*)");
    }
    // An interface-typed source reads from its implementors, never from the
    // interface's own view — see `IrSource::poly`.
    match source_fanout(src) {
        Some(fanout) => format!("(\n{}\n)", emit_poly_union(&fanout.implementors, &fanout.columns)),
        None => junction_ref(module_of(&src.type_name), &src.table),
    }
}

/// The table a statement writes to: only ever the one named, never the
/// subtypes a read of it would take in.
fn target_ref(src: &IrSource) -> String {
    match src.table.strip_prefix("@cte:") {
        Some(cte_name) => qi(cte_name),
        None => qn(module_of(&src.type_name), &src.table),
    }
}

thread_local! {
    /// The statement being emitted's `IrOutput::subtype_fanouts`.
    static SUBTYPE_FANOUTS: std::cell::RefCell<HashMap<(String, String), IrPolyFanout>> =
        std::cell::RefCell::new(HashMap::new());

    /// Every CTE of the statement being emitted that holds one value per
    /// iteration, and the iterator whose key says which.
    static CORRELATED_CTES: std::cell::RefCell<HashMap<String, String>> =
        std::cell::RefCell::new(HashMap::new());
}

/// The iterator a read of `cte_name` has to pin to one iteration of, if that
/// binding is per-iteration at all.
fn correlated_cte_iterator(cte_name: &str) -> Option<String> {
    CORRELATED_CTES.with(|cell| cell.borrow().get(cte_name).cloned())
}

/// Installs the per-iteration bindings for the emission running inside
/// `emit`, restoring the previous set after.
fn with_correlated_ctes<T>(ctes: &[IrCteDef], emit: impl FnOnce() -> T) -> T {
    fn collect(ctes: &[IrCteDef], into: &mut HashMap<String, String>) {
        for c in ctes {
            if let Some(iterator) = &c.correlated_to {
                into.insert(c.name.clone(), iterator.clone());
            }
            match &c.stmt {
                IrStmt::For(f) => collect(&f.body_ctes, into),
                IrStmt::Insert(ins) => collect(&ins.nested_ctes, into),
                IrStmt::Update(upd) => collect(&upd.nested_ctes, into),
                _ => {}
            }
        }
    }
    let mut collected = HashMap::new();
    collect(ctes, &mut collected);
    let previous = CORRELATED_CTES.with(|cell| cell.replace(collected));
    let result = emit();
    CORRELATED_CTES.with(|cell| cell.replace(previous));
    result
}

/// Installs `fanouts` for the emission running inside `emit`, restoring the
/// previous set after, so a nested emission of another output cannot leak.
fn with_subtype_fanouts<T>(fanouts: &HashMap<(String, String), IrPolyFanout>, emit: impl FnOnce() -> T) -> T {
    let previous = SUBTYPE_FANOUTS.with(|cell| cell.replace(fanouts.clone()));
    let result = emit();
    SUBTYPE_FANOUTS.with(|cell| cell.replace(previous));
    result
}

/// The fan-out a read of `src` goes through: its own, or the one a concrete
/// type with subtypes needs.
fn source_fanout(src: &IrSource) -> Option<IrPolyFanout> {
    if let Some(fanout) = &src.poly {
        return Some(fanout.clone());
    }
    if src.table.starts_with('@') {
        return None;
    }
    let key = (module_of(&src.type_name).to_string(), src.table.clone());
    SUBTYPE_FANOUTS.with(|cell| cell.borrow().get(&key).cloned())
}

/// A junction table, or the CTE standing in for one. Same `@cte:` sentinel
/// `source_ref` honours: a multi-link populated by the statement's own nested
/// DML reads the junction rows back out of the CTE that wrote them, since
/// Postgres does not show them in the table yet.
/// The name the emitter gives the CTE wrapping a `SELECT (INSERT/UPDATE …)`'s
/// own mutation, and the stem of the junction CTEs beside it
/// (`_dml__ml_add_0`). Public because the compiler has to name those same
/// CTEs when it points a shape at them — see
/// `read_nested_links_from_their_ctes`.
pub const DML_CTE: &str = "_dml";

fn junction_ref(module: &str, junction_table: &str) -> String {
    if let Some(cte_name) = junction_table.strip_prefix("@cte:") {
        return qi(cte_name);
    }
    let Some((tables, columns)) = crate::ir::parse_inherited_junction(junction_table) else {
        return qn(module, junction_table);
    };
    let columns = columns.iter().map(|c| qi(c)).collect::<Vec<_>>().join(", ");
    let branches = tables
        .iter()
        .map(|(module, table)| format!("SELECT {columns} FROM {}", qn(module, table)))
        .collect::<Vec<_>>()
        .join(" UNION ALL ");
    format!("({branches})")
}

/// The discriminator for a path select's object result.
///
/// The terminal row source is the one the result came from, so that is the one
/// whose fan-out decides whether the type is read off the row or fixed.
fn result_type_disc(path: &IrPathSelect, alias: &str, type_name: &str) -> String {
    match terminal_source(path, alias) {
        Some(source) => source_type_disc(source),
        None => type_disc(type_name),
    }
}

fn terminal_source<'a>(path: &'a IrPathSelect, alias: &str) -> Option<&'a IrSource> {
    if path.root.alias == alias {
        return Some(&path.root);
    }
    path.joins
        .iter()
        .map(path_join_target)
        .find(|target| target.alias == alias)
}

fn path_join_target(join: &IrPathJoin) -> &IrSource {
    match join {
        IrPathJoin::Single { target, .. }
        | IrPathJoin::Multi { target, .. }
        | IrPathJoin::BacklinkSingle { target, .. }
        | IrPathJoin::BacklinkMulti { target, .. }
        | IrPathJoin::Function { target, .. }
        | IrPathJoin::Lateral { target, .. } => target,
    }
}

/// The type discriminator for a row from `src`: the fanned-out column when the
/// source reads from implementors, else the one type its table holds.
fn source_type_disc(src: &IrSource) -> String {
    match source_fanout(src) {
        Some(_) => format!("{}.\"__type__\"", qi(&src.alias)),
        None => type_disc(&src.type_name),
    }
}

// ── Polymorphic UNION ALL ───────────────────────────────────────────────────

fn emit_poly_union(implementors: &[IrPolyImplementor], columns: &[String]) -> String {
    let col_list = columns.iter().map(|c| qi(c)).collect::<Vec<_>>().join(", ");
    implementors
        .iter()
        .map(|imp| {
            format!(
                "    SELECT {}::text AS \"__type__\", {} FROM {}",
                sql_str(&imp.type_name),
                col_list,
                qn(&imp.module, &imp.table),
            )
        })
        .collect::<Vec<_>>()
        .join("\n    UNION ALL\n")
}

// ── SELECT statement ────────────────────────────────────────────────────────

fn emit_select_stmt(sel: &IrSelect, ctes: &[IrCteDef]) -> SqlOutput {
    match sel.rows.as_slice() {
        [IrRowSource::Bound { source, shape }] => emit_bound_select(sel, source, shape),
        rows if rows.len() > 1 && rows.iter().all(|r| matches!(r, IrRowSource::Bound { .. })) => {
            emit_bound_union_select(sel, rows)
        }
        rows => emit_free_rows(sel, rows, ctes),
    }
}

/// The branches of `select (a union b)` all carry the same object type, so
/// they share a column set: union them in the FROM clause and project, filter
/// and order the combined set once, under the shape's single alias.
/// `common_columns`, when set, are the columns of the ancestor branches of
/// different types share: each branch projects those and its rows' concrete
/// type, rather than columns the others do not have.
fn bound_union_from_clause(rows: &[IrRowSource], common_columns: &[String]) -> String {
    let columns = common_columns.iter().map(|c| qi(c)).collect::<Vec<_>>().join(", ");
    rows.iter()
        .map(|row| match row {
            IrRowSource::Bound { source, .. } if !common_columns.is_empty() => {
                let own_type = match source_fanout(source) {
                    Some(_) => "\"__type__\"".to_string(),
                    None => format!("{} AS \"__type__\"", type_disc(&source.type_name)),
                };
                format!("    SELECT {own_type}, {columns} FROM {}", source_ref(source))
            }
            IrRowSource::Bound { source, .. } => format!("    SELECT * FROM {}", source_ref(source)),
            IrRowSource::Free(_) => unreachable!("caller checked every row is bound"),
        })
        .collect::<Vec<_>>()
        .join("\n    UNION ALL\n")
}

fn emit_bound_union_select(sel: &IrSelect, rows: &[IrRowSource]) -> SqlOutput {
    let IrRowSource::Bound { source, shape } = &rows[0] else {
        unreachable!("caller checked every row is bound")
    };
    let alias = &source.alias;
    let (pointer_exprs, shape_pointers) = build_shape(shape, alias);
    let row_type = if sel.poly_columns.is_empty() {
        type_disc(&source.type_name)
    } else {
        format!("{}.\"__type__\"", qi(alias))
    };
    let mut parts = vec![row_type];
    parts.extend(pointer_exprs);

    let mut sql = format!(
        "SELECT {}(\n    {}\n) AS result\nFROM (\n{}\n) AS {}",
        if sel.distinct { "DISTINCT " } else { "" },
        parts.join(",\n    "),
        bound_union_from_clause(rows, &sel.poly_columns),
        qi(alias),
    );
    append_filter(&mut sql, &sel.filter);
    append_order_by(&mut sql, &sel.order_by);
    append_offset_limit(&mut sql, &sel.offset, &sel.limit);

    SqlOutput {
        sql,
        shape: ShapeDescriptor {
            root: ShapeNode::Object {
                name: String::new(),
                type_name: Some(source.type_name.clone()),
                position: 0,
                cardinality: Cardinality::Many,
                pointers: prepend_type(shape_pointers),
                has_implicit_id: shape_has_implicit_id(shape),
            },
        },
        inference_plan: None,
    }
}

/// Schema-bound SELECT — has a FROM clause, projected column shape, and
/// (optionally) a DML source / polymorphic fan-out. `source`/`shape` are the
/// single `IrRowSource::Bound` entry destructured by `emit_select_stmt`;
/// everything else (filter/order_by/offset/limit/distinct/dml_source/
/// polymorphic/poly_*) stays on the outer `IrSelect`.
fn emit_bound_select(sel: &IrSelect, source: &IrSource, shape: &[IrShapePointer]) -> SqlOutput {
    let alias = &source.alias;
    let (pointer_exprs, shape_pointers) = build_shape(shape, alias);

    let type_expr = if sel.polymorphic {
        format!("{}.\"__type__\"", qi(alias))
    } else {
        type_disc(&source.type_name)
    };
    let mut parts = vec![type_expr];
    parts.extend(pointer_exprs);
    let tuple = parts.join(",\n    ");

    let distinct = if sel.distinct { "DISTINCT " } else { "" };

    // SELECT-over-DML: wrap inner statement in a CTE, select from it.
    let from_clause = if let Some(dml) = &sel.dml_source {
        let mut cte_parts = match dml.as_ref() {
            IrStmt::Update(upd) if update_has_any_multilink(upd) => emit_update_multilink_ctes(upd, DML_CTE),
            IrStmt::Insert(ins) if insert_has_any_multilink(ins) => emit_insert_multilink_ctes(ins, DML_CTE),
            // Nested DML the row's values read from: hoisted ahead of the row
            // itself, since a data-modifying WITH only works at the top level.
            IrStmt::Insert(ins) if !ins.nested_ctes.is_empty() => {
                let mut parts = emit_user_cte_parts(&ins.nested_ctes);
                parts.push(format!("\"_dml\" AS (\n{}\n)", emit_dml_as_cte_source(dml)));
                parts
            }
            IrStmt::Update(upd) if !upd.nested_ctes.is_empty() => {
                let mut parts = emit_user_cte_parts(&upd.nested_ctes);
                parts.push(format!("\"_dml\" AS (\n{}\n)", emit_dml_as_cte_source(dml)));
                parts
            }
            IrStmt::Update(upd) if !upd.poly_implementors.is_empty() => emit_poly_update_dml_ctes(upd, "_dml"),
            IrStmt::Delete(del) if !del.poly_implementors.is_empty() => emit_poly_delete_dml_ctes(del, "_dml"),
            _ => vec![format!("\"_dml\" AS (\n{}\n)", emit_dml_as_cte_source(dml))],
        };
        let (enqueue_v, enqueue_s) = match dml.as_ref() {
            IrStmt::Insert(ins) => (ins.enqueue_vector.as_slice(), ins.enqueue_search.as_slice()),
            IrStmt::Update(upd) => (upd.enqueue_vector.as_slice(), upd.enqueue_search.as_slice()),
            _ => (&[][..], &[][..]),
        };
        cte_parts.extend(enqueue_ctes(enqueue_v, "_dml"));
        cte_parts.extend(enqueue_search_ctes(enqueue_s, "_dml", enqueue_v.len()));
        format!(
            "WITH\n{}\nSELECT {}(\n    {}\n) AS result\nFROM \"_dml\" AS {}",
            cte_parts.join(",\n"),
            distinct,
            tuple,
            qi(alias)
        )
    } else if sel.polymorphic && !source.table.starts_with("@cte:") {
        // Polymorphic interface with no CTE indirection: fan out to implementor tables.
        let union_sql = emit_poly_union(&sel.poly_implementors, &sel.poly_columns);
        format!(
            "SELECT {}(\n    {}\n) AS result\nFROM (\n{}\n) AS {}",
            distinct,
            tuple,
            union_sql,
            qi(alias)
        )
    } else {
        // Concrete table or CTE (pre-filtered): query directly.
        // For CTE-backed polymorphic sources, __type__ is already present in the CTE result.
        format!(
            "SELECT {}(\n    {}\n) AS result\nFROM {} AS {}",
            distinct,
            tuple,
            source_ref(source),
            qi(alias)
        )
    };

    let mut sql = from_clause;
    append_filter(&mut sql, &sel.filter);
    append_order_by(&mut sql, &sel.order_by);
    append_offset_limit(&mut sql, &sel.offset, &sel.limit);
    append_lock_clause(&mut sql, &sel.lock);

    let root_pointers = prepend_type(shape_pointers);
    SqlOutput {
        sql,
        shape: ShapeDescriptor {
            root: ShapeNode::Object {
                name: String::new(),
                type_name: Some(source.type_name.clone()),
                position: 0,
                cardinality: Cardinality::Many,
                pointers: root_pointers,
                has_implicit_id: shape_has_implicit_id(shape),
            },
        },
        inference_plan: None,
    }
}

/// Emit a DML statement for use as a CTE source, using `RETURNING *` to expose
/// all columns to the outer SELECT.  The DML's own returning shape is ignored.
/// `FROM` over the nested-DML CTEs `reading` takes a column of, prefixed by
/// `lead`; empty when it takes none. One read through a scalar subquery
/// (`(SELECT "id" FROM cte)`) needs no join, and joining it anyway drops every
/// row the moment that CTE is empty -- which a guarded branch is.
fn nested_cte_from(ctes: &[IrCteDef], reading: &str, lead: &str) -> String {
    let joined: Vec<String> = ctes
        .iter()
        .map(|c| qi(&c.name))
        .filter(|name| reading.contains(&format!("{name}.")))
        .collect();
    if joined.is_empty() {
        return String::new();
    }
    format!("{lead}FROM {}", joined.join(", "))
}

/// Each operand's rows read through the bare `v` every scalar statement
/// exposes as a CTE source, so operands of different kinds line up.
fn emit_scalar_union(branches: &[IrStmt]) -> String {
    let operands: Vec<String> = branches
        .iter()
        .enumerate()
        .map(|(i, branch)| {
            format!(
                "SELECT \"v\" FROM (\n{}\n) AS \"_u{i}\"",
                emit_dml_as_cte_source(branch)
            )
        })
        .collect();
    format!(
        "SELECT ROW(v) AS result, v FROM (\n{}\n) AS _scalar",
        operands.join("\nUNION ALL\n")
    )
}

fn emit_dml_as_cte_source(stmt: &IrStmt) -> String {
    match stmt {
        IrStmt::ScalarUnion(branches) => format!("    {}", emit_scalar_union(branches)),
        IrStmt::Insert(ins) => {
            let rewrite_cols: std::collections::HashSet<&str> =
                ins.rewrites.iter().map(|r| r.column.as_str()).collect();
            let cols: Vec<String> = ins
                .assignments
                .iter()
                .filter(|(c, _)| !rewrite_cols.contains(c.as_str()))
                .map(|(c, _)| format!("    {}", qi(c)))
                .chain(ins.rewrites.iter().map(|r| format!("    {}", qi(&r.column))))
                .collect();
            let vals: Vec<String> = ins
                .assignments
                .iter()
                .filter(|(c, _)| !rewrite_cols.contains(c.as_str()))
                .map(|(_, e)| format!("    {}", emit_expr(e)))
                .chain(ins.rewrites.iter().map(|r| format!("    {}", emit_expr(&r.expr))))
                .collect();
            let mut sql = if !ins.nested_ctes.is_empty() || ins.guard.is_some() {
                // The row reads columns out of hoisted nested-DML CTEs, so it
                // selects from them rather than listing literal values. Those
                // CTEs are emitted by the caller: Postgres only allows a
                // data-modifying WITH at the top level, never inside a CTE.
                // A guarded insert likewise has to be a `SELECT … WHERE`, so
                // that it writes nothing when the condition is false.
                format!("    {}", emit_insert_row_sql(ins))
            } else if cols.is_empty() {
                // See `emit_insert_row_sql`: a row of nothing but defaults.
                format!("    INSERT INTO {} DEFAULT VALUES", target_ref(&ins.target))
            } else {
                format!(
                    "    INSERT INTO {} (\n{}\n    ) VALUES (\n{}\n    )",
                    target_ref(&ins.target),
                    cols.join(",\n"),
                    vals.join(",\n"),
                )
            };
            if let Some(conflict) = &ins.unless_conflict {
                emit_conflict(&mut sql, conflict);
            }
            sql.push_str("\n    RETURNING *");
            sql
        }
        IrStmt::Update(upd) => {
            // Callers must route an update with any multi-link (junction
            // table) mutation through emit_update_multilink_ctes instead —
            // junction INSERT/DELETE CTEs are data-modifying and Postgres
            // requires those to sit at the *top level* of the query, so they
            // can't be nested inside this function's single-CTE-body return
            // value. See emit_cte_prefix and emit_select_stmt's dml_source
            // branch, both of which check has_any_multilink before calling
            // this function at all.
            let alias = &upd.target.alias;
            let sets = update_set_fragments(&upd.assignments, &upd.rewrites, "    ");
            let mut sql = format!(
                "    UPDATE {} AS {}\n    SET\n{}",
                target_ref(&upd.target),
                qi(alias),
                sets.join(",\n"),
            );
            // The nested CTEs are emitted alongside this body (see
            // `emit_user_cte_parts`), so the reference to them needs a FROM.
            sql.push_str(&nested_cte_from(&upd.nested_ctes, &sets.join(","), "\n    "));
            append_filter(&mut sql, &upd.filter);
            // Qualified: past a FROM, a bare `*` carries the joined relation's
            // columns too, and a reader of this CTE then finds two `id`s.
            sql.push_str(&format!("\n    RETURNING {}.*", qi(alias)));
            sql
        }
        IrStmt::Delete(del) => {
            let alias = &del.target.alias;
            let mut sql = format!("    DELETE FROM {} AS {}", target_ref(&del.target), qi(alias),);
            append_filter(&mut sql, &del.filter);
            sql.push_str("\n    RETURNING *");
            sql
        }
        IrStmt::Select(inner) => match inner.rows.as_slice() {
            // `x := (select (select T filter …) filter …)` — the inner select is
            // what this one reads, the same `_dml` the top-level emitter reads
            // it from. Read from the table instead, it returned every row of
            // the type, with neither filter applied.
            [IrRowSource::Bound { source, .. }]
                if let Some(read) = inner.dml_source.as_deref()
                    && !matches!(read, IrStmt::Insert(_) | IrStmt::Update(_) | IrStmt::Delete(_)) =>
            {
                // A subquery, not a `WITH`: this may already sit inside one.
                let mut sql = format!(
                    "    SELECT {}* FROM (\n{}\n    ) AS {}",
                    if inner.distinct { "DISTINCT " } else { "" },
                    emit_dml_as_cte_source(read),
                    qi(&source.alias),
                );
                append_filter(&mut sql, &inner.filter);
                append_order_by(&mut sql, &inner.order_by);
                append_offset_limit(&mut sql, &inner.offset, &inner.limit);
                sql
            }
            [IrRowSource::Bound { source, .. }] => {
                // SELECT-over-SELECT: expose raw columns so the outer SELECT can
                // project its own shape from them, mirroring DML's RETURNING *.
                let from = if inner.polymorphic && !source.table.starts_with("@cte:") {
                    format!(
                        "(\n{}\n    ) AS {}",
                        emit_poly_union(&inner.poly_implementors, &inner.poly_columns),
                        qi(&source.alias),
                    )
                } else {
                    format!("{} AS {}", source_ref(source), qi(&source.alias))
                };
                let mut sql = format!(
                    "    SELECT {}* FROM {}",
                    if inner.distinct { "DISTINCT " } else { "" },
                    from
                );
                append_filter(&mut sql, &inner.filter);
                append_order_by(&mut sql, &inner.order_by);
                append_offset_limit(&mut sql, &inner.offset, &inner.limit);
                sql
            }
            // Object union: same FROM-clause union as `emit_bound_union_select`,
            // but exposing raw columns like every other arm here.
            rows if rows.len() > 1 && rows.iter().all(|r| matches!(r, IrRowSource::Bound { .. })) => {
                let IrRowSource::Bound { source, .. } = &rows[0] else {
                    unreachable!("checked by the guard")
                };
                let mut sql = format!(
                    "    SELECT {}* FROM (\n{}\n    ) AS {}",
                    if inner.distinct { "DISTINCT " } else { "" },
                    bound_union_from_clause(rows, &inner.poly_columns),
                    qi(&source.alias),
                );
                append_filter(&mut sql, &inner.filter);
                append_order_by(&mut sql, &inner.order_by);
                append_offset_limit(&mut sql, &inner.offset, &inner.limit);
                sql
            }
            // Free rows: full UNION-ALL emission, same as before the merge
            // (this arm used to be a separate IrStmt::FreeSelect match). Only
            // `.sql` is used here, and CtePassthrough's SQL text doesn't
            // depend on the ctes list (only its *shape* resolution does), so
            // an empty slice is fine.
            _ => emit_select_stmt(inner, &[]).sql,
        },
        IrStmt::FunctionSelect(sel) => {
            // Expose raw columns so the outer SELECT can project its own shape,
            // mirroring how IrStmt::Select works as a CTE source.
            let args_sql = sel.fn_args.iter().map(emit_expr).collect::<Vec<_>>().join(", ");
            let fn_call = format!("{}.{}({})", pg_schema(&sel.fn_module), qi(&sel.fn_name), args_sql);
            let mut sql = format!("    SELECT * FROM {} AS {}", fn_call, qi(&sel.alias));
            append_filter(&mut sql, &sel.filter);
            append_order_by(&mut sql, &sel.order_by);
            append_offset_limit(&mut sql, &sel.offset, &sel.limit);
            sql
        }
        IrStmt::For(f) => {
            // `for x in S union (...)` as a set source: iterate in the FROM
            // clause and expose the body's raw columns, the same contract the
            // other arms honour. An empty literal iterator yields no rows, but
            // still has to name a column for the outer SELECT to project.
            if matches!(&f.iterator, IrForIterator::Values { exprs, .. } if exprs.is_empty()) {
                return "    SELECT NULL AS \"id\" WHERE FALSE".to_string();
            }
            let (values_from, _) = emit_for_iterator(&f.iterator, &format!("_for_{}", f.var_name));
            let body =
                prefix_body_ctes(emit_dml_as_cte_source(&f.body), &uncorrelated_body_ctes(f)).replace('\n', "\n    ");
            format!(
                "    SELECT \"_body\".*\n    FROM {}\n    CROSS JOIN LATERAL (\n    {}\n    ) AS \"_body\"",
                values_from.replace('\n', "\n    "),
                body,
            )
        }
        IrStmt::Group(grp) if matches!(grp.output, IrGroupOutput::Elements) => emit_group_elements_source(grp),
        IrStmt::Group(_) | IrStmt::VectorSearch(_) | IrStmt::FtsSearch(_) => {
            unreachable!("cannot appear as a CTE source")
        }
        IrStmt::PathSelect(ps) => {
            // A path traversal as a CTE body. What it projects depends on
            // where the path ends: a scalar becomes a `result` column (which
            // is what `IrExpr::CteRef { scalar: true }` reads back), an
            // object exposes its raw columns so the outer query can both
            // traverse it (`@cte:` source) and read its `id`.
            let distinct = if ps.distinct { "DISTINCT " } else { "" };
            let projection = match &ps.result {
                // Same two-column convention a free scalar binding uses:
                // `result` is the ROW() composite top-level decoding wants,
                // `v` the bare value `IrExpr::CteRef { scalar: true }` reads.
                IrPathResult::Scalar(e, _) => format!("{}{} AS v", distinct, emit_expr(e)),
                IrPathResult::Object { alias, .. } => format!("{}{}.*", distinct, qi(alias)),
            };
            let mut sql = format!(
                "SELECT {}\n    FROM {}",
                projection,
                emit_path_joins(&ps.root, &ps.joins)
            );
            append_filter(&mut sql, &ps.filter);
            append_order_by(&mut sql, &ps.order_by);
            append_offset_limit(&mut sql, &ps.offset, &ps.limit);
            match &ps.result {
                IrPathResult::Scalar(e, _) => {
                    let row_value = if enum_type_of_expr(e).is_some() { "v::text" } else { "v" };
                    format!("    SELECT ROW({row_value}) AS result, v FROM ({sql}) AS _scalar")
                }
                IrPathResult::Object { .. } => format!("    {sql}"),
            }
        }
    }
}

// ── Polymorphic DML-as-CTE fan-out ──────────────────────────────────────────
//
// A DML statement targeting an interface/abstract type can't be a single
// `UPDATE`/`DELETE ... RETURNING *` the way emit_dml_as_cte_source handles a
// concrete-type target — Postgres has no single physical relation backing
// the interface (each concrete implementor has its own table), and a bare
// `RETURNING *` off whichever one table `source_ref` happened to resolve
// would only ever see that one implementor's rows and never carry a
// `__type__` discriminator column at all (confirmed live: querying such a
// CTE from an outer polymorphic SELECT raised "column t1.__type__ does not
// exist", since the outer SELECT — see emit_select_stmt and the "CTE-backed
// polymorphic sources" comment there — assumes any CTE tied to a polymorphic
// type already exposes one).
//
// Mirrors emit_poly_update_stmt/emit_poly_delete_stmt's per-implementor
// UNION ALL fan-out (one CTE per concrete table, each RETURNING-ing rows,
// unioned together with a compile-time-known type literal per branch — the
// type is known per branch even though it can vary per row across the whole
// statement, since each branch only ever touches one implementor's table).
// The difference here: those two functions only need `id` back (a
// standalone DML's own default RETURNING shape), while a DML-as-CTE source
// must expose whatever columns the *outer* SELECT might project — so this
// returns the interface's full common column set (poly_columns: every
// property + link FK the type declares, guaranteed present on every
// implementor table) instead of just id.
//
// Not handled: a multi-link mutation (+=/-=/:=) on the interface's own
// pointer combined with poly_implementors on the same statement — that
// would need per-implementor junction-table CTEs too, which
// emit_update_multilink_ctes doesn't do either today. Callers check
// update_has_any_multilink first and keep routing that (rarer) combination
// through the existing (equally not-poly-aware) path rather than silently
// mishandling it here.

fn emit_poly_update_dml_ctes(upd: &IrUpdate, name: &str) -> Vec<String> {
    let alias = &upd.target.alias;
    let sets = update_set_fragments(&upd.assignments, &upd.rewrites, "");
    let col_list = upd.poly_columns.iter().map(|c| qi(c)).collect::<Vec<_>>().join(", ");

    let mut cte_parts = vec![];
    let mut union_parts = vec![];
    for (i, imp) in upd.poly_implementors.iter().enumerate() {
        let cte_name = format!("{}__u{}", name, i);
        let mut upd_sql = format!(
            "UPDATE {} AS {}\nSET {}",
            qn(&imp.module, &imp.table),
            qi(alias),
            sets.join(", "),
        );
        append_filter(&mut upd_sql, &upd.filter);
        upd_sql.push_str(&format!("\nRETURNING {}", col_list));
        cte_parts.push(format!("\"{}\" AS (\n{}\n)", cte_name, upd_sql));

        union_parts.push(format!(
            "SELECT {}::text AS \"__type__\", {} FROM \"{}\"",
            sql_str(&imp.type_name),
            col_list,
            cte_name,
        ));
    }
    cte_parts.push(format!("\"{}\" AS (\n{}\n)", name, union_parts.join("\nUNION ALL\n")));
    cte_parts
}

/// One implementor's own copy of a polymorphic update: it writes to that
/// implementor's table, and to that implementor's own junction tables — an
/// interface's junction is a view over them, which Postgres refuses to
/// insert into.
fn concrete_poly_update(upd: &IrUpdate, imp: &IrPolyImplementor) -> IrUpdate {
    let interface_prefix = format!("{}.", upd.target.table);
    let own_prefix = format!("{}.", imp.table);
    let own_junction = |table: &mut String, module: &mut String| {
        if let Some(link) = table.strip_prefix(&interface_prefix) {
            *table = format!("{own_prefix}{link}");
            module.clone_from(&imp.module);
        }
    };
    let mut concrete = upd.clone();
    concrete.poly_implementors = vec![];
    concrete.target.table = imp.table.clone();
    concrete.target.type_name = imp.type_name.clone();
    concrete.target.poly = None;
    concrete
        .multi_link_clears
        .iter_mut()
        .for_each(|c| own_junction(&mut c.junction_table, &mut c.module));
    concrete
        .multi_link_replaces
        .iter_mut()
        .chain(concrete.multi_link_appends.iter_mut())
        .chain(concrete.multi_link_removals.iter_mut())
        .for_each(|m| own_junction(&mut m.junction_table, &mut m.module));
    concrete
}

/// `emit_poly_update_dml_ctes` for an update that also mutates a multi-link:
/// each implementor gets the whole sibling set `emit_update_multilink_ctes`
/// builds, against its own tables.
fn emit_poly_update_multilink_ctes(upd: &IrUpdate, name: &str) -> Vec<String> {
    let col_list = upd.poly_columns.iter().map(|c| qi(c)).collect::<Vec<_>>().join(", ");
    let mut cte_parts = emit_user_cte_parts(&upd.nested_ctes);
    let mut union_parts = vec![];
    for (i, imp) in upd.poly_implementors.iter().enumerate() {
        let cte_name = format!("{}__u{}", name, i);
        cte_parts.extend(emit_update_multilink_cte_parts(
            &concrete_poly_update(upd, imp),
            &cte_name,
            false,
        ));
        union_parts.push(format!(
            "SELECT {}::text AS \"__type__\", {} FROM \"{}\"",
            sql_str(&imp.type_name),
            col_list,
            cte_name,
        ));
    }
    cte_parts.push(format!("\"{}\" AS (\n{}\n)", name, union_parts.join("\nUNION ALL\n")));
    cte_parts
}

fn emit_poly_delete_dml_ctes(del: &IrDelete, name: &str) -> Vec<String> {
    let alias = &del.target.alias;
    let col_list = del.poly_columns.iter().map(|c| qi(c)).collect::<Vec<_>>().join(", ");

    let mut cte_parts = vec![];
    let mut union_parts = vec![];
    for (i, imp) in del.poly_implementors.iter().enumerate() {
        let cte_name = format!("{}__d{}", name, i);
        let mut del_sql = format!("DELETE FROM {} AS {}", qn(&imp.module, &imp.table), qi(alias),);
        append_filter(&mut del_sql, &del.filter);
        del_sql.push_str(&format!("\nRETURNING {}", col_list));
        cte_parts.push(format!("\"{}\" AS (\n{}\n)", cte_name, del_sql));

        union_parts.push(format!(
            "SELECT {}::text AS \"__type__\", {} FROM \"{}\"",
            sql_str(&imp.type_name),
            col_list,
            cte_name,
        ));
    }
    cte_parts.push(format!("\"{}\" AS (\n{}\n)", name, union_parts.join("\nUNION ALL\n")));
    cte_parts
}

// ── User CTE helpers ────────────────────────────────────────────────────────

fn update_has_any_multilink(upd: &IrUpdate) -> bool {
    !upd.multi_link_clears.is_empty()
        || !upd.multi_link_replaces.is_empty()
        || !upd.multi_link_appends.is_empty()
        || !upd.multi_link_removals.is_empty()
}

/// Extra `AND target NOT IN (...)` clause for a junction clear paired with
/// a replace at the same index (`friends := expr` — see `IrUpdate.
/// multi_link_clears`'s own doc comment on the shared-index pairing).
///
/// Without this, a target present in *both* the old and new sets gets
/// deleted here and then silently fails to come back: the paired INSERT
/// (`emit_ml_append_cte`) runs its `ON CONFLICT` check against the same
/// statement-wide snapshot every CTE in one `WITH` shares, which still
/// shows the row this DELETE is *about to* remove as present — Postgres
/// treats that as a real conflict and skips the insert (`ON CONFLICT DO
/// NOTHING`) or updates the row this DELETE is also deleting (`DO UPDATE`,
/// the link-properties case), not "insert the row back." Confirmed live:
/// reassigning a junction-backed single link to its own already-set value,
/// or replacing a multi-link with a set overlapping its current one,
/// silently dropped every overlapping member. Excluding still-wanted
/// targets from the DELETE instead means they're never touched at all, so
/// there's nothing for the INSERT to conflict with.
fn ml_clear_exclusion(rep: Option<&IrMultiLinkMutation>) -> String {
    match rep {
        Some(rep) => {
            let vals_ref = emit_multilink_values_subquery(&rep.values, &[]);
            format!(
                " AND {} NOT IN (SELECT \"_v\".\"id\" FROM {} AS \"_v\")",
                qi(&rep.target_col),
                vals_ref
            )
        }
        None => String::new(),
    }
}

/// Builds the CTE chain for an UPDATE with multi-link (junction table)
/// mutations, bound to a single external `name` (a user WITH-binding, or the
/// implicit "_dml" wrapper for `SELECT (UPDATE ...)`). Junction INSERT/DELETE
/// statements are themselves data-modifying, and Postgres requires
/// data-modifying CTEs to sit at the *top level* of the query — they can't
/// be nested inside another CTE's own body (confirmed live: nesting them
/// raises "WITH clause containing a data-modifying statement must be at the
/// top level"). So this returns a flat list of sibling CTE parts, prefixed
/// by `name` to stay collision-free alongside any other bound statements in
/// the same WITH block, ending in a `"{name}" AS (SELECT * FROM
/// "{name}__ids")` passthrough — every other reference to `name` keeps
/// seeing the updated row's full columns exactly as `RETURNING *` would have
/// exposed them.
fn emit_update_multilink_ctes(upd: &IrUpdate, name: &str) -> Vec<String> {
    emit_update_multilink_cte_parts(upd, name, true)
}

/// `emit_nested` is false for a caller that has already emitted
/// `upd.nested_ctes` beside this one: a polymorphic update emits one branch
/// per implementor off the same `IrUpdate`, and a `WITH` may name each
/// hoisted statement only once.
fn emit_update_multilink_cte_parts(upd: &IrUpdate, name: &str, emit_nested: bool) -> Vec<String> {
    let alias = &upd.target.alias;
    let has_scalar_changes = !upd.assignments.is_empty() || !upd.rewrites.is_empty();
    let ids_name = format!("{}__ids", name);
    // The nested statements its values read from, ahead of it.
    let mut parts: Vec<String> = if emit_nested {
        emit_user_cte_parts(&upd.nested_ctes)
    } else {
        vec![]
    };

    if has_scalar_changes {
        let sets = update_set_fragments(&upd.assignments, &upd.rewrites, "");
        let mut upd_sql = format!(
            "UPDATE {} AS {}\nSET {}",
            target_ref(&upd.target),
            qi(alias),
            sets.join(", "),
        );
        // A nested statement's CTE sits beside this one, so reading its id
        // needs a FROM — the same reason the plain update path has one.
        upd_sql.push_str(&nested_cte_from(&upd.nested_ctes, &sets.join(","), "\n"));
        append_filter(&mut upd_sql, &upd.filter);
        // See `emit_dml_as_cte_source`: past a FROM, a bare `*` would carry the
        // joined relation's columns into this CTE beside the target's.
        upd_sql.push_str(&format!("\nRETURNING {}.*", qi(alias)));
        parts.push(format!("\"{}\" AS (\n{}\n)", ids_name, upd_sql));
    } else {
        let mut sel = format!(
            "SELECT {}.* FROM {} AS {}",
            qi(alias),
            target_ref(&upd.target),
            qi(alias),
        );
        append_filter(&mut sel, &upd.filter);
        parts.push(format!("\"{}\" AS (\n{}\n)", ids_name, sel));
    }

    for (i, clr) in upd.multi_link_clears.iter().enumerate() {
        let exclude = ml_clear_exclusion(upd.multi_link_replaces.get(i));
        let del = format!(
            "DELETE FROM {} WHERE {} IN (SELECT id FROM \"{}\"){}",
            qn(&clr.module, &clr.junction_table),
            qi(&clr.source_col),
            ids_name,
            exclude,
        );
        parts.push(format!("\"{}__clr_{}\" AS (\n{}\n)", name, i, del));
    }
    for (i, app) in upd.multi_link_appends.iter().enumerate() {
        parts.push(emit_ml_append_cte(
            app,
            &ids_name,
            &upd.target.alias,
            &format!("{}__ml_add_{}", name, i),
        ));
    }
    for (i, rem) in upd.multi_link_removals.iter().enumerate() {
        parts.push(emit_ml_remove_cte(
            rem,
            &ids_name,
            &upd.target.alias,
            &format!("{}__ml_rm_{}", name, i),
        ));
    }
    for (i, rep) in upd.multi_link_replaces.iter().enumerate() {
        parts.push(emit_ml_append_cte(
            rep,
            &ids_name,
            &upd.target.alias,
            &format!("{}__ml_rep_{}", name, i),
        ));
    }

    parts.push(format!("\"{}\" AS (\n    SELECT * FROM \"{}\"\n)", name, ids_name));
    parts
}

fn insert_has_any_multilink(ins: &IrInsert) -> bool {
    !ins.multi_link_appends.is_empty()
}

/// Builds the CTE chain for an INSERT with multi-link (junction table)
/// assignments at creation time — same top-level-CTE constraint as
/// `emit_update_multilink_ctes` (junction INSERTs are themselves
/// data-modifying). The row itself is inserted first (`{name}__ids`, via
/// `RETURNING *`) so the junction CTEs can reference its freshly-generated id.
fn emit_insert_multilink_ctes(ins: &IrInsert, name: &str) -> Vec<String> {
    let ids_name = format!("{}__ids", name);
    let mut parts: Vec<String> = emit_user_cte_parts(&ins.nested_ctes);

    let mut insert_sql = emit_insert_row_sql(ins);
    if let Some(conflict) = &ins.unless_conflict {
        emit_conflict(&mut insert_sql, conflict);
    }
    insert_sql.push_str("\nRETURNING *");
    parts.push(format!("\"{}\" AS (\n{}\n)", ids_name, insert_sql));

    for (i, app) in ins.multi_link_appends.iter().enumerate() {
        parts.push(emit_ml_append_cte(
            app,
            &ids_name,
            &ins.target.alias,
            &format!("{}__ml_add_{}", name, i),
        ));
    }

    parts.push(format!("\"{}\" AS (\n    SELECT * FROM \"{}\"\n)", name, ids_name));
    parts
}

/// A `for` whose body mutates, bound to a WITH name: the same iteration-driven
/// shapes `emit_for_insert`/`emit_for_update` build at the top level, but split
/// into sibling CTE parts. A data-modifying statement can neither sit inside
/// another CTE's body nor in a LATERAL, so the loop cannot be emitted as the
/// binding's own subquery the way a select-bodied one is.
/// A `for` whose body mutates, bound to a WITH name. `outer_alias` names an
/// enclosing loop's iteration when this loop reads it, which makes every row
/// this one writes belong to one iteration of it rather than to all of them. `outer_alias` names
/// that loop's iteration when this loop reads it, which makes every row this
/// one writes belong to one iteration of it rather than to all of them.
fn emit_for_dml_ctes_within(f: &IrFor, name: &str, outer_alias: Option<&str>) -> Vec<String> {
    let iter_alias = format!("_for_{}", f.var_name);
    let (_, plain_iter_cte) = emit_for_iterator(&f.iterator, &iter_alias);
    let correlated = outer_alias;
    let iter_cte = match correlated {
        Some(outer) => emit_nested_for_iterator(&f.iterator, &iter_alias, outer),
        None => plain_iter_cte,
    };
    // The iterator first: a body CTE may read the loop variable, and a WITH
    // name is only in scope for the bindings that follow it.
    let mut parts: Vec<String> = vec![iter_cte];
    parts.extend(emit_user_cte_parts(&f.body_ctes));
    let ids_name = format!("{}__ids", name);

    // Rows written under a correlated loop carry their iteration key, so the
    // enclosing statement can pair them with the row that iteration wrote.
    if correlated.is_some()
        && let IrStmt::Insert(ins) = f.body.as_ref()
        && ins.multi_link_appends.is_empty()
    {
        let rows_name = format!("{}__rows", name);
        let outer_source = format!("{}.{}", qi(&iter_alias), qi(OUTER_KEY));
        parts.push(emit_iteration_rows_cte(ins, &rows_name, &iter_alias, &outer_source));
        parts.push(emit_insert_from_rows_cte(ins, &ids_name, &rows_name));
        parts.push(format!("\"{}\" AS (\n    SELECT * FROM \"{}\"\n)", name, rows_name));
        return parts;
    }

    match f.body.as_ref() {
        IrStmt::Insert(ins) => {
            let rewrite_cols: std::collections::HashSet<&str> =
                ins.rewrites.iter().map(|r| r.column.as_str()).collect();
            let cols: Vec<String> = ins
                .assignments
                .iter()
                .filter(|(c, _)| !rewrite_cols.contains(c.as_str()))
                .map(|(c, _)| qi(c))
                .chain(ins.rewrites.iter().map(|r| qi(&r.column)))
                .collect();
            let values: Vec<String> = ins
                .assignments
                .iter()
                .filter(|(c, _)| !rewrite_cols.contains(c.as_str()))
                .map(|(_, e)| emit_expr(e))
                .chain(ins.rewrites.iter().map(|r| emit_expr(&r.expr)))
                .collect();
            // A nested loop that carried the iteration key back means these
            // rows need one too, to pair with.
            let rows_name = format!("{}__rows", name);
            let correlated_targets =
                correlated_append_indices(&ins.multi_link_appends, &[&f.body_ctes, &ins.nested_ctes]);
            if correlated_targets.is_empty() {
                let mut sql = format!(
                    "INSERT INTO {} ({})\nSELECT {} FROM {}",
                    target_ref(&ins.target),
                    cols.join(", "),
                    values.join(", "),
                    qi(&iter_alias),
                );
                if let Some(conflict) = &ins.unless_conflict {
                    emit_conflict(&mut sql, conflict);
                }
                sql.push_str("\nRETURNING *");
                parts.push(format!("\"{}\" AS (\n{}\n)", ids_name, sql));
            } else {
                let outer_source = match correlated {
                    Some(_) => format!("{}.{}", qi(&iter_alias), qi(OUTER_KEY)),
                    None => format!("{}.{}", qi(&iter_alias), qi("v")),
                };
                parts.push(emit_iteration_rows_cte(ins, &rows_name, &iter_alias, &outer_source));
                parts.push(emit_insert_from_rows_cte(ins, &ids_name, &rows_name));
            }
            for (i, append) in ins.multi_link_appends.iter().enumerate() {
                let cte_name = format!("{}__ml_add_{}", name, i);
                if correlated_targets.contains(&i) {
                    let IrMultiLinkValueSource::CteRef(target) = &append.values.source else {
                        unreachable!("only a CTE-backed value can carry an iteration key")
                    };
                    parts.push(emit_correlated_ml_append_cte(append, &rows_name, target, &cte_name));
                } else {
                    parts.push(emit_ml_append_cte(append, &ids_name, &ins.target.alias, &cte_name));
                }
            }
        }
        IrStmt::Update(upd) => {
            let alias = &upd.target.alias;
            if upd.assignments.is_empty() && upd.rewrites.is_empty() && !upd.multi_link_appends.is_empty() {
                const ITER_COL: &str = "_iter";
                // A nested insert gets its id generated here rather than by the
                // column default, so the junction below can name the row this
                // iteration inserted. `RETURNING` offers no order to pair on.
                let generated: Vec<Option<&IrInsert>> = upd
                    .multi_link_appends
                    .iter()
                    .map(|a| per_iteration_insert(a, &upd.nested_ctes))
                    .collect();
                let new_cols: String = generated
                    .iter()
                    .enumerate()
                    .filter_map(|(i, ins)| {
                        ins.map(|ins| {
                            format!(
                                ", {} AS {}",
                                ins.id_default_sql.as_deref().unwrap_or("uuidv7()"),
                                qi(&format!("_new_{i}"))
                            )
                        })
                    })
                    .collect();
                let mut ids_sql = format!(
                    "\"{}\" AS (\nSELECT {}.*, {}.\"v\" AS {}{} FROM {} AS {}, {}",
                    ids_name,
                    qi(alias),
                    qi(&iter_alias),
                    qi(ITER_COL),
                    new_cols,
                    target_ref(&upd.target),
                    qi(alias),
                    qi(&iter_alias),
                );
                append_filter(&mut ids_sql, &upd.filter);
                ids_sql.push_str("\n)");
                parts.push(ids_sql);
                for (i, append) in upd.multi_link_appends.iter().enumerate() {
                    match generated[i] {
                        Some(ins) => {
                            parts.push(emit_for_nested_insert_cte(
                                ins,
                                &ids_name,
                                &iter_alias,
                                &format!("{}__ml_ins_{}", name, i),
                                &format!("_new_{i}"),
                            ));
                            parts.push(emit_for_ml_append_cte(
                                append,
                                &ids_name,
                                &format!("{}__ml_add_{}", name, i),
                                &format!("_new_{i}"),
                            ));
                        }
                        None => parts.push(emit_for_ml_append_cte(
                            append,
                            &ids_name,
                            &format!("{}__ml_add_{}", name, i),
                            ITER_COL,
                        )),
                    }
                }
            } else {
                let sets = update_set_fragments(&upd.assignments, &upd.rewrites, "");
                let mut sql = format!(
                    "UPDATE {} AS {}\nSET {}\nFROM {}",
                    target_ref(&upd.target),
                    qi(alias),
                    sets.join(", "),
                    qi(&iter_alias),
                );
                append_filter(&mut sql, &upd.filter);
                sql.push_str(&format!("\nRETURNING {}.*", qi(alias)));
                parts.push(format!("\"{}\" AS (\n{}\n)", ids_name, sql));
            }
        }
        // `for line in … union (for component in … union (insert …))` — one
        // iterator CTE per loop, the inner carrying the outer's key, and the
        // insert driven from the two joined on it.
        IrStmt::For(inner) => {
            let inner_alias = format!("_for_{}", inner.var_name);
            parts.push(emit_nested_for_iterator(&inner.iterator, &inner_alias, &iter_alias));
            parts.extend(emit_user_cte_parts(&inner.body_ctes));
            let IrStmt::Insert(ins) = inner.body.as_ref() else {
                unreachable!("a nested for-loop's own body is an insert: {:?}", inner.body)
            };
            let rewrite_cols: std::collections::HashSet<&str> =
                ins.rewrites.iter().map(|r| r.column.as_str()).collect();
            let cols: Vec<String> = ins
                .assignments
                .iter()
                .filter(|(c, _)| !rewrite_cols.contains(c.as_str()))
                .map(|(c, _)| qi(c))
                .chain(ins.rewrites.iter().map(|r| qi(&r.column)))
                .collect();
            let values: Vec<String> = ins
                .assignments
                .iter()
                .filter(|(c, _)| !rewrite_cols.contains(c.as_str()))
                .map(|(_, e)| emit_expr(e))
                .chain(ins.rewrites.iter().map(|r| emit_expr(&r.expr)))
                .collect();
            let mut sql = format!(
                "INSERT INTO {} ({})\nSELECT {} FROM {}",
                target_ref(&ins.target),
                cols.join(", "),
                values.join(", "),
                nested_for_from(&inner_alias, &iter_alias),
            );
            if let Some(conflict) = &ins.unless_conflict {
                emit_conflict(&mut sql, conflict);
            }
            sql.push_str("\nRETURNING *");
            parts.push(format!("\"{}\" AS (\n{}\n)", ids_name, sql));
            for (i, append) in ins.multi_link_appends.iter().enumerate() {
                parts.push(emit_ml_append_cte(
                    append,
                    &ids_name,
                    &ins.target.alias,
                    &format!("{}__ml_add_{}", name, i),
                ));
            }
        }
        // Unreachable: only a mutating body is routed here.
        other => unreachable!("for-loop body is not a mutation: {other:?}"),
    }

    parts.push(format!("\"{}\" AS (\n    SELECT * FROM \"{}\"\n)", name, ids_name));
    parts
}

/// Expands a list of user-bound WITH names into their top-level CTE parts —
/// usually one part per name, except an UPDATE or INSERT with any multi-link
/// mutation expands into several sibling parts (see
/// emit_update_multilink_ctes / emit_insert_multilink_ctes), and an UPDATE or
/// DELETE targeting an interface/abstract type similarly fans out into one
/// part per concrete implementor (see emit_poly_update_dml_ctes /
/// emit_poly_delete_dml_ctes) — neither can be nested inside a single name's
/// own CTE body.
fn emit_user_cte_parts(ctes: &[IrCteDef]) -> Vec<String> {
    let mut parts: Vec<String> = vec![];
    for c in ctes {
        // A hoisted statement may have hoisted one of its own —
        // `credentials := (insert access::Credentials { … })` inside an update.
        // Those CTEs are emitted beside it rather than inside its body, which
        // is where its own SQL expects to read them from.
        // The multi-link builders emit a statement's own, ahead of it.
        let nested: &[IrCteDef] = match &c.stmt {
            IrStmt::Insert(ins) if insert_has_any_multilink(ins) => &[],
            IrStmt::Update(upd) if update_has_any_multilink(upd) => &[],
            IrStmt::Insert(ins) => &ins.nested_ctes,
            IrStmt::Update(upd) => &upd.nested_ctes,
            _ => &[],
        };
        if !nested.is_empty() {
            parts.extend(emit_user_cte_parts(nested));
        }
        if let IrStmt::Update(upd) = &c.stmt {
            if !upd.poly_implementors.is_empty() {
                parts.extend(if update_has_any_multilink(upd) {
                    emit_poly_update_multilink_ctes(upd, &c.name)
                } else {
                    emit_poly_update_dml_ctes(upd, &c.name)
                });
                continue;
            }
            if update_has_any_multilink(upd) {
                parts.extend(emit_update_multilink_ctes(upd, &c.name));
                continue;
            }
        }
        if let IrStmt::Insert(ins) = &c.stmt
            && insert_has_any_multilink(ins)
        {
            parts.extend(emit_insert_multilink_ctes(ins, &c.name));
            continue;
        }
        if let IrStmt::Delete(del) = &c.stmt
            && !del.poly_implementors.is_empty()
        {
            parts.extend(emit_poly_delete_dml_ctes(del, &c.name));
            continue;
        }
        if let IrStmt::For(f) = &c.stmt
            && matches!(f.body.as_ref(), IrStmt::Insert(_) | IrStmt::Update(_) | IrStmt::For(_))
        {
            parts.extend(emit_for_dml_ctes_within(f, &c.name, c.correlated_to.as_deref()));
            continue;
        }
        let body = emit_dml_as_cte_source(&c.stmt);
        parts.push(match &c.correlated_to {
            // Evaluated once per iteration and keyed by it, since a CTE is
            // evaluated once for the whole statement whatever it reads.
            Some(iterator) => format!(
                "\"{}\" AS (\nSELECT {}.\"v\" AS {}, \"_row\".*\nFROM {}\nCROSS JOIN LATERAL (\n{}\n) AS \"_row\"\n)",
                c.name,
                qi(iterator),
                qi(OUTER_KEY),
                qi(iterator),
                body,
            ),
            None => format!("\"{}\" AS (\n{}\n)", c.name, body),
        });
    }
    parts
}

/// Emit `WITH "name" AS (...), ...` prefix (WITH keyword + trailing newline included).
fn emit_cte_prefix(ctes: &[IrCteDef]) -> String {
    format!("WITH\n{}\n", emit_user_cte_parts(ctes).join(",\n"))
}

/// Collects every link-property name assigned anywhere in a multilink value
/// tree (both sides of any nested `union`), in first-seen order — the SQL
/// layer needs one consistent column list across every unioned branch, even
/// when different targets set different (or no) properties.
pub(crate) fn collect_link_prop_names(vals: &IrMultiLinkValues, names: &mut Vec<String>) {
    for (name, _) in &vals.link_props {
        if !names.contains(name) {
            names.push(name.clone());
        }
    }
    match &vals.source {
        IrMultiLinkValueSource::Union(a, b) => {
            collect_link_prop_names(a, names);
            collect_link_prop_names(b, names);
        }
        IrMultiLinkValueSource::Asserted { inner, .. } => collect_link_prop_names(inner, names),
        _ => {}
    }
}

/// `, <expr> AS "name"` for each entry in `prop_names` — the value assigned to
/// *this* node if any, else a bare `NULL` (Postgres infers its type from the
/// other union branch's typed value in the same column position; if no
/// branch ever sets it, the INSERT's target column type coerces it).
fn emit_link_prop_cols(vals: &IrMultiLinkValues, prop_names: &[String]) -> String {
    prop_names
        .iter()
        .map(|name| match vals.link_props.iter().find(|(n, _)| n == name) {
            Some((_, expr)) => format!(", {} AS {}", emit_expr(expr), qi(name)),
            None => format!(", NULL AS {}", qi(name)),
        })
        .collect()
}

/// Emit `(SELECT id[, prop, ...] FROM ...)` sub-expression for a multi-link
/// values source. `prop_names` is the full set of link-property names used
/// anywhere in the enclosing mutation (see `collect_link_prop_names`) — every
/// leaf projects all of them so a `union` of heterogeneous branches has a
/// consistent column list.
fn emit_multilink_values_subquery(vals: &IrMultiLinkValues, prop_names: &[String]) -> String {
    emit_multilink_values_inner(vals, prop_names, true)
}

/// `bare_cte_ok` is false wherever the result has to be a SELECT in its own
/// right — a `UNION ALL` operand, or the body of an assert. A CTE reference
/// otherwise emits just its quoted name, which only works where the caller
/// aliases it into a FROM.
fn emit_multilink_values_inner(vals: &IrMultiLinkValues, prop_names: &[String], bare_cte_ok: bool) -> String {
    if let IrMultiLinkValueSource::Union(a, b) = &vals.source {
        return format!(
            "({}\nUNION ALL\n{})",
            emit_multilink_values_inner(a, prop_names, false),
            emit_multilink_values_inner(b, prop_names, false),
        );
    }

    // The assert wraps the whole value, so the ids are collected once, passed
    // through the check, and unnested back into rows. Reading the inner value
    // twice — once for the check, once for the rows — would run any `insert`
    // inside it twice.
    if let IrMultiLinkValueSource::Asserted {
        fn_name,
        inner,
        message,
    } = &vals.source
    {
        let inner_sql = emit_multilink_values_inner(inner, &[], false);
        let prop_cols = emit_link_prop_cols(vals, prop_names);
        return format!(
            "(SELECT unnest(\"_pylon\".{}(ARRAY(SELECT \"_v\".\"id\" FROM {} AS \"_v\"){})) AS \"id\"{})",
            qi(fn_name),
            inner_sql,
            assert_message_arg(message),
            prop_cols,
        );
    }

    let prop_cols = emit_link_prop_cols(vals, prop_names);

    match &vals.source {
        IrMultiLinkValueSource::CteRef(name) => {
            if prop_cols.is_empty() && bare_cte_ok {
                // Just the CTE name; will be aliased at the call site.
                format!("\"{}\"", name)
            } else {
                format!("(SELECT \"_s\".\"id\"{} FROM \"{}\" AS \"_s\")", prop_cols, name)
            }
        }
        IrMultiLinkValueSource::Select(s) => {
            // Compiler only ever constructs this variant for a single
            // schema-bound row (compile_subquery_exists's callers reject
            // free rows before wrapping) — see compiler.rs's multilink-
            // value-source sites.
            let [IrRowSource::Bound { source, .. }] = s.rows.as_slice() else {
                unreachable!("IrMultiLinkValueSource::Select is always schema-bound")
            };
            let alias = &source.alias;
            let mut sql = format!(
                "(SELECT {}.\"id\"{} FROM {} AS {}",
                qi(alias),
                prop_cols,
                source_ref(source),
                qi(alias)
            );
            append_filter(&mut sql, &s.filter);
            sql.push(')');
            sql
        }
        IrMultiLinkValueSource::PathSelect(ps) => {
            let root_alias = &ps.root.alias;
            // The targets are the rows the walk lands on, not the one it
            // starts from.
            let target_alias = match &ps.result {
                IrPathResult::Object { alias, .. } => alias,
                IrPathResult::Scalar(..) => root_alias,
            };
            let mut sql = format!(
                "(SELECT {}.\"id\"{} FROM {} AS {}",
                qi(target_alias),
                prop_cols,
                source_ref(&ps.root),
                qi(root_alias)
            );
            for join in &ps.joins {
                sql.push_str(&emit_path_join_sql(join));
            }
            append_filter(&mut sql, &ps.filter);
            sql.push(')');
            sql
        }
        IrMultiLinkValueSource::Function(fs) => {
            let args = fs.fn_args.iter().map(emit_expr).collect::<Vec<_>>().join(", ");
            let mut sql = format!(
                "(SELECT {}.\"id\"{} FROM {}.{}({}) AS {}",
                qi(&fs.alias),
                prop_cols,
                pg_schema(&fs.fn_module),
                qi(&fs.fn_name),
                args,
                qi(&fs.alias),
            );
            append_filter(&mut sql, &fs.filter);
            sql.push(')');
            sql
        }
        IrMultiLinkValueSource::Union(..) | IrMultiLinkValueSource::Asserted { .. } => {
            unreachable!("handled above")
        }
    }
}

/// SQL fragment for a single path join (used in multilink values emission).
fn emit_path_join_sql(join: &IrPathJoin) -> String {
    match join {
        IrPathJoin::Lateral { inner, target } => {
            let projected = match &inner.result {
                IrPathResult::Object { alias, .. } => alias.clone(),
                IrPathResult::Scalar(..) => unreachable!("a lateral step is only built for an object-valued computed"),
            };
            let mut body = format!(
                "SELECT {}.* FROM {}",
                qi(&projected),
                emit_path_joins(&inner.root, &inner.joins)
            );
            append_filter(&mut body, &inner.filter);
            append_order_by(&mut body, &inner.order_by);
            append_offset_limit(&mut body, &inner.offset, &inner.limit);
            format!(" JOIN LATERAL (\n{}\n) AS {} ON TRUE", body, qi(&target.alias))
        }
        IrPathJoin::Function {
            fn_module,
            fn_name,
            args,
            target,
        } => {
            let args_sql = args.iter().map(emit_expr).collect::<Vec<_>>().join(", ");
            format!(
                " CROSS JOIN LATERAL {}.{}({}) AS {}",
                pg_schema(fn_module),
                qi(fn_name),
                args_sql,
                qi(&target.alias),
            )
        }
        IrPathJoin::Single {
            source_alias,
            fk_col,
            target,
        } => {
            format!(
                " JOIN {} AS {} ON {}.\"id\" = {}.{}",
                source_ref(target),
                qi(&target.alias),
                qi(&target.alias),
                qi(source_alias),
                qi(fk_col)
            )
        }
        IrPathJoin::Multi {
            source_alias,
            junction_alias,
            join: ml_join,
            target,
        } => {
            let (jt_ref, src_col, tgt_col) = match ml_join {
                IrMultiLinkJoin::Standard { junction_table, module } => (
                    junction_ref(module, junction_table),
                    "source".to_string(),
                    "target".to_string(),
                ),
                IrMultiLinkJoin::Through {
                    junction_table,
                    module,
                    source_col,
                    target_col,
                } => (
                    junction_ref(module, junction_table),
                    source_col.clone(),
                    target_col.clone(),
                ),
                // A forward `IrPathJoin::Multi` step is always built from a
                // real multi-link (`compile_path_select`'s middle-step
                // handling), never a backlink — the Backlink* variants only
                // ever appear inside `IrShapePointer::MultiLink.join`.
                IrMultiLinkJoin::BacklinkFk { .. } | IrMultiLinkJoin::BacklinkJunction { .. } => {
                    unreachable!("a forward multi-link path step never uses a backlink join variant")
                }
            };
            format!(
                " JOIN {} AS {} ON {}.{} = {}.\"id\" JOIN {} AS {} ON {}.{} = {}.\"id\"",
                jt_ref,
                qi(junction_alias),
                qi(junction_alias),
                qi(&src_col),
                qi(source_alias),
                source_ref(target),
                qi(&target.alias),
                qi(junction_alias),
                qi(&tgt_col),
                qi(&target.alias),
            )
        }
        IrPathJoin::BacklinkSingle {
            source_alias,
            fk_col,
            target,
        } => {
            format!(
                " JOIN {} AS {} ON {}.{} = {}.\"id\"",
                source_ref(target),
                qi(&target.alias),
                qi(&target.alias),
                qi(fk_col),
                qi(source_alias),
            )
        }
        IrPathJoin::BacklinkMulti {
            source_alias,
            junction_alias,
            junction_table,
            module,
            owner_col,
            current_col,
            target,
        } => {
            format!(
                " JOIN {} AS {} ON {}.{} = {}.\"id\" JOIN {} AS {} ON {}.\"id\" = {}.{}",
                junction_ref(module, junction_table),
                qi(junction_alias),
                qi(junction_alias),
                qi(current_col),
                qi(source_alias),
                source_ref(target),
                qi(&target.alias),
                qi(&target.alias),
                qi(junction_alias),
                qi(owner_col),
            )
        }
    }
}

/// Emit the CTE clause for a junction table INSERT (append / replace-insert).
/// `ids_name` is the row-source CTE to join against (usually "_ids", but
/// callers hoisting this into a shared top-level WITH block alongside other
/// user-bound statements pass a name prefixed for collision-safety instead).
/// The loop variable an append's value stands for, when the value is exactly
/// that variable (`memberships += membership`).
///
/// Such a value is compiled as a select over the target table narrowed to the
/// key the variable holds. Inside a `for` that key is already carried on the
/// driving rows, so the junction insert can read it from there instead of
/// correlating to a CTE it has no range-table entry for.
pub(crate) fn append_value_is_the_loop_variable(values: &IrMultiLinkValues, var: &str) -> bool {
    let IrMultiLinkValueSource::Select(sel) = &values.source else {
        return false;
    };
    let Some(IrExpr::BinOp(cmp)) = sel.filter.as_ref() else {
        return false;
    };
    matches!(
        (&cmp.left, &cmp.op, &cmp.right),
        (
            IrExpr::ColumnRef { column, .. },
            crate::parse::ast::BinOpKind::Eq,
            IrExpr::ForVar { name, .. },
        ) if column == "id" && name == var
    ) && values.link_props.is_empty()
}

/// `emit_ml_append_cte` for a `for` body: the target comes from the column the
/// driving rows carry rather than from a correlated subquery.
/// The nested insert a multi-link append's value names, when it is one of this
/// update's own hoisted CTEs. Inside a `for` body such an insert runs once per
/// iteration, so it cannot stay a standalone CTE.
pub(crate) fn per_iteration_insert<'c>(mutation: &IrMultiLinkMutation, nested: &'c [IrCteDef]) -> Option<&'c IrInsert> {
    let IrMultiLinkValueSource::CteRef(name) = &mutation.values.source else {
        return None;
    };
    if !mutation.values.link_props.is_empty() {
        return None;
    }
    nested.iter().find(|c| &c.name == name).and_then(|c| match &c.stmt {
        IrStmt::Insert(ins) if ins.id_default_sql.is_some() && ins.multi_link_appends.is_empty() => Some(ins),
        _ => None,
    })
}

/// A nested insert inside a `for` body: one row per driving row, with its id
/// taken from the column the driving CTE generated so the junction can pair
/// with it. The iterator CTE is joined in because the inserted values are
/// correlated subqueries that read the loop variable.
fn emit_for_nested_insert_cte(
    ins: &IrInsert,
    ids_name: &str,
    iter_alias: &str,
    cte_name: &str,
    id_col: &str,
) -> String {
    let mut cols = vec![qi("id")];
    let mut values = vec![format!("\"{}\".{}", ids_name, qi(id_col))];
    for (col, expr) in &ins.assignments {
        cols.push(qi(col));
        values.push(emit_expr(expr));
    }
    format!(
        "\"{}\" AS (\nINSERT INTO {} ({})\nSELECT {} FROM \"{}\", {}\nWHERE \"{}\".{} = {}.\"v\"\nRETURNING {}\n)",
        cte_name,
        target_ref(&ins.target),
        cols.join(", "),
        values.join(", "),
        ids_name,
        qi(iter_alias),
        ids_name,
        qi("_iter"),
        qi(iter_alias),
        qi("id"),
    )
}

fn emit_for_ml_append_cte(mutation: &IrMultiLinkMutation, ids_name: &str, cte_name: &str, iter_col: &str) -> String {
    format!(
        "\"{}\" AS (\nINSERT INTO {} ({}, {})\nSELECT \"{}\".\"id\", \"{}\".{} FROM \"{}\"\nON CONFLICT DO NOTHING\nRETURNING {}, {}\n)",
        cte_name,
        qn(&mutation.module, &mutation.junction_table),
        qi(&mutation.source_col),
        qi(&mutation.target_col),
        ids_name,
        ids_name,
        qi(iter_col),
        ids_name,
        qi(&mutation.source_col),
        qi(&mutation.target_col),
    )
}

/// `ids_alias` is the name the rows being written go by inside the value:
/// a walk off the row (`+= (select .notifications …)`) is correlated to it.
fn emit_ml_append_cte(mutation: &IrMultiLinkMutation, ids_name: &str, ids_alias: &str, cte_name: &str) -> String {
    let mut prop_names = vec![];
    collect_link_prop_names(&mutation.values, &mut prop_names);
    let vals_ref = emit_multilink_values_subquery(&mutation.values, &prop_names);

    let extra_cols: String = prop_names.iter().map(|n| format!(", {}", qi(n))).collect();
    let extra_select: String = prop_names.iter().map(|n| format!(", \"_v\".{}", qi(n))).collect();

    // Re-checking an already-linked target with a new property value must
    // update in place rather than error — a bare `DO NOTHING` (the no-
    // properties case) would silently keep the old value instead.
    let conflict_clause = if mutation.single {
        // A junction-backed single link's table has `PRIMARY KEY (source)`
        // alone (D3, cardinality-one) — the paired `_clr_i` DELETE (see
        // `ml_clear_exclusion`) removes any existing row for this source in
        // a sibling CTE that shares this whole statement's one snapshot, so
        // this INSERT still "sees" that old row and treats it as a
        // same-key conflict regardless of what the new target is — a plain
        // `DO NOTHING` would then silently drop a genuine reassignment.
        // `source` alone is always this table's conflict target, so this
        // must always be an upsert, not just when link properties exist.
        let mut sets = vec![format!(
            "{} = EXCLUDED.{}",
            qi(&mutation.target_col),
            qi(&mutation.target_col)
        )];
        sets.extend(prop_names.iter().map(|n| format!("{} = EXCLUDED.{}", qi(n), qi(n))));
        format!(
            "ON CONFLICT ({}) DO UPDATE SET {}",
            qi(&mutation.source_col),
            sets.join(", ")
        )
    } else if prop_names.is_empty() {
        "ON CONFLICT DO NOTHING".to_string()
    } else {
        let sets: Vec<String> = prop_names
            .iter()
            .map(|n| format!("{} = EXCLUDED.{}", qi(n), qi(n)))
            .collect();
        format!(
            "ON CONFLICT ({}, {}) DO UPDATE SET {}",
            qi(&mutation.source_col),
            qi(&mutation.target_col),
            sets.join(", "),
        )
    };

    // A subquery value may read the row it is written for, so it joins
    // laterally; a bare CTE name cannot take LATERAL.
    let lateral = if vals_ref.starts_with('(') { "LATERAL " } else { "" };
    let ins = format!(
        "INSERT INTO {} ({}, {}{})\nSELECT {}.\"id\", \"_v\".\"id\"{} FROM \"{}\" AS {} CROSS JOIN {}{} AS \"_v\"\n{}\nRETURNING {}, {}",
        qn(&mutation.module, &mutation.junction_table),
        qi(&mutation.source_col),
        qi(&mutation.target_col),
        extra_cols,
        qi(ids_alias),
        extra_select,
        ids_name,
        qi(ids_alias),
        lateral,
        vals_ref,
        conflict_clause,
        qi(&mutation.source_col),
        qi(&mutation.target_col),
    );
    format!("\"{}\" AS (\n{}\n)", cte_name, ins)
}

/// Emit the CTE clause for a junction table DELETE (remove). See
/// emit_ml_append_cte re: `ids_name` and `ids_alias`.
fn emit_ml_remove_cte(mutation: &IrMultiLinkMutation, ids_name: &str, ids_alias: &str, cte_name: &str) -> String {
    // Removal never carries link-property assignments (rejected at compile
    // time in ir/compiler.rs), so no extra columns are ever needed here.
    let vals_ref = emit_multilink_values_subquery(&mutation.values, &[]);
    // A walk off the row (`-= (select .line_items …)`) is correlated to
    // `ids_alias`, so the rows being written are joined in under that name
    // rather than read through an uncorrelated `IN`.
    let del = format!(
        "DELETE FROM {} AS \"_jn\"\nUSING \"{}\" AS {}\nWHERE \"_jn\".{} = {}.\"id\"\n  AND \"_jn\".{} IN (SELECT \"id\" FROM {})\nRETURNING \"_jn\".{}, \"_jn\".{}",
        qn(&mutation.module, &mutation.junction_table),
        ids_name,
        qi(ids_alias),
        qi(&mutation.source_col),
        qi(ids_alias),
        qi(&mutation.target_col),
        vals_ref,
        qi(&mutation.source_col),
        qi(&mutation.target_col),
    );
    format!("\"{}\" AS (\n{}\n)", cte_name, del)
}

// ── Conflict helper ─────────────────────────────────────────────────────────

use crate::ir::IrConflict;

fn emit_conflict(sql: &mut String, conflict: &IrConflict) {
    // `unless conflict on (.email, .invitee)` names several columns, not one
    // composite value: Postgres infers the arbiter index from a column list,
    // and would look for an expression index on the tuple instead.
    let on_sql = conflict.on.as_ref().map(|e| match e {
        IrExpr::Tuple(elements) => format!("({})", elements.iter().map(emit_expr).collect::<Vec<_>>().join(", ")),
        other => format!("({})", emit_expr(other)),
    });
    match (&on_sql, &conflict.do_update) {
        (None, None) => sql.push_str(" ON CONFLICT DO NOTHING"),
        (Some(on), None) => sql.push_str(&format!(" ON CONFLICT {} DO NOTHING", on)),
        (None, Some(updates)) => {
            sql.push_str(&format!(" ON CONFLICT DO UPDATE SET {}", do_update_sets(updates)));
        }
        (Some(on), Some(updates)) => {
            sql.push_str(&format!(
                " ON CONFLICT {} DO UPDATE SET {}",
                on,
                do_update_sets(updates),
            ));
        }
    }
    if conflict.do_update.is_some()
        && let Some(predicate) = &conflict.do_update_where
    {
        sql.push_str(&format!(" WHERE {}", emit_expr(predicate)));
    }
}

fn do_update_sets(updates: &[(String, IrExpr)]) -> String {
    updates
        .iter()
        .map(|(col, expr)| format!("{} = {}", qi(col), emit_expr(expr)))
        .collect::<Vec<_>>()
        .join(", ")
}

// ── FREE SELECT ─────────────────────────────────────────────────────────────

/// Types returned as plain top-level columns rather than wrapped in the
/// `result` ROW() composite, so each keeps its own natural top-level shape.
fn is_integer_expr(expr: &IrExpr) -> bool {
    match expr {
        IrExpr::ColumnRef { pg_type, .. } => matches!(
            pg_type.as_str(),
            "int2" | "int4" | "int8" | "integer" | "bigint" | "smallint"
        ),
        IrExpr::Literal(crate::ir::IrLiteral::Int(_)) => true,
        IrExpr::BinOp(op) => is_integer_expr(&op.left) && is_integer_expr(&op.right),
        _ => false,
    }
}

fn is_raw_scalar(expr: &IrExpr) -> bool {
    matches!(expr, IrExpr::Array(_))
        || matches!(expr, IrExpr::TypeCast(c) if c.pg_type == "jsonb")
        || matches!(expr, IrExpr::NamedTuple { .. })
        || matches!(expr, IrExpr::Tuple(_))
        || matches!(expr, IrExpr::JsonbField { .. })
        || matches!(expr, IrExpr::JsonbIndex { .. })
}

/// Free (non-schema-bound) SELECT rows — one or more UNION ALL branches with
/// a ROW(…) wrapper. `rows` is `sel.rows` with any leading `Bound` entries
/// already ruled out by `emit_select_stmt`'s dispatch (a schema object and a
/// free literal can never appear in the same UNION — `check_union_type_compat`
/// rejects that at compile time), so every entry here is `IrRowSource::Free`.
fn emit_free_rows(sel: &IrSelect, rows: &[IrRowSource], ctes: &[IrCteDef]) -> SqlOutput {
    use crate::query::ShapeNode;

    let items: Vec<&IrFreeExpr> = rows
        .iter()
        .map(|r| match r {
            IrRowSource::Free(item) => item,
            IrRowSource::Bound { .. } => unreachable!("mixed Bound/Free rows rejected at compile time"),
        })
        .collect();

    if items.is_empty() {
        return SqlOutput {
            sql: "SELECT NULL AS result WHERE FALSE".to_string(),
            shape: ShapeDescriptor {
                root: ShapeNode::Scalar {
                    name: String::new(),
                    position: 0,
                },
            },
            inference_plan: None,
        };
    }

    // assert_exists / assert_distinct: set-returning — emit as unnest, not UNION ALL
    if items.len() == 1
        && let IrFreeExpr::AssertSet {
            fn_name,
            inner,
            message,
        } = items[0]
    {
        let array_sql = emit_array_source(inner) + &assert_message_arg(message);
        let mut sql = format!(
            "SELECT ROW(v) AS result FROM unnest(\"_pylon\".{}({})) AS _assert(v)",
            fn_name, array_sql,
        );
        if sel.distinct {
            sql = format!("SELECT DISTINCT * FROM ({}) AS \"_distinct\"", sql);
        }
        append_order_by(&mut sql, &sel.order_by);
        append_offset_limit(&mut sql, &sel.offset, &sel.limit);
        return SqlOutput {
            sql,
            shape: ShapeDescriptor {
                root: ShapeNode::Scalar {
                    name: String::new(),
                    position: 0,
                },
            },
            inference_plan: None,
        };
    }

    let shape_root = free_item_shape(items.first().unwrap(), ctes);

    let branches: Vec<String> = items
        .iter()
        .map(|item| match item {
            IrFreeExpr::Scalar(expr) => {
                // An object standing as the whole result is already the
                // composite row `result` is meant to be — wrapping it in
                // another ROW() would bury its columns one level down from
                // where its shape node says they are.
                if matches!(expr, IrExpr::ObjectSubquery(_)) {
                    return format!("SELECT v AS result, v FROM (SELECT {} AS v) AS _obj", emit_expr(expr));
                }
                // Arrays and jsonb are returned as plain top-level columns
                // rather than wrapped, so they keep their own shape node.
                if is_raw_scalar(expr) {
                    // `v` alongside it for the same reason the wrapped form
                    // has one: a `with` binding of an array is read by name.
                    format!("SELECT v AS result, v FROM (SELECT {} AS v) AS _raw", emit_expr(expr))
                } else {
                    // `result` is a ROW() composite for top-level decoding.
                    // `v` is the unwrapped scalar for use in CteRef expression context.
                    // Wrap in a subquery so volatile functions (nextval, etc.) are called once.
                    // Enum values also need a ::text cast inside the ROW() — same
                    // unregistered-OID problem as arrays/jsonb — but the bare `v`
                    // column stays natively typed for CteRef expression use.
                    let e = emit_expr(expr);
                    let row_value = if enum_type_of_expr(expr).is_some() {
                        "v::text"
                    } else {
                        "v"
                    };
                    // In a set of several, an element that is empty is not
                    // in the set at all (`{a.id, b.id}` with no `a` is just
                    // `b.id`) — where SQL would carry it as a NULL row.
                    let present = if items.len() > 1 { " WHERE v IS NOT NULL" } else { "" };
                    format!("SELECT ROW({row_value}) AS result, v FROM (SELECT {e} AS v) AS _scalar{present}")
                }
            }
            IrFreeExpr::FreeObject(fields) => {
                // Each field is computed once in an inner subquery (so a
                // volatile expression like nextval() isn't evaluated twice) and
                // exposed both packed into the `result` composite (for whole-
                // object passthrough/decoding) and as its own named column (for
                // `IrExpr::CteFieldRef` — `with x := {a := ...} select x.a`).
                let inner_cols: Vec<String> = fields
                    .iter()
                    .enumerate()
                    .map(|(i, (_, e))| format!("{} AS \"_f{}\"", emit_expr(e), i))
                    .collect();
                let row_items: Vec<String> = fields
                    .iter()
                    .enumerate()
                    .map(|(i, (_, e))| {
                        if enum_type_of_expr(e).is_some() {
                            format!("\"_f{}\"::text", i)
                        } else {
                            format!("\"_f{}\"", i)
                        }
                    })
                    .collect();
                let named_cols: Vec<String> = fields
                    .iter()
                    .enumerate()
                    .map(|(i, (name, _))| format!("\"_f{}\" AS {}", i, qi(name)))
                    .collect();
                format!(
                    "SELECT ROW({}) AS result, {} FROM (SELECT {}) AS _obj",
                    row_items.join(", "),
                    named_cols.join(", "),
                    inner_cols.join(", "),
                )
            }
            IrFreeExpr::Tuple(exprs) => {
                if exprs.len() == 1 {
                    format!("SELECT ROW({}) AS result", emit_free_field_expr(&exprs[0]))
                } else {
                    let parts: Vec<String> = exprs.iter().map(emit_free_field_expr).collect();
                    format!("SELECT ({}) AS result", parts.join(", "))
                }
            }
            IrFreeExpr::NamedTupleRow(fields) => {
                let parts: Vec<String> = fields.iter().map(|(_, e)| emit_free_field_expr(e)).collect();
                if parts.len() == 1 {
                    format!("SELECT ROW({}) AS result", parts[0])
                } else {
                    format!("SELECT ({}) AS result", parts.join(", "))
                }
            }
            IrFreeExpr::AssertSet { .. } => unreachable!("AssertSet is handled by early return above"),
            IrFreeExpr::CtePassthrough(name) => format!("SELECT \"result\" FROM {}", qi(name)),
        })
        .collect();

    let union_sql = branches.join("\nUNION ALL\n");

    let mut sql = if sel.distinct {
        // Wrap UNION ALL in an outer SELECT DISTINCT to deduplicate.
        format!("SELECT DISTINCT * FROM (\n{}\n) AS \"_distinct\"", union_sql)
    } else {
        union_sql
    };
    // A free SELECT's FILTER has no row of its own to sit beside, so it gates
    // the result the branches already produced rather than joining their own
    // WHERE clauses.
    if let Some(filter) = &sel.filter {
        sql = format!(
            "SELECT * FROM (\n{}\n) AS \"_filtered\"\nWHERE {}",
            sql,
            emit_expr(filter)
        );
    }
    append_order_by(&mut sql, &sel.order_by);
    append_offset_limit(&mut sql, &sel.offset, &sel.limit);

    SqlOutput {
        sql,
        shape: ShapeDescriptor { root: shape_root },
        inference_plan: None,
    }
}

/// Determine whether `expr`'s runtime SQL type is a custom/enum type whose
/// OID is assigned by the database rather than fixed — and
/// if so, the Postgres-schema-qualified enum type name for `ShapeNode::Enum`
/// tagging (same convention `pg_quoted_to_pylon` expects). Covers both a
/// column reference to an enum-typed property (quoted pg_type) and a bare
/// enum literal (`default::Gender.Male`) — the two ways an enum value can
/// appear as a free scalar/object/tuple field, which (unlike a schema
/// object's `emit_scalar`) has no dedicated per-pointer descriptor to carry
/// this, so it must be recovered from the expression itself.
fn enum_type_of_expr(expr: &IrExpr) -> Option<QualifiedPgType> {
    match expr {
        IrExpr::ColumnRef { pg_type, .. } => QualifiedPgType::of(pg_type),
        IrExpr::EnumLiteral { pg_type, .. } => Some(QualifiedPgType {
            name: pg_quoted_to_pylon(pg_type),
            is_array: false,
        }),
        _ => None,
    }
}

/// The enum type a shape position decodes as — wider than `enum_type_of_expr`,
/// which only answers for the values that need a `::text` cast emitting.
///
/// `[is Concrete].*` wraps each of the concrete type's properties in a
/// correlated subquery of its own, so the enum sits a level down and the
/// subquery's own emission has already cast it. Without looking through to it
/// the outer shape stays a plain `Scalar` and the label is handed back as a
/// bare string rather than the enum member.
fn enum_type_of_shape_expr(expr: &IrExpr) -> Option<QualifiedPgType> {
    if let Some(qualified) = enum_type_of_expr(expr) {
        return Some(qualified);
    }
    let IrExpr::Subquery(select) = expr else {
        return None;
    };
    let [IrRowSource::Bound { shape, .. }] = select.rows.as_slice() else {
        return None;
    };
    let [IrShapePointer::Scalar(scalar)] = shape.as_slice() else {
        return None;
    };
    QualifiedPgType::of(&scalar.pg_type)
}

/// Emit `expr` for use as a free object/tuple field value — casts to
/// `::text` when it's enum-typed (see `enum_type_of_expr`), which keeps the
/// value's shape independent of the enum's database-assigned type OID.
/// (`pylon-pgcon` also discovers enum OIDs at connect time now, so this cast
/// is belt-and-braces rather than load-bearing for decoding.)
fn emit_free_field_expr(expr: &IrExpr) -> String {
    match enum_type_of_expr(expr) {
        Some(qualified) => format!("{}{}", emit_expr(expr), qualified.text_cast()),
        None => emit_expr(expr),
    }
}

/// Shape node for one free scalar/object/tuple field — `Enum` when the
/// value is enum-typed (see `enum_type_of_expr`), else a plain `Scalar`.
fn free_field_shape_node(name: &str, position: usize, expr: &IrExpr) -> crate::query::ShapeNode {
    use crate::query::{Cardinality, ShapeNode};
    if let IrExpr::ObjectSubquery(sel) = expr {
        let [IrRowSource::Bound { source, shape }] = sel.rows.as_slice() else {
            unreachable!("an object subquery is always schema-bound")
        };
        let (_, nodes) = build_shape(shape, &source.alias);
        return ShapeNode::Object {
            name: name.to_string(),
            type_name: Some(source.type_name.clone()),
            position,
            cardinality: Cardinality::Optional,
            pointers: prepend_type(nodes),
            has_implicit_id: shape_has_implicit_id(shape),
        };
    }
    // A walk landing on one object: `(select … limit 1).latest_data { … }`.
    if matches!(expr, IrExpr::ObjectPathSubquery(_) | IrExpr::ObjectPathUnion { .. }) {
        return expr_shape_node(name, position, expr);
    }
    // A set of objects keeps its rows' own shape, as it does on a schema
    // shape; described as a scalar, each row hydrates as a bare tuple.
    if let IrExpr::ArrayFromSelect(source) = expr
        && matches!(
            source.as_ref(),
            IrArraySource::ObjectSelect(_)
                | IrArraySource::ObjectFunction(_)
                | IrArraySource::PathSelect(_)
                | IrArraySource::Group(_)
        )
    {
        return expr_shape_node(name, position, expr);
    }
    match enum_type_of_shape_expr(expr) {
        Some(qualified) => qualified.shape_node(name.to_string(), position),
        None => ShapeNode::Scalar {
            name: name.to_string(),
            position,
        },
    }
}

/// Shape node for an arbitrary compiled expression appearing at a shape
/// position (a computed pointer, or a free scalar/tuple/object field) —
/// `NamedTuple`/`JsonScalar` when the expression produces jsonb (tuple
/// literals, nested free objects, tuple-typed casts), `RawScalar` when
/// it is returned as a bare top-level column, `Enum` when
/// enum-typed, else a plain `Scalar`.
fn expr_shape_node(name: &str, position: usize, expr: &IrExpr) -> crate::query::ShapeNode {
    use crate::query::{Cardinality, ShapeNode};
    match expr {
        IrExpr::ObjectPathUnion { branches, multi, .. } => {
            let first = branches.first().expect("a union has at least one branch");
            let IrPathResult::Object {
                alias,
                type_name,
                shape,
            } = &first.result
            else {
                unreachable!("an object path union's branches always land on objects")
            };
            let (_, pointer_nodes) = build_shape(shape, alias);
            let object = ShapeNode::Object {
                name: if *multi { String::new() } else { name.to_string() },
                type_name: Some(type_name.clone()),
                position: if *multi { 0 } else { position },
                cardinality: if *multi {
                    Cardinality::Many
                } else {
                    Cardinality::Optional
                },
                pointers: prepend_type(pointer_nodes),
                has_implicit_id: shape_has_implicit_id(shape),
            };
            if *multi {
                ShapeNode::Array {
                    name: name.to_string(),
                    position,
                    element: Box::new(object),
                }
            } else {
                object
            }
        }
        IrExpr::ObjectPathSubquery(ps) => {
            let IrPathResult::Object {
                alias,
                type_name,
                shape,
            } = &ps.result
            else {
                unreachable!("an object path subquery always lands on an object")
            };
            let (_, pointer_nodes) = build_shape(shape, alias);
            ShapeNode::Object {
                name: name.to_string(),
                type_name: Some(type_name.clone()),
                position,
                cardinality: Cardinality::Optional,
                pointers: prepend_type(pointer_nodes),
                has_implicit_id: shape_has_implicit_id(shape),
            }
        }
        IrExpr::TypeCast(c) if c.tuple_shape.is_some() => {
            let shape = c.tuple_shape.as_ref().unwrap();
            ShapeNode::NamedTuple {
                name: name.to_string(),
                position,
                type_name: shape.type_name.clone(),
                members: Some(shape.members.clone()),
                is_free_object: false,
            }
        }
        // A `<json>` cast at a shape position is an ordinary column of the row
        // tuple, so it needs the pointer's own name and position like any
        // other leaf. `JsonScalar` is the *root-only* form, where the result
        // column itself is the value and there is no pointer to name — it
        // carries neither field, and `free_item_shape` is what selects it.
        IrExpr::TypeCast(c) if c.pg_type == "jsonb" => ShapeNode::Scalar {
            name: name.to_string(),
            position,
        },
        IrExpr::NamedTuple { is_free_object, .. } => ShapeNode::NamedTuple {
            name: name.to_string(),
            position,
            type_name: None,
            members: None,
            is_free_object: *is_free_object,
        },
        // A path select aggregated into an array keeps its rows' own shape, so
        // the elements hydrate as objects rather than as opaque scalars.
        IrExpr::ArrayFromSelect(source) => match source.as_ref() {
            IrArraySource::ObjectFunction(fs) => {
                let (_, nodes) = build_shape(&fs.shape, &fs.alias);
                ShapeNode::Array {
                    name: name.to_string(),
                    position,
                    element: Box::new(ShapeNode::Object {
                        name: String::new(),
                        type_name: Some(fs.type_name.clone()),
                        position: 0,
                        cardinality: Cardinality::Many,
                        pointers: prepend_type(nodes),
                        has_implicit_id: shape_has_implicit_id(&fs.shape),
                    }),
                }
            }
            IrArraySource::Group(grp) => ShapeNode::Array {
                name: name.to_string(),
                position,
                element: Box::new(emit_group(grp).shape.root),
            },
            IrArraySource::ObjectSelect(sel) => {
                let [IrRowSource::Bound { source, shape }] = sel.rows.as_slice() else {
                    unreachable!("IrArraySource::ObjectSelect is always schema-bound")
                };
                let (_, nodes) = build_shape(shape, &source.alias);
                ShapeNode::Array {
                    name: name.to_string(),
                    position,
                    element: Box::new(ShapeNode::Object {
                        name: String::new(),
                        type_name: Some(source.type_name.clone()),
                        position: 0,
                        cardinality: Cardinality::Many,
                        pointers: prepend_type(nodes),
                        has_implicit_id: shape_has_implicit_id(shape),
                    }),
                }
            }
            IrArraySource::PathSelect(ps) => match &ps.result {
                IrPathResult::Object {
                    alias,
                    type_name,
                    shape,
                } => {
                    let (_, pointer_nodes) = build_shape(shape, alias);
                    ShapeNode::Array {
                        name: name.to_string(),
                        position,
                        element: Box::new(ShapeNode::Object {
                            name: String::new(),
                            type_name: Some(type_name.clone()),
                            position: 0,
                            cardinality: Cardinality::Many,
                            pointers: prepend_type(pointer_nodes),
                            has_implicit_id: shape_has_implicit_id(shape),
                        }),
                    }
                }
                IrPathResult::Scalar(..) => ShapeNode::Scalar {
                    name: name.to_string(),
                    position,
                },
            },
            _ => ShapeNode::Scalar {
                name: name.to_string(),
                position,
            },
        },
        e if is_raw_scalar(e) => ShapeNode::RawScalar,
        e => free_field_shape_node(name, position, e),
    }
}

fn free_item_shape(item: &IrFreeExpr, ctes: &[IrCteDef]) -> crate::query::ShapeNode {
    use crate::query::{Cardinality, ShapeNode};
    match item {
        IrFreeExpr::Scalar(e) => match e {
            // Top level: the result column *is* the jsonb value rather than a
            // field inside a row tuple. A tuple-shaped cast is excluded because
            // `expr_shape_node` decodes that as a `NamedTuple`.
            IrExpr::TypeCast(c) if c.pg_type == "jsonb" && c.tuple_shape.is_none() => ShapeNode::JsonScalar,
            _ => expr_shape_node("", 0, e),
        },
        IrFreeExpr::FreeObject(fields) => ShapeNode::Object {
            name: String::new(),
            type_name: None,
            position: 0,
            cardinality: Cardinality::Many,
            pointers: fields
                .iter()
                .enumerate()
                .map(|(i, (name, e))| free_field_shape_node(name, i, e))
                .collect(),
            has_implicit_id: false,
        },
        IrFreeExpr::Tuple(exprs) => ShapeNode::Tuple {
            position: 0,
            elements: exprs
                .iter()
                .enumerate()
                .map(|(i, e)| free_field_shape_node("", i, e))
                .collect(),
            names: None,
        },
        IrFreeExpr::NamedTupleRow(fields) => ShapeNode::Tuple {
            position: 0,
            elements: fields
                .iter()
                .enumerate()
                .map(|(i, (name, e))| free_field_shape_node(name, i, e))
                .collect(),
            names: Some(fields.iter().map(|(name, _)| name.clone()).collect()),
        },
        IrFreeExpr::AssertSet { .. } => ShapeNode::Scalar {
            name: String::new(),
            position: 0,
        },
        // The referenced CTE's own `result` column carries whatever shape its
        // defining free row has (a bare scalar, a multi-field free object, a
        // tuple, or even another passthrough) — resolve it by looking the CTE
        // up rather than assuming Scalar, otherwise a multi-field free object
        // bound in a `WITH` clause decodes as just its first field's value.
        IrFreeExpr::CtePassthrough(name) => ctes
            .iter()
            .find(|c| &c.name == name)
            .and_then(|c| match &c.stmt {
                IrStmt::Select(sel) => match sel.rows.first() {
                    Some(IrRowSource::Free(inner)) => Some(free_item_shape(inner, ctes)),
                    _ => None,
                },
                _ => None,
            })
            .unwrap_or(ShapeNode::Scalar {
                name: String::new(),
                position: 0,
            }),
    }
}

// ── PATH SELECT ─────────────────────────────────────────────────────────────

fn emit_path_joins(root: &IrSource, joins: &[IrPathJoin]) -> String {
    let mut parts = vec![format!("{} AS {}", source_ref(root), qi(&root.alias))];
    for join in joins {
        match join {
            IrPathJoin::Single {
                source_alias,
                fk_col,
                target,
            } => {
                parts.push(format!(
                    "JOIN {} AS {} ON {}.{} = {}.\"id\"",
                    source_ref(target),
                    qi(&target.alias),
                    qi(source_alias),
                    qi(fk_col),
                    qi(&target.alias),
                ));
            }
            IrPathJoin::Multi {
                source_alias,
                junction_alias,
                join,
                target,
            } => {
                match join {
                    IrMultiLinkJoin::Standard { junction_table, module } => {
                        parts.push(format!(
                            "JOIN {} AS {} ON {}.\"source\" = {}.\"id\"",
                            junction_ref(module, junction_table),
                            qi(junction_alias),
                            qi(junction_alias),
                            qi(source_alias),
                        ));
                        parts.push(format!(
                            "JOIN {} AS {} ON {}.\"id\" = {}.\"target\"",
                            source_ref(target),
                            qi(&target.alias),
                            qi(&target.alias),
                            qi(junction_alias),
                        ));
                    }
                    IrMultiLinkJoin::Through {
                        junction_table,
                        module,
                        source_col,
                        target_col,
                    } => {
                        parts.push(format!(
                            "JOIN {} AS {} ON {}.{} = {}.\"id\"",
                            junction_ref(module, junction_table),
                            qi(junction_alias),
                            qi(junction_alias),
                            qi(source_col),
                            qi(source_alias),
                        ));
                        parts.push(format!(
                            "JOIN {} AS {} ON {}.\"id\" = {}.{}",
                            source_ref(target),
                            qi(&target.alias),
                            qi(&target.alias),
                            qi(junction_alias),
                            qi(target_col),
                        ));
                    }
                    // See the identical comment in `emit_path_join_sql` above.
                    IrMultiLinkJoin::BacklinkFk { .. } | IrMultiLinkJoin::BacklinkJunction { .. } => {
                        unreachable!("a forward multi-link path step never uses a backlink join variant")
                    }
                }
            }
            IrPathJoin::BacklinkSingle {
                source_alias,
                fk_col,
                target,
            } => {
                parts.push(format!(
                    "JOIN {} AS {} ON {}.{} = {}.\"id\"",
                    source_ref(target),
                    qi(&target.alias),
                    qi(&target.alias),
                    qi(fk_col),
                    qi(source_alias),
                ));
            }
            // LATERAL, not a plain join: the arguments read from the alias
            // the traversal has reached, which a non-lateral function call
            // in FROM cannot see.
            IrPathJoin::Function {
                fn_module,
                fn_name,
                args,
                target,
            } => {
                let args_sql = args.iter().map(emit_expr).collect::<Vec<_>>().join(", ");
                parts.push(format!(
                    "CROSS JOIN LATERAL {}.{}({}) AS {}",
                    pg_schema(fn_module),
                    qi(fn_name),
                    args_sql,
                    qi(&target.alias),
                ));
            }
            IrPathJoin::Lateral { inner, target } => {
                let projected = match &inner.result {
                    IrPathResult::Object { alias, .. } => alias.clone(),
                    IrPathResult::Scalar(..) => {
                        unreachable!("a lateral step is only built for an object-valued computed")
                    }
                };
                let mut body = format!(
                    "SELECT {}.* FROM {}",
                    qi(&projected),
                    emit_path_joins(&inner.root, &inner.joins)
                );
                append_filter(&mut body, &inner.filter);
                append_order_by(&mut body, &inner.order_by);
                append_offset_limit(&mut body, &inner.offset, &inner.limit);
                parts.push(format!("JOIN LATERAL (\n{}\n) AS {} ON TRUE", body, qi(&target.alias)));
            }
            IrPathJoin::BacklinkMulti {
                source_alias,
                junction_alias,
                junction_table,
                module,
                owner_col,
                current_col,
                target,
            } => {
                parts.push(format!(
                    "JOIN {} AS {} ON {}.{} = {}.\"id\"",
                    junction_ref(module, junction_table),
                    qi(junction_alias),
                    qi(junction_alias),
                    qi(current_col),
                    qi(source_alias),
                ));
                parts.push(format!(
                    "JOIN {} AS {} ON {}.\"id\" = {}.{}",
                    source_ref(target),
                    qi(&target.alias),
                    qi(&target.alias),
                    qi(junction_alias),
                    qi(owner_col),
                ));
            }
        }
    }
    parts.join("\n")
}

/// Emit `ARRAY(SELECT scalar FROM source [JOINs] [WHERE filter])`.
fn emit_array_source(src: &IrArraySource) -> String {
    match src {
        IrArraySource::Select(s) => {
            // compile_subquery_to_array_source only ever constructs this
            // variant for a single schema-bound row.
            let [IrRowSource::Bound { source, shape }] = s.rows.as_slice() else {
                unreachable!("IrArraySource::Select is always schema-bound")
            };
            let scalar = match shape.first() {
                Some(IrShapePointer::Scalar(sf)) => format!("{}.{}", qi(&source.alias), qi(&sf.column)),
                _ => format!("{}.\"id\"", qi(&source.alias)),
            };
            let mut sql = format!("SELECT {} FROM {} AS {}", scalar, source_ref(source), qi(&source.alias));
            append_filter(&mut sql, &s.filter);
            format!("ARRAY({})", sql)
        }
        IrArraySource::ObjectFunction(fs) => {
            let (exprs, _) = build_shape(&fs.shape, &fs.alias);
            let mut parts = vec![sql_str(&fs.type_name) + "::text"];
            parts.extend(exprs);
            let args_sql = fs.fn_args.iter().map(emit_expr).collect::<Vec<_>>().join(", ");
            let mut sql = format!(
                "SELECT (\n    {}\n) FROM {}.{}({}) AS {}",
                parts.join(",\n    "),
                pg_schema(&fs.fn_module),
                qi(&fs.fn_name),
                args_sql,
                qi(&fs.alias),
            );
            append_filter(&mut sql, &fs.filter);
            append_order_by(&mut sql, &fs.order_by);
            append_offset_limit(&mut sql, &fs.offset, &fs.limit);
            format!("ARRAY({})", sql)
        }
        IrArraySource::ObjectSelect(s) => {
            let [IrRowSource::Bound { source, shape }] = s.rows.as_slice() else {
                unreachable!("IrArraySource::ObjectSelect is always schema-bound")
            };
            let (exprs, _) = build_shape(shape, &source.alias);
            let mut parts = vec![source_type_disc(source)];
            parts.extend(exprs);
            let mut sql = format!(
                "SELECT (\n    {}\n) FROM {} AS {}",
                parts.join(",\n    "),
                source_ref(source),
                qi(&source.alias)
            );
            append_filter(&mut sql, &s.filter);
            append_order_by(&mut sql, &s.order_by);
            append_offset_limit(&mut sql, &s.offset, &s.limit);
            format!("ARRAY({})", sql)
        }
        IrArraySource::Group(grp) => format!("ARRAY({})", emit_group(grp).sql),
        IrArraySource::StmtColumn { stmt, column } => format!(
            "ARRAY(SELECT {} FROM (\n{}\n) AS \"_rows\")",
            qi(column),
            emit_dml_as_cte_source(stmt),
        ),
        IrArraySource::PathSelect(ps) => {
            let scalar = match &ps.result {
                IrPathResult::Scalar(e, _) => emit_expr(e),
                // Aggregating objects keeps the whole row, not the id: this is
                // how a computed pointer that walks through a multi-link comes
                // back as hydrated rows rather than a list of uuids.
                IrPathResult::Object {
                    alias,
                    type_name,
                    shape,
                } => {
                    let (pointer_exprs, _) = build_shape(shape, alias);
                    let mut parts = vec![result_type_disc(ps, alias, type_name)];
                    parts.extend(pointer_exprs);
                    format!("(\n    {}\n)", parts.join(",\n    "))
                }
            };
            let from_sql = emit_path_joins(&ps.root, &ps.joins);
            let mut sql = format!(
                "SELECT {}{} FROM {}",
                if ps.distinct { "DISTINCT " } else { "" },
                scalar,
                from_sql
            );
            append_filter(&mut sql, &ps.filter);
            append_order_by(&mut sql, &ps.order_by);
            append_offset_limit(&mut sql, &ps.offset, &ps.limit);
            format!("ARRAY({})", sql)
        }
        IrArraySource::RawExpr {
            source,
            poly_implementors,
            poly_columns,
            expr,
        } => {
            let from_sql = if !poly_implementors.is_empty() {
                format!(
                    "(\n{}\n) AS {}",
                    emit_poly_union(poly_implementors, poly_columns),
                    qi(&source.alias)
                )
            } else {
                format!("{} AS {}", source_ref(source), qi(&source.alias))
            };
            format!("ARRAY(SELECT {} FROM {})", emit_expr(expr), from_sql)
        }
    }
}

/// Emit a group-key expression, casting schema-qualified enum ColumnRefs to `::text`
/// so they decode outside of a typed composite.
fn emit_key_expr(expr: &IrExpr) -> String {
    if let IrExpr::ColumnRef { alias, column, pg_type } = expr
        && let Some(qualified) = QualifiedPgType::of(pg_type)
    {
        let col_ref = if alias.is_empty() {
            qi(column)
        } else {
            format!("{}.{}", qi(alias), qi(column))
        };
        return format!("{}{}", col_ref, qualified.text_cast());
    }
    emit_expr(expr)
}

/// The rows a group reads, as `(FROM, WHERE)`. A per-group OFFSET/LIMIT
/// can't be a plain LIMIT — that would cut whole groups — so the rows are
/// ranked within each key first and only the wanted slice kept.
fn group_rows(grp: &IrGroup) -> (String, Option<String>) {
    let alias = &grp.source.alias;
    if grp.limit.is_none() && grp.offset.is_none() {
        return (
            format!("{} AS {}", source_ref(&grp.source), qi(alias)),
            grp.filter.as_ref().map(emit_expr),
        );
    }
    let ranked = format!(
        "(SELECT {}.*, row_number() OVER (PARTITION BY {}{}) AS \"__rk\"\n    FROM {} AS {}{}) AS {}",
        qi(alias),
        group_by_sql(grp),
        group_element_order(grp),
        source_ref(&grp.source),
        qi(alias),
        grp.filter
            .as_ref()
            .map(|f| format!("\n    WHERE {}", emit_expr(f)))
            .unwrap_or_default(),
        qi(alias),
    );
    let lower = grp.offset.as_ref().map(emit_expr).unwrap_or_else(|| "0".to_string());
    let mut conds = vec![format!("\"__rk\" > {}", lower)];
    if let Some(l) = &grp.limit {
        conds.push(format!("\"__rk\" <= {} + {}", lower, emit_expr(l)));
    }
    (ranked, Some(conds.join(" AND ")))
}

fn group_by_sql(grp: &IrGroup) -> String {
    grp.keys
        .iter()
        .map(|(_, key_expr)| emit_expr(key_expr))
        .collect::<Vec<_>>()
        .join(", ")
}

fn group_element_order(grp: &IrGroup) -> String {
    if grp.order_by.is_empty() {
        return String::new();
    }
    let s: Vec<_> = grp.order_by.iter().map(emit_sort_clause).collect();
    format!(" ORDER BY {}", s.join(", "))
}

/// A `for` over a group's elements used as a CTE body: the rows themselves,
/// so the outer query can traverse them like any object binding.
fn emit_group_elements_source(grp: &IrGroup) -> String {
    let (from_sql, cond) = group_rows(grp);
    let mut sql = format!("SELECT {}.*\n    FROM {}", qi(&grp.source.alias), from_sql);
    if let Some(cond) = cond {
        sql.push_str(&format!("\n    WHERE {}", cond));
    }
    sql
}

fn emit_group(grp: &IrGroup) -> SqlOutput {
    match &grp.output {
        IrGroupOutput::Groups => emit_group_rows(grp),
        IrGroupOutput::Projection(projection) => emit_group_projection(
            grp,
            &projection.pointers,
            &projection.order_by,
            &projection.offset,
            &projection.limit,
        ),
        IrGroupOutput::Elements => emit_group_elements(grp),
    }
}

fn emit_group_projection(
    grp: &IrGroup,
    pointers: &[IrShapePointer],
    order_by: &[IrSort],
    offset: &Option<IrExpr>,
    limit: &Option<IrExpr>,
) -> SqlOutput {
    let (exprs, nodes) = build_shape(pointers, &grp.source.alias);
    let mut parts = vec!["NULL::text".to_string()];
    parts.extend(exprs);
    let (from_sql, cond) = group_rows(grp);
    let mut sql = format!(
        "SELECT (\n    {}\n) AS \"result\"\nFROM {}",
        parts.join(",\n    "),
        from_sql
    );
    if let Some(cond) = cond {
        sql.push_str(&format!("\nWHERE {}", cond));
    }
    sql.push_str(&format!("\nGROUP BY {}", group_by_sql(grp)));
    if !order_by.is_empty() {
        let sorts: Vec<_> = order_by.iter().map(emit_sort_clause).collect();
        sql.push_str(&format!("\nORDER BY {}", sorts.join(", ")));
    }
    append_offset_limit(&mut sql, offset, limit);
    SqlOutput {
        sql,
        shape: ShapeDescriptor {
            root: ShapeNode::Object {
                name: String::new(),
                type_name: None,
                position: 0,
                cardinality: Cardinality::Many,
                pointers: nodes,
                has_implicit_id: false,
            },
        },
        inference_plan: None,
    }
}

fn emit_group_elements(grp: &IrGroup) -> SqlOutput {
    let (exprs, nodes) = build_shape(&grp.shape, &grp.source.alias);
    let mut parts = vec![source_type_disc(&grp.source)];
    parts.extend(exprs);
    let (from_sql, cond) = group_rows(grp);
    let mut sql = format!(
        "SELECT (\n    {}\n) AS \"result\"\nFROM {}",
        parts.join(",\n    "),
        from_sql
    );
    if let Some(cond) = cond {
        sql.push_str(&format!("\nWHERE {}", cond));
    }
    SqlOutput {
        sql,
        shape: ShapeDescriptor {
            root: ShapeNode::Object {
                name: String::new(),
                type_name: Some(grp.source.type_name.clone()),
                position: 0,
                cardinality: Cardinality::Many,
                pointers: prepend_type(nodes),
                has_implicit_id: shape_has_implicit_id(&grp.shape),
            },
        },
        inference_plan: None,
    }
}

fn emit_group_rows(grp: &IrGroup) -> SqlOutput {
    let alias = &grp.source.alias;
    let (shape_exprs, shape_nodes) = build_shape(&grp.shape, alias);

    // Build the elements ROW: type discriminator at pos 0, then shape pointers.
    let mut elem_row_parts = vec![source_type_disc(&grp.source)];
    elem_row_parts.extend(shape_exprs);
    let elem_row = elem_row_parts.join(",\n            ");

    // Key positions: NULL at pos 0 (type slot), keys at 1..=N, grouping at N+1, elements at N+2.
    let n_keys = grp.keys.len();
    let grouping_pos = n_keys + 1;
    let elements_pos = n_keys + 2;

    // Build key SQL expressions and ShapeNodes.
    let mut key_exprs_sql: Vec<String> = vec![];
    let mut key_nodes: Vec<ShapeNode> = vec![];
    for (i, (key_name, key_expr)) in grp.keys.iter().enumerate() {
        let pos = i + 1;
        if let IrExpr::ColumnRef { pg_type, .. } = key_expr
            && let Some(qualified) = QualifiedPgType::of(pg_type)
        {
            key_exprs_sql.push(emit_key_expr(key_expr));
            key_nodes.push(qualified.shape_node(key_name.clone(), pos));
            continue;
        }
        key_exprs_sql.push(emit_expr(key_expr));
        key_nodes.push(ShapeNode::Scalar {
            name: key_name.clone(),
            position: pos,
        });
    }

    // Build the outer SELECT tuple.
    let mut outer_parts = vec!["NULL::text".to_string()];
    outer_parts.extend(key_exprs_sql.clone());
    let key_names_sql = grp
        .keys
        .iter()
        .map(|(name, _)| format!("'{}'", name))
        .collect::<Vec<_>>()
        .join(", ");
    outer_parts.push(format!("ARRAY[{}]::text[]", key_names_sql));
    outer_parts.push(format!(
        "array_agg(ROW(\n            {}\n        )::record{})",
        elem_row,
        group_element_order(grp)
    ));

    let outer_tuple = outer_parts.join(",\n    ");
    let (from_sql, cond) = group_rows(grp);
    let mut sql = format!("SELECT (\n    {}\n) AS \"result\"\nFROM {}", outer_tuple, from_sql,);
    if let Some(cond) = cond {
        sql.push_str(&format!("\nWHERE {}", cond));
    }
    sql.push_str(&format!("\nGROUP BY {}", group_by_sql(grp)));

    // ShapeNode for each element (Object with the selected pointers).
    let element_node = ShapeNode::Object {
        name: String::new(),
        type_name: Some(grp.source.type_name.clone()),
        position: 0,
        cardinality: Cardinality::Many,
        pointers: prepend_type(shape_nodes),
        has_implicit_id: shape_has_implicit_id(&grp.shape),
    };

    let root = ShapeNode::Group {
        key_nodes,
        grouping_position: grouping_pos,
        elements_position: elements_pos,
        element: Box::new(element_node),
    };

    SqlOutput {
        sql,
        shape: ShapeDescriptor { root },
        inference_plan: None,
    }
}

fn emit_poly_union_type_only(implementors: &[IrPolyImplementor]) -> String {
    implementors
        .iter()
        .map(|imp| {
            format!(
                "    SELECT {}::text AS \"__type__\" FROM {}",
                sql_str(&imp.type_name),
                qn(&imp.module, &imp.table),
            )
        })
        .collect::<Vec<_>>()
        .join("\n    UNION ALL\n")
}

fn emit_path_select(sel: &IrPathSelect) -> SqlOutput {
    let distinct = if sel.distinct { "DISTINCT " } else { "" };
    let from_sql = if !sel.poly_implementors.is_empty() {
        format!(
            "(\n{}\n) AS {}",
            emit_poly_union_type_only(&sel.poly_implementors),
            qi(&sel.root.alias),
        )
    } else {
        emit_path_joins(&sel.root, &sel.joins)
    };

    let (result_expr, shape_root) = match &sel.result {
        IrPathResult::Scalar(ir_expr, tuple_shape) => {
            // Named tuples / jsonb field accesses can't be decoded inside ROW() — emit raw.
            let is_nt = matches!(ir_expr, IrExpr::NamedTuple { .. })
                || matches!(ir_expr, IrExpr::Tuple(_))
                || matches!(ir_expr, IrExpr::JsonbField { .. })
                || matches!(ir_expr, IrExpr::JsonbIndex { .. })
                || matches!(ir_expr, IrExpr::ColumnRef { pg_type, .. } if pg_type.starts_with("__nt__:"))
                || tuple_shape.is_some();
            if is_nt {
                let expr_sql = format!("{} AS result", emit_expr(ir_expr));
                let shape = if matches!(
                    ir_expr,
                    IrExpr::JsonbField { .. } | IrExpr::JsonbIndex { .. } | IrExpr::Tuple(_)
                ) {
                    ShapeNode::RawScalar
                } else if let Some(shape) = tuple_shape {
                    // A bare tuple-typed property reference (nominal or
                    // structural) — real member shape already resolved at
                    // compile time (see resolve_property_tuple_shape), same
                    // as a `Type { tuple_property }` shape query gets via
                    // emit_scalar, instead of falling back to `members: None`.
                    ShapeNode::NamedTuple {
                        name: String::new(),
                        position: 0,
                        type_name: shape.type_name.clone(),
                        members: Some(shape.members.clone()),
                        is_free_object: false,
                    }
                } else {
                    let type_name = match ir_expr {
                        IrExpr::ColumnRef { pg_type, .. } => pg_type.strip_prefix("__nt__:").map(|s| s.to_string()),
                        _ => None,
                    };
                    ShapeNode::NamedTuple {
                        name: String::new(),
                        position: 0,
                        type_name,
                        members: None,
                        is_free_object: false,
                    }
                };
                (expr_sql, shape)
            } else {
                // Schema-qualified types (enums, domains) have unknown OIDs inside ROW() —
                // cast to text so the value's shape doesn't depend on a runtime OID.
                if let IrExpr::ColumnRef { pg_type, .. } = ir_expr {
                    if let Some(qualified) = QualifiedPgType::of(pg_type) {
                        let expr = format!("ROW({}{}) AS result", emit_expr(ir_expr), qualified.text_cast());
                        let shape = qualified.shape_node(String::new(), 0);
                        (expr, shape)
                    } else {
                        let expr = format!("ROW({}) AS result", emit_expr(ir_expr));
                        (
                            expr,
                            ShapeNode::Scalar {
                                name: String::new(),
                                position: 0,
                            },
                        )
                    }
                } else {
                    let expr = format!("ROW({}) AS result", emit_expr(ir_expr));
                    (
                        expr,
                        ShapeNode::Scalar {
                            name: String::new(),
                            position: 0,
                        },
                    )
                }
            }
        }
        IrPathResult::Object {
            alias,
            type_name,
            shape,
        } => {
            let (pointer_exprs, pointer_nodes) = build_shape(shape, alias);
            let mut parts = vec![result_type_disc(sel, alias, type_name)];
            parts.extend(pointer_exprs);
            let expr = format!("(\n    {}\n) AS result", parts.join(",\n    "));
            let shape_root = ShapeNode::Object {
                name: String::new(),
                type_name: Some(type_name.clone()),
                position: 0,
                cardinality: Cardinality::Many,
                pointers: prepend_type(pointer_nodes),
                has_implicit_id: shape_has_implicit_id(shape),
            };
            (expr, shape_root)
        }
    };

    let mut sql = format!("SELECT {}{}\nFROM {}", distinct, result_expr, from_sql);
    append_filter(&mut sql, &sel.filter);
    append_order_by(&mut sql, &sel.order_by);
    append_offset_limit(&mut sql, &sel.offset, &sel.limit);

    SqlOutput {
        sql,
        shape: ShapeDescriptor { root: shape_root },
        inference_plan: None,
    }
}

// ── FOR LOOP ─────────────────────────────────────────────────────────────────

/// The loop's iterator as a relation with a single column named `v`, in both
/// the forms the two body kinds need: one for a `FROM` clause, one for a CTE.
/// True for an expression PostgreSQL hands back as `jsonb`, whose value a
/// scalar cast has to reach into rather than render.
fn yields_jsonb(expr: &IrExpr) -> bool {
    match expr {
        IrExpr::JsonbField { .. } | IrExpr::JsonbIndex { .. } | IrExpr::NamedTuple { .. } => true,
        IrExpr::TypeCast(c) => c.pg_type == "jsonb",
        IrExpr::ColumnRef { pg_type, .. } | IrExpr::FnParam { pg_type, .. } => pg_type == "jsonb",
        IrExpr::CteRef { pg_type, .. } => pg_type.as_deref() == Some("jsonb"),
        IrExpr::FunctionCall(f) if f.schema.is_none() => {
            let mut overloads = crate::stdlib::registry().iter().filter(|d| d.name == f.name).peekable();
            overloads.peek().is_some() && overloads.all(|d| matches!(d.return_type, crate::stdlib::PylonType::Json))
        }
        _ => false,
    }
}

fn emit_for_iterator(it: &IrForIterator, iter_alias: &str) -> (String, String) {
    match it {
        IrForIterator::Values { exprs, pg_type } => {
            let rows: Vec<String> = exprs
                .iter()
                .map(|e| format!("({}::{})", emit_expr(e), pg_type))
                .collect();
            (
                format!("(VALUES {}) AS {}(\"v\")", rows.join(", "), qi(iter_alias)),
                format!("{}(\"v\") AS (VALUES {})", qi(iter_alias), rows.join(", ")),
            )
        }
        IrForIterator::Query { stmt, scalar } => {
            let inner = emit_dml_as_cte_source(stmt);
            // A scalar-yielding select already exposes `v` (the same
            // convention a scalar `with` binding uses); an object-yielding
            // one exposes its raw columns, so the loop variable binds `id`.
            let body = if *scalar {
                inner
            } else {
                format!("    SELECT \"id\" AS v FROM (\n{}\n    ) AS _src", inner)
            };
            (
                format!("(\n{}\n) AS {}", body, qi(iter_alias)),
                format!("{} AS (\n{}\n)", qi(iter_alias), body),
            )
        }
        IrForIterator::SetReturning { expr, pg_type } => {
            let body = format!("    SELECT {}::{} AS v", emit_expr(expr), pg_type);
            (
                format!("(\n{}\n) AS {}", body, qi(iter_alias)),
                format!("{} AS (\n{}\n)", qi(iter_alias), body),
            )
        }
    }
}

/// The iterator CTE of a `for` nested in another one. The enclosing iterator
/// goes in its FROM, so the walk it iterates can read the outer loop variable,
/// and its key rides along as `_outer` for the body to join back on — the bond
/// Gel's own `merge_iterator` puts between a chain of iterator CTEs.
fn emit_nested_for_iterator(it: &IrForIterator, iter_alias: &str, outer_alias: &str) -> String {
    const OUTER: &str = "_outer";
    match it {
        IrForIterator::Values { exprs, pg_type } => {
            let rows: Vec<String> = exprs
                .iter()
                .map(|e| format!("({}::{})", emit_expr(e), pg_type))
                .collect();
            format!(
                "{} AS (\nSELECT {}.\"v\" AS {}, \"_vals\".\"v\" AS v\nFROM {}, (VALUES {}) AS \"_vals\"(\"v\")\n)",
                qi(iter_alias),
                qi(outer_alias),
                qi(OUTER),
                qi(outer_alias),
                rows.join(", "),
            )
        }
        IrForIterator::Query { stmt, scalar } => {
            let inner = emit_dml_as_cte_source(stmt);
            let projected = if *scalar { "\"_src\".\"v\"" } else { "\"_src\".\"id\"" };
            format!(
                "{} AS (\nSELECT {}.\"v\" AS {}, {} AS v\nFROM {}\nCROSS JOIN LATERAL (\n{}\n) AS \"_src\"\n)",
                qi(iter_alias),
                qi(outer_alias),
                qi(OUTER),
                projected,
                qi(outer_alias),
                inner,
            )
        }
        IrForIterator::SetReturning { expr, pg_type } => format!(
            "{} AS (\nSELECT {}.\"v\" AS {}, \"_vals\".\"v\" AS v\nFROM {}, LATERAL (SELECT {}::{} AS v) AS \"_vals\"\n)",
            qi(iter_alias),
            qi(outer_alias),
            qi(OUTER),
            qi(outer_alias),
            emit_expr(expr),
            pg_type,
        ),
    }
}

/// `FROM <inner iterator> JOIN <outer iterator> ON …` — both loop variables in
/// scope for the body, each still read as `"_for_<name>"."v"`.
fn nested_for_from(inner_alias: &str, outer_alias: &str) -> String {
    format!(
        "{} JOIN {} ON {}.\"v\" = {}.\"_outer\"",
        qi(inner_alias),
        qi(outer_alias),
        qi(outer_alias),
        qi(inner_alias),
    )
}

/// Put the bindings a `for` body declared in front of the body itself: they may
/// read the loop variable, so they cannot sit in the enclosing WITH clause.
/// A `for` body's CTEs as they must read inside the LATERAL that runs the body.
///
/// `emit_user_cte_parts` keys a binding that reads the loop variable by the
/// iteration, joining `_for_<slot>` to carry its key — right where the binding
/// sits in the statement's own WITH beside the iterator CTE, as the DML forms
/// emit it. Inside the LATERAL there is no such relation: `_for_<slot>` is the
/// outer FROM item's alias, so `FROM "_for_<slot>"` is SQL PostgreSQL rejects
/// ("relation does not exist"), and spelling it legally would widen the binding
/// to every iteration's rows at once. The LATERAL has already pinned the
/// iteration to one row, and the correlated column reference the body carries
/// reads it, so the key is neither needed nor correct here.
fn uncorrelated_body_ctes(f: &IrFor) -> Vec<IrCteDef> {
    f.body_ctes
        .iter()
        .map(|cte| IrCteDef {
            correlated_to: None,
            ..cte.clone()
        })
        .collect()
}

fn prefix_body_ctes(sql: String, body_ctes: &[IrCteDef]) -> String {
    if body_ctes.is_empty() {
        return sql;
    }
    merge_into_existing_with(&sql, &emit_user_cte_parts(body_ctes))
        .unwrap_or_else(|| format!("{}{}", emit_cte_prefix(body_ctes), sql))
}

fn emit_for_stmt(f: &IrFor, user_ctes: &[IrCteDef]) -> SqlOutput {
    let iter_alias = format!("_for_{}", f.var_name);

    if matches!(&f.iterator, IrForIterator::Values { exprs, .. } if exprs.is_empty()) {
        let empty = SqlOutput {
            sql: "SELECT NULL AS result WHERE FALSE".to_string(),
            shape: ShapeDescriptor {
                root: ShapeNode::Scalar {
                    name: String::new(),
                    position: 0,
                },
            },
            inference_plan: None,
        };
        return empty;
    }

    let (values_from, iter_cte) = emit_for_iterator(&f.iterator, &iter_alias);

    match f.body.as_ref() {
        IrStmt::Insert(ins) => emit_for_insert(ins, &iter_alias, &iter_cte, user_ctes, &f.body_ctes),
        IrStmt::Update(upd) => emit_for_update(upd, &iter_alias, &iter_cte, user_ctes, &f.body_ctes),
        body => {
            let body_out = match body {
                IrStmt::Select(sel) => emit_select_stmt(sel, user_ctes),
                IrStmt::PathSelect(sel) => emit_path_select(sel),
                // Unreachable: `Compiler::compile_for` rejects every other
                // body kind with a PyQL error before an `IrFor` is built.
                other => unreachable!("for-loop body should have been rejected at compile time: {other:?}"),
            };
            let indent_body = prefix_body_ctes(body_out.sql, &uncorrelated_body_ctes(f)).replace('\n', "\n    ");
            let cte_prefix = if !user_ctes.is_empty() {
                emit_cte_prefix(user_ctes)
            } else {
                String::new()
            };
            let sql = format!(
                "{}SELECT \"_body\".result\nFROM {}\nCROSS JOIN LATERAL (\n    {}\n) AS \"_body\"",
                cte_prefix, values_from, indent_body,
            );
            SqlOutput {
                sql,
                shape: body_out.shape,
                inference_plan: None,
            }
        }
    }
}

/// A `for` whose body updates: driven from the iteration with `UPDATE … FROM`,
/// since a LATERAL -- what the select-bodied form uses -- cannot hold DML.
/// Both the SET clause and the filter may read the loop variable, and both are
/// already compiled against the iteration alias.
fn emit_for_update(
    upd: &IrUpdate,
    iter_alias: &str,
    iter_cte: &str,
    user_ctes: &[IrCteDef],
    body_ctes: &[IrCteDef],
) -> SqlOutput {
    let mut cte_parts: Vec<String> = emit_user_cte_parts(user_ctes);
    cte_parts.extend(emit_user_cte_parts(body_ctes));
    cte_parts.push(iter_cte.to_string());

    let alias = &upd.target.alias;
    // With only junction rows to write there is nothing to SET, so the driving
    // rows are selected rather than updated -- the same shape a multi-link-only
    // update takes outside a loop.
    if upd.assignments.is_empty() && upd.rewrites.is_empty() && !upd.multi_link_appends.is_empty() {
        const ITER_COL: &str = "_iter";
        let ids_name = "_ids";
        // A nested insert gets its id generated here rather than by the column
        // default, so the junction below can name the row *this* iteration
        // inserted — `RETURNING` offers no order to pair on.
        let generated: Vec<Option<&IrInsert>> = upd
            .multi_link_appends
            .iter()
            .map(|a| per_iteration_insert(a, &upd.nested_ctes))
            .collect();
        let new_cols: String = generated
            .iter()
            .enumerate()
            .filter_map(|(i, ins)| {
                ins.map(|ins| {
                    format!(
                        ", {} AS {}",
                        ins.id_default_sql.as_deref().unwrap_or("uuidv7()"),
                        qi(&format!("_new_{i}"))
                    )
                })
            })
            .collect();
        let mut ids_sql = format!(
            "\"{}\" AS (\nSELECT {}.*, {}.\"v\" AS {}{} FROM {} AS {}, {}",
            ids_name,
            qi(alias),
            qi(iter_alias),
            qi(ITER_COL),
            new_cols,
            target_ref(&upd.target),
            qi(alias),
            qi(iter_alias),
        );
        append_filter(&mut ids_sql, &upd.filter);
        ids_sql.push_str("\n)");
        cte_parts.push(ids_sql);
        for (i, append) in upd.multi_link_appends.iter().enumerate() {
            match generated[i] {
                Some(ins) => {
                    cte_parts.push(emit_for_nested_insert_cte(
                        ins,
                        ids_name,
                        iter_alias,
                        &format!("_ml_ins_{i}"),
                        &format!("_new_{i}"),
                    ));
                    cte_parts.push(emit_for_ml_append_cte(
                        append,
                        ids_name,
                        &format!("_ml_add_{i}"),
                        &format!("_new_{i}"),
                    ));
                }
                None => cte_parts.push(emit_for_ml_append_cte(
                    append,
                    ids_name,
                    &format!("_ml_add_{i}"),
                    ITER_COL,
                )),
            }
        }
        let (pointer_exprs, shape_nodes) = build_shape(&upd.returning, ids_name);
        let mut parts = vec![type_disc(&upd.target.type_name)];
        parts.extend(pointer_exprs);
        let sql = format!(
            "WITH {}\nSELECT (\n    {}\n) AS result\nFROM \"{}\"",
            cte_parts.join(",\n"),
            parts.join(",\n    "),
            ids_name,
        );
        return SqlOutput {
            sql,
            shape: crate::query::ShapeDescriptor {
                root: ShapeNode::Object {
                    name: String::new(),
                    type_name: Some(upd.target.type_name.clone()),
                    position: 0,
                    cardinality: Cardinality::Many,
                    pointers: prepend_type(shape_nodes),
                    has_implicit_id: shape_has_implicit_id(&upd.returning),
                },
            },
            inference_plan: None,
        };
    }
    let sets: Vec<String> = upd
        .assignments
        .iter()
        .map(|(col, expr)| format!("{} = {}", qi(col), emit_expr(expr)))
        .chain(
            upd.rewrites
                .iter()
                .map(|rw| format!("{} = {}", qi(&rw.column), emit_expr(&rw.expr))),
        )
        .collect();
    let mut sql = format!(
        "WITH {}\nUPDATE {} AS {}\nSET {}\nFROM {}",
        cte_parts.join(",\n"),
        target_ref(&upd.target),
        qi(alias),
        sets.join(", "),
        qi(iter_alias),
    );
    append_filter(&mut sql, &upd.filter);
    let mut parts = vec![type_disc(&upd.target.type_name)];
    let (pointer_exprs, shape_nodes) = build_shape(&upd.returning, alias);
    parts.extend(pointer_exprs);
    sql.push_str(&format!("\nRETURNING (\n    {}\n) AS result", parts.join(",\n    ")));
    SqlOutput {
        sql,
        shape: crate::query::ShapeDescriptor {
            root: ShapeNode::Object {
                name: String::new(),
                type_name: Some(upd.target.type_name.clone()),
                position: 0,
                cardinality: Cardinality::Many,
                pointers: prepend_type(shape_nodes),
                has_implicit_id: shape_has_implicit_id(&upd.returning),
            },
        },
        inference_plan: None,
    }
}

/// The iteration key a correlated nested CTE carries, so rows written per
/// iteration can be paired back with the row that iteration wrote. `RETURNING`
/// can only name the inserted row's own columns, so the pairing is computed in
/// a staging CTE beside the insert rather than read back out of it.
const OUTER_KEY: &str = "_outer";

/// `SELECT <id> AS "id", <iter>."v" AS "_outer", <value> AS "<column>", …` —
/// the rows one iteration-driven insert will write, with their ids generated
/// here rather than by the column default so a junction can name them before
/// the insert has run.
fn emit_iteration_rows_cte(ins: &IrInsert, cte_name: &str, iter_alias: &str, outer_source: &str) -> String {
    let rewrite_cols: std::collections::HashSet<&str> = ins.rewrites.iter().map(|r| r.column.as_str()).collect();
    let mut selected = vec![
        format!(
            "{} AS {}",
            ins.id_default_sql.as_deref().unwrap_or("uuidv7()"),
            qi("id")
        ),
        format!("{} AS {}", outer_source, qi(OUTER_KEY)),
    ];
    selected.extend(
        ins.assignments
            .iter()
            .filter(|(c, _)| !rewrite_cols.contains(c.as_str()))
            .map(|(c, e)| format!("{} AS {}", emit_expr(e), qi(c)))
            .chain(
                ins.rewrites
                    .iter()
                    .map(|r| format!("{} AS {}", emit_expr(&r.expr), qi(&r.column))),
            ),
    );
    format!(
        "\"{}\" AS (\nSELECT {}\nFROM {}\n)",
        cte_name,
        selected.join(", "),
        qi(iter_alias),
    )
}

/// The insert that reads its rows back out of the staging CTE above.
fn emit_insert_from_rows_cte(ins: &IrInsert, cte_name: &str, rows_name: &str) -> String {
    let rewrite_cols: std::collections::HashSet<&str> = ins.rewrites.iter().map(|r| r.column.as_str()).collect();
    let cols: Vec<String> = std::iter::once(qi("id"))
        .chain(
            ins.assignments
                .iter()
                .filter(|(c, _)| !rewrite_cols.contains(c.as_str()))
                .map(|(c, _)| qi(c)),
        )
        .chain(ins.rewrites.iter().map(|r| qi(&r.column)))
        .collect();
    let mut sql = format!(
        "INSERT INTO {} ({})\nSELECT {} FROM \"{}\"",
        target_ref(&ins.target),
        cols.join(", "),
        cols.join(", "),
        rows_name,
    );
    if let Some(conflict) = &ins.unless_conflict {
        emit_conflict(&mut sql, conflict);
    }
    sql.push_str("\nRETURNING *");
    format!("\"{}\" AS (\n{}\n)", cte_name, sql)
}

/// A junction append whose targets were written per iteration: pair the two
/// staging CTEs on the iteration key rather than cross-joining every row
/// written against every row that wrote one.
fn emit_correlated_ml_append_cte(
    mutation: &IrMultiLinkMutation,
    rows_name: &str,
    targets_name: &str,
    cte_name: &str,
) -> String {
    format!(
        "\"{}\" AS (\nINSERT INTO {} ({}, {})\nSELECT \"_src\".{}, \"_tgt\".{} FROM \"{}\" AS \"_src\" JOIN \"{}\" AS \"_tgt\" ON \"_tgt\".{} = \"_src\".{}\nON CONFLICT DO NOTHING\nRETURNING {}, {}\n)",
        cte_name,
        qn(&mutation.module, &mutation.junction_table),
        qi(&mutation.source_col),
        qi(&mutation.target_col),
        qi("id"),
        qi("id"),
        rows_name,
        targets_name,
        qi(OUTER_KEY),
        qi(OUTER_KEY),
        qi(&mutation.source_col),
        qi(&mutation.target_col),
    )
}

/// Which of `appends` name a CTE that wrote its rows per iteration, and so
/// pair on the iteration key instead of cross-joining.
fn correlated_append_indices(appends: &[IrMultiLinkMutation], defs: &[&[IrCteDef]]) -> Vec<usize> {
    appends
        .iter()
        .enumerate()
        .filter(|(_, append)| match &append.values.source {
            IrMultiLinkValueSource::CteRef(target) => defs
                .iter()
                .flat_map(|group| group.iter())
                .any(|c| &c.name == target && c.correlated_to.is_some()),
            _ => false,
        })
        .map(|(i, _)| i)
        .collect()
}

fn emit_for_insert(
    ins: &IrInsert,
    iter_alias: &str,
    iter_cte: &str,
    user_ctes: &[IrCteDef],
    body_ctes: &[IrCteDef],
) -> SqlOutput {
    let rewrite_cols: std::collections::HashSet<&str> = ins.rewrites.iter().map(|r| r.column.as_str()).collect();

    let cols: Vec<String> = ins
        .assignments
        .iter()
        .filter(|(c, _)| !rewrite_cols.contains(c.as_str()))
        .map(|(c, _)| qi(c))
        .chain(ins.rewrites.iter().map(|r| qi(&r.column)))
        .collect();
    let sel_exprs: Vec<String> = ins
        .assignments
        .iter()
        .filter(|(c, _)| !rewrite_cols.contains(c.as_str()))
        .map(|(_, e)| emit_expr(e))
        .chain(ins.rewrites.iter().map(|r| emit_expr(&r.expr)))
        .collect();

    let mut cte_parts: Vec<String> = emit_user_cte_parts(user_ctes);
    // The iterator first: a body CTE, and any DML nested in the row's values,
    // may read the loop variable, and a WITH name is only in scope for the
    // bindings that follow it.
    cte_parts.push(iter_cte.to_string());
    cte_parts.extend(emit_user_cte_parts(body_ctes));
    cte_parts.extend(emit_user_cte_parts(&ins.nested_ctes));
    // A nested loop that carried the iteration key back means this insert's
    // own rows need one too, to pair with.
    let correlated_targets = correlated_append_indices(&ins.multi_link_appends, &[body_ctes, &ins.nested_ctes]);

    let mut insert_sql = format!(
        "INSERT INTO {} ({})\nSELECT {} FROM {}",
        target_ref(&ins.target),
        cols.join(", "),
        sel_exprs.join(", "),
        qi(iter_alias),
    );
    if let Some(conflict) = &ins.unless_conflict {
        emit_conflict(&mut insert_sql, conflict);
    }
    let (shape, returning_sql) = emit_returning_shape(&ins.target, &ins.returning, false);

    if ins.multi_link_appends.is_empty() {
        let mut sql = format!("WITH {}\n{}", cte_parts.join(",\n"), insert_sql);
        if let Some(r) = returning_sql {
            sql.push_str(&r);
        }
        return SqlOutput {
            sql,
            shape,
            inference_plan: None,
        };
    }

    // A junction row needs the id of the row this iteration just wrote, so
    // the insert becomes a CTE the junction inserts read back from.
    const IDS: &str = "_for_dml__ids";
    const ROWS: &str = "_for_dml__rows";
    if correlated_targets.is_empty() {
        insert_sql.push_str("\nRETURNING *");
        cte_parts.push(format!("\"{}\" AS (\n{}\n)", IDS, insert_sql));
    } else {
        let outer_source = format!("{}.{}", qi(iter_alias), qi("v"));
        cte_parts.push(emit_iteration_rows_cte(ins, ROWS, iter_alias, &outer_source));
        cte_parts.push(emit_insert_from_rows_cte(ins, IDS, ROWS));
    }
    for (i, append) in ins.multi_link_appends.iter().enumerate() {
        let cte_name = format!("_for_dml__ml_add_{}", i);
        if correlated_targets.contains(&i) {
            let IrMultiLinkValueSource::CteRef(target) = &append.values.source else {
                unreachable!("only a CTE-backed value can carry an iteration key")
            };
            cte_parts.push(emit_correlated_ml_append_cte(append, ROWS, target, &cte_name));
        } else {
            cte_parts.push(emit_ml_append_cte(append, IDS, &ins.target.alias, &cte_name));
        }
    }

    let projection = match &returning_sql {
        // `emit_returning_shape` spells the projection as a RETURNING clause;
        // read off a CTE it is the same tuple in a SELECT list.
        Some(r) => r.trim_start_matches('\n').replacen("RETURNING ", "SELECT ", 1),
        None => "SELECT 1".to_string(),
    };
    SqlOutput {
        sql: format!("WITH {}\n{}\nFROM \"{}\"", cte_parts.join(",\n"), projection, IDS),
        shape,
        inference_plan: None,
    }
}

// ── INSERT ──────────────────────────────────────────────────────────────────

// ── Vector index enqueue helpers ─────────────────────────────────────────────

/// Build one `"_eqN" AS (INSERT INTO _pylon."IndexOutbox" ...)` CTE string.
fn enqueue_cte_sql(eq: &VectorEnqueueInfo, source_cte: &str, cte_name: &str) -> String {
    let index_name_sql = match &eq.index_name {
        None => "NULL".to_string(),
        Some(name) => sql_str(name),
    };
    format!(
        concat!(
            "\"{}\" AS (\n",
            "    INSERT INTO _pylon.\"IndexOutbox\"\n",
            "        (object_id, type_name, index_kind, index_name)\n",
            "    SELECT \"id\", {}, 'Vector'::_pylon.\"IndexKind\", {}\n",
            "    FROM \"{}\"\n",
            "    ON CONFLICT (object_id, index_kind, index_name)\n",
            "    DO UPDATE SET status = 'Pending', enqueued_at = now()\n",
            ")",
        ),
        cte_name,
        sql_str(&eq.type_name),
        index_name_sql,
        source_cte,
    )
}

/// Build the full list of enqueue CTE strings for a set of vector indexes.
fn enqueue_ctes(enqueue: &[VectorEnqueueInfo], source_cte: &str) -> Vec<String> {
    enqueue
        .iter()
        .enumerate()
        .map(|(i, eq)| enqueue_cte_sql(eq, source_cte, &format!("_eq{}", i)))
        .collect()
}

/// The `_pylon."IndexOutbox".index_kind` enum label for a search backend —
/// `Postgres`-backed indexes never reach this (see `collect_search_enqueue`),
/// so there's no corresponding `IndexKind` value for that variant.
fn search_backend_index_kind(backend: &crate::schema::SearchBackend) -> &'static str {
    match backend {
        crate::schema::SearchBackend::OpenSearch => "OpenSearch",
        crate::schema::SearchBackend::Meilisearch => "Meilisearch",
        crate::schema::SearchBackend::Postgres => {
            unreachable!("Postgres-backed search indexes are never collected into SearchEnqueueInfo")
        }
    }
}

/// Build one OpenSearch/Meilisearch outbox CTE string.
fn enqueue_search_cte_sql(eq: &SearchEnqueueInfo, source_cte: &str, cte_name: &str) -> String {
    let index_name_sql = match &eq.index_name {
        None => "NULL".to_string(),
        Some(name) => sql_str(name),
    };
    format!(
        concat!(
            "\"{}\" AS (\n",
            "    INSERT INTO _pylon.\"IndexOutbox\"\n",
            "        (object_id, type_name, index_kind, index_name, operation)\n",
            "    SELECT \"id\", {}, '{}'::_pylon.\"IndexKind\", {}, {}\n",
            "    FROM \"{}\"\n",
            "    ON CONFLICT (object_id, index_kind, index_name)\n",
            "    DO UPDATE SET status = 'Pending', operation = EXCLUDED.operation, enqueued_at = now()\n",
            ")",
        ),
        cte_name,
        sql_str(&eq.type_name),
        search_backend_index_kind(&eq.backend),
        index_name_sql,
        sql_str(eq.operation),
        source_cte,
    )
}

/// Build the full list of OpenSearch/Meilisearch enqueue CTE strings.
fn enqueue_search_ctes(enqueue: &[SearchEnqueueInfo], source_cte: &str, offset: usize) -> Vec<String> {
    enqueue
        .iter()
        .enumerate()
        .map(|(i, eq)| enqueue_search_cte_sql(eq, source_cte, &format!("_es{}", offset + i)))
        .collect()
}

/// Build `SELECT (...) AS result FROM "cte_name"` plus its ShapeDescriptor,
/// mirroring `emit_returning_shape` but for the CTE-wrapper SELECT path.
fn shape_select_from_cte(
    target: &IrSource,
    returning: &[IrShapePointer],
    cte_name: &str,
) -> (ShapeDescriptor, Option<String>) {
    if returning.is_empty() {
        return (
            ShapeDescriptor {
                root: ShapeNode::Scalar {
                    name: String::new(),
                    position: 0,
                },
            },
            None,
        );
    }
    let (pointer_exprs, shape_pointers) = build_shape(returning, "");
    let mut parts = vec![type_disc(&target.type_name)];
    parts.extend(pointer_exprs);
    let tuple = parts.join(",\n    ");
    let sql = format!("SELECT (\n    {}\n) AS result\nFROM {}", tuple, qi(cte_name));
    let root_pointers = prepend_type(shape_pointers);
    let shape = ShapeDescriptor {
        root: ShapeNode::Object {
            name: String::new(),
            type_name: Some(target.type_name.clone()),
            position: 0,
            cardinality: Cardinality::Required,
            pointers: root_pointers,
            has_implicit_id: shape_has_implicit_id(returning),
        },
    };
    (shape, Some(sql))
}

/// Builds the base `INSERT INTO t (cols) VALUES (...)` row source — or,
/// when `ins.nested_ctes` is non-empty, `INSERT INTO t (cols) SELECT ...
/// FROM cte1, cte2, ...` instead, since a VALUES list can't reference a
/// CTE's columns (needed when a link's value is sourced from a hoisted
/// nested INSERT/UPDATE/DELETE — see `IrInsert::nested_ctes`). Each nested
/// CTE is guaranteed exactly one row, so an implicit cross join is safe.
/// Shared by the plain wrap path and the multi-link junction wrap path
/// below, both of which need this same VALUES/SELECT choice.
fn emit_insert_row_sql(ins: &IrInsert) -> String {
    let rewrite_cols: std::collections::HashSet<&str> = ins.rewrites.iter().map(|r| r.column.as_str()).collect();
    let cols: Vec<String> = ins
        .assignments
        .iter()
        .filter(|(c, _)| !rewrite_cols.contains(c.as_str()))
        .map(|(c, _)| qi(c))
        .chain(ins.rewrites.iter().map(|r| qi(&r.column)))
        .collect();
    let vals: Vec<String> = ins
        .assignments
        .iter()
        .filter(|(c, _)| !rewrite_cols.contains(c.as_str()))
        .map(|(_, e)| emit_expr(e))
        .chain(ins.rewrites.iter().map(|r| emit_expr(&r.expr)))
        .collect();
    // A guarded insert writes nothing when its condition is false, so the row
    // comes from a `SELECT` that returns none -- `VALUES` always yields one.
    if let Some(guard) = &ins.guard {
        let from_ctes = nested_cte_from(&ins.nested_ctes, &vals.join(","), " ");
        let projection = if cols.is_empty() {
            String::new()
        } else {
            format!(" {}", vals.join(", "))
        };
        let column_list = if cols.is_empty() {
            String::new()
        } else {
            format!(" ({})", cols.join(", "))
        };
        return format!(
            "INSERT INTO {}{} SELECT{}{} WHERE {}",
            target_ref(&ins.target),
            column_list,
            projection,
            from_ctes,
            emit_expr(guard),
        );
    }
    if cols.is_empty() && ins.nested_ctes.is_empty() {
        // `insert Preferences {}` — a row made entirely of its own defaults.
        // An empty column list is not SQL; `DEFAULT VALUES` is how it is said.
        format!("INSERT INTO {} DEFAULT VALUES", target_ref(&ins.target))
    } else if ins.nested_ctes.is_empty() {
        format!(
            "INSERT INTO {} ({}) VALUES ({})",
            target_ref(&ins.target),
            cols.join(", "),
            vals.join(", ")
        )
    } else {
        format!(
            "INSERT INTO {} ({}) SELECT {}{}",
            target_ref(&ins.target),
            cols.join(", "),
            vals.join(", "),
            nested_cte_from(&ins.nested_ctes, &vals.join(","), " "),
        )
    }
}

fn emit_insert_stmt(ins: &IrInsert) -> SqlOutput {
    if ins.enqueue_vector.is_empty()
        && ins.enqueue_search.is_empty()
        && !insert_has_any_multilink(ins)
        && ins.nested_ctes.is_empty()
    {
        let mut sql = emit_insert_row_sql(ins);
        if let Some(conflict) = &ins.unless_conflict {
            emit_conflict(&mut sql, conflict);
        }
        let (shape, returning_sql) = emit_returning_shape(&ins.target, &ins.returning, false);
        if let Some(r) = returning_sql {
            sql.push_str(&r);
        }
        return SqlOutput {
            sql,
            shape,
            inference_plan: None,
        };
    }

    // Wrap path: needed for outbox enqueue CTEs, junction-table population
    // (see emit_insert_multilink_ctes), and/or hoisted nested-DML CTEs (see
    // IrInsert::nested_ctes) — all of which require either the row's own id
    // or a WITH prefix, so the plain single-statement INSERT above can't be
    // used.
    let mut cte_parts = if insert_has_any_multilink(ins) {
        emit_insert_multilink_ctes(ins, "_w")
    } else {
        let mut cte_parts = emit_user_cte_parts(&ins.nested_ctes);
        let mut insert_sql = emit_insert_row_sql(ins);
        if let Some(conflict) = &ins.unless_conflict {
            emit_conflict(&mut insert_sql, conflict);
        }
        insert_sql.push_str("\nRETURNING \"id\"");
        cte_parts.push(format!("\"_w\" AS (\n{}\n)", insert_sql));
        cte_parts
    };
    cte_parts.extend(enqueue_ctes(&ins.enqueue_vector, "_w"));
    cte_parts.extend(enqueue_search_ctes(&ins.enqueue_search, "_w", ins.enqueue_vector.len()));

    let (shape, select_sql) = shape_select_from_cte(&ins.target, &ins.returning, "_w");
    let sql = format!(
        "WITH\n{}\n{}",
        cte_parts.join(",\n"),
        select_sql.unwrap_or_else(|| "SELECT * FROM \"_w\"".to_string()),
    );
    SqlOutput {
        sql,
        shape,
        inference_plan: None,
    }
}

// ── UPDATE ──────────────────────────────────────────────────────────────────

fn emit_poly_update_stmt(upd: &IrUpdate, user_ctes: &[IrCteDef]) -> SqlOutput {
    let alias = &upd.target.alias;
    let sets = update_set_fragments(&upd.assignments, &upd.rewrites, "");

    // Nested-DML CTEs first — a single hoisted CTE (e.g. a nested INSERT)
    // is materialized once and safely referenced from every per-implementor
    // branch below via its own FROM clause; at most one branch's filter
    // will ever actually match a row, since each row belongs to exactly
    // one concrete implementor table.
    let mut cte_parts: Vec<String> = emit_user_cte_parts(user_ctes);
    cte_parts.extend(emit_user_cte_parts(&upd.nested_ctes));
    let mut union_parts = vec![];
    let from_ctes = nested_cte_from(&upd.nested_ctes, &sets.join(","), "\n");

    let has_any_multilink = !upd.multi_link_clears.is_empty()
        || !upd.multi_link_replaces.is_empty()
        || !upd.multi_link_appends.is_empty()
        || !upd.multi_link_removals.is_empty();
    for (i, imp) in upd.poly_implementors.iter().enumerate() {
        let cte_name = format!("_u{}", i);
        // Junction rows live in each implementor's own junction table, so
        // with any to write each implementor is updated as the concrete type
        // it is.
        if has_any_multilink {
            cte_parts.extend(emit_update_multilink_cte_parts(
                &concrete_poly_update(upd, imp),
                &cte_name,
                false,
            ));
        } else {
            let mut upd_sql = format!(
                "UPDATE {} AS {}\nSET {}{}",
                qn(&imp.module, &imp.table),
                qi(alias),
                sets.join(", "),
                from_ctes,
            );
            append_filter(&mut upd_sql, &upd.filter);
            upd_sql.push_str(&format!("\nRETURNING {}.\"id\"", qi(alias)));
            cte_parts.push(format!("\"{}\" AS (\n{}\n)", cte_name, upd_sql));
        }

        let r_alias = format!("_r{}", i);
        union_parts.push(format!(
            "SELECT ROW({}::text, {}.\"id\") AS result FROM \"{}\" AS {}",
            sql_str(&imp.type_name),
            qi(&r_alias),
            cte_name,
            qi(&r_alias),
        ));
    }

    let sql = format!("WITH\n{}\n{}", cte_parts.join(",\n"), union_parts.join("\nUNION ALL\n"),);

    let (shape, _) = emit_returning_shape(&upd.target, &upd.returning, true);
    SqlOutput {
        sql,
        shape,
        inference_plan: None,
    }
}

fn emit_update_stmt(upd: &IrUpdate, user_ctes: &[IrCteDef]) -> SqlOutput {
    if !upd.poly_implementors.is_empty() {
        return emit_poly_update_stmt(upd, user_ctes);
    }
    let alias = &upd.target.alias;
    let (shape, returning_sql) = emit_returning_shape(&upd.target, &upd.returning, true);

    let has_any_multilink = !upd.multi_link_clears.is_empty()
        || !upd.multi_link_replaces.is_empty()
        || !upd.multi_link_appends.is_empty()
        || !upd.multi_link_removals.is_empty();

    if !has_any_multilink && upd.enqueue_vector.is_empty() && upd.enqueue_search.is_empty() {
        // No junction changes, no enqueue — plain UPDATE (possibly with a
        // user CTE prefix and/or hoisted nested-DML CTEs from a link value
        // like `author := (select (insert Person {...}) { id })` — see
        // IrUpdate::nested_ctes). A nested CTE's `id` column is referenced
        // from the SET clause, which needs a `FROM` clause to see it —
        // ordinary `UPDATE ... SET ...` has no FROM of its own otherwise.
        let sets = update_set_fragments(&upd.assignments, &upd.rewrites, "");
        let mut sql = format!(
            "UPDATE {} AS {}\nSET {}",
            target_ref(&upd.target),
            qi(alias),
            sets.join(", "),
        );
        sql.push_str(&nested_cte_from(&upd.nested_ctes, &sets.join(","), "\n"));
        append_filter(&mut sql, &upd.filter);
        if let Some(r) = returning_sql {
            sql.push_str(&r);
        }
        // The user's bindings first: a nested statement may read them, and a
        // CTE only sees the ones before it.
        let combined_ctes: Vec<IrCteDef> = user_ctes
            .iter()
            .cloned()
            .chain(upd.nested_ctes.iter().cloned())
            .collect();
        if !combined_ctes.is_empty() {
            sql = format!("{}{}", emit_cte_prefix(&combined_ctes), sql);
        }
        return SqlOutput {
            sql,
            shape,
            inference_plan: None,
        };
    }

    if !has_any_multilink && (!upd.enqueue_vector.is_empty() || !upd.enqueue_search.is_empty()) {
        // No junction changes but need to enqueue — wrap UPDATE in a CTE.
        let sets = update_set_fragments(&upd.assignments, &upd.rewrites, "");
        let mut upd_sql = format!(
            "    UPDATE {} AS {}\n    SET {}",
            target_ref(&upd.target),
            qi(alias),
            sets.join(", "),
        );
        upd_sql.push_str(&nested_cte_from(&upd.nested_ctes, &sets.join(","), "\n    "));
        append_filter(&mut upd_sql, &upd.filter);
        upd_sql.push_str("\n    RETURNING \"id\"");

        let mut cte_parts: Vec<String> = emit_user_cte_parts(user_ctes);
        cte_parts.extend(emit_user_cte_parts(&upd.nested_ctes));
        cte_parts.push(format!("\"_w\" AS (\n{}\n)", upd_sql));
        cte_parts.extend(enqueue_ctes(&upd.enqueue_vector, "_w"));
        cte_parts.extend(enqueue_search_ctes(&upd.enqueue_search, "_w", upd.enqueue_vector.len()));

        let (shape2, select_sql) = shape_select_from_cte(&upd.target, &upd.returning, "_w");
        let sql = format!(
            "WITH\n{}\n{}",
            cte_parts.join(",\n"),
            select_sql.unwrap_or_else(|| "SELECT * FROM \"_w\"".to_string()),
        );
        return SqlOutput {
            sql,
            shape: shape2,
            inference_plan: None,
        };
    }

    // CTE-based UPDATE for junction table mutations.
    let result_expr = if !upd.returning.is_empty() {
        let (pointer_exprs, _) = build_shape(&upd.returning, alias);
        let mut parts = vec![type_disc(&upd.target.type_name)];
        parts.extend(pointer_exprs);
        parts.join(",\n    ")
    } else {
        format!("{}.id", qi(alias))
    };

    let has_scalar_changes = !upd.assignments.is_empty() || !upd.rewrites.is_empty();

    // Nested-DML CTEs first (see IrUpdate::nested_ctes) — only relevant when
    // has_scalar_changes, since nested_ctes is only ever populated while
    // compiling scalar assignments, which is exactly the condition below.
    let mut cte_parts: Vec<String> = emit_user_cte_parts(user_ctes);
    cte_parts.extend(emit_user_cte_parts(&upd.nested_ctes));

    // _ids: the target rows (updated or selected).
    if has_scalar_changes {
        let sets = update_set_fragments(&upd.assignments, &upd.rewrites, "");
        let mut upd_sql = format!(
            "UPDATE {} AS {}\nSET {}",
            target_ref(&upd.target),
            qi(alias),
            sets.join(", "),
        );
        upd_sql.push_str(&nested_cte_from(&upd.nested_ctes, &sets.join(","), "\n"));
        append_filter(&mut upd_sql, &upd.filter);
        upd_sql.push_str("\nRETURNING *");
        cte_parts.push(format!("\"_ids\" AS (\n{}\n)", upd_sql));
    } else {
        let mut sel = format!(
            "SELECT {}.* FROM {} AS {}",
            qi(alias),
            target_ref(&upd.target),
            qi(alias),
        );
        append_filter(&mut sel, &upd.filter);
        cte_parts.push(format!("\"_ids\" AS (\n{}\n)", sel));
    }

    // Junction clears (`:= {}` and `:= expr` — the clear part of replace).
    for (i, clr) in upd.multi_link_clears.iter().enumerate() {
        let exclude = ml_clear_exclusion(upd.multi_link_replaces.get(i));
        let del = format!(
            "DELETE FROM {} WHERE {} IN (SELECT id FROM \"_ids\"){}",
            qn(&clr.module, &clr.junction_table),
            qi(&clr.source_col),
            exclude,
        );
        cte_parts.push(format!("\"_clr_{}\" AS (\n{}\n)", i, del));
    }

    // Junction appends (`+=`).
    for (i, app) in upd.multi_link_appends.iter().enumerate() {
        cte_parts.push(emit_ml_append_cte(
            app,
            "_ids",
            &upd.target.alias,
            &format!("_ml_add_{}", i),
        ));
    }

    // Junction removals (`-=`).
    for (i, rem) in upd.multi_link_removals.iter().enumerate() {
        cte_parts.push(emit_ml_remove_cte(
            rem,
            "_ids",
            &upd.target.alias,
            &format!("_ml_rm_{}", i),
        ));
    }

    // Junction inserts for replace (`:= expr` — insert after the clear).
    for (i, rep) in upd.multi_link_replaces.iter().enumerate() {
        cte_parts.push(emit_ml_append_cte(
            rep,
            "_ids",
            &upd.target.alias,
            &format!("_ml_rep_{}", i),
        ));
    }

    // Enqueue CTEs (source is _ids which has all columns including id).
    cte_parts.extend(enqueue_ctes(&upd.enqueue_vector, "_ids"));
    cte_parts.extend(enqueue_search_ctes(
        &upd.enqueue_search,
        "_ids",
        upd.enqueue_vector.len(),
    ));

    let sql = format!(
        "WITH\n{}\nSELECT (\n    {}\n) AS result\nFROM \"_ids\" AS {}",
        cte_parts.join(",\n"),
        result_expr,
        qi(alias),
    );
    SqlOutput {
        sql,
        shape,
        inference_plan: None,
    }
}

// ── DELETE ──────────────────────────────────────────────────────────────────

fn emit_delete_stmt(del: &IrDelete) -> SqlOutput {
    if !del.poly_implementors.is_empty() {
        return emit_poly_delete_stmt(del);
    }
    let alias = &del.target.alias;

    if del.enqueue_search.is_empty() {
        let mut sql = format!("DELETE FROM {} AS {}", target_ref(&del.target), qi(alias),);
        append_filter(&mut sql, &del.filter);
        let (shape, returning_sql) = emit_returning_shape(&del.target, &del.returning, true);
        if let Some(r) = returning_sql {
            sql.push_str(&r);
        }
        return SqlOutput {
            sql,
            shape,
            inference_plan: None,
        };
    }

    // Wrap DELETE in a CTE to enqueue OpenSearch delete jobs.
    let mut del_sql = format!("    DELETE FROM {} AS {}", target_ref(&del.target), qi(alias),);
    append_filter(&mut del_sql, &del.filter);
    del_sql.push_str("\n    RETURNING \"id\"");

    let mut cte_parts = vec![format!("\"_del\" AS (\n{}\n)", del_sql)];
    cte_parts.extend(enqueue_search_ctes(&del.enqueue_search, "_del", 0));

    let (shape, select_sql) = shape_select_from_cte(&del.target, &del.returning, "_del");
    let sql = format!(
        "WITH\n{}\n{}",
        cte_parts.join(",\n"),
        select_sql.unwrap_or_else(|| "SELECT * FROM \"_del\"".to_string()),
    );
    SqlOutput {
        sql,
        shape,
        inference_plan: None,
    }
}

fn emit_poly_delete_stmt(del: &IrDelete) -> SqlOutput {
    let alias = &del.target.alias;
    let mut cte_parts = vec![];
    let mut union_parts = vec![];

    for (i, imp) in del.poly_implementors.iter().enumerate() {
        let cte_name = format!("_d{}", i);
        let mut del_sql = format!("DELETE FROM {} AS {}", qn(&imp.module, &imp.table), qi(alias),);
        append_filter(&mut del_sql, &del.filter);
        del_sql.push_str(&format!("\nRETURNING {}.\"id\"", qi(alias)));
        cte_parts.push(format!("\"{}\" AS (\n{}\n)", cte_name, del_sql));

        let r_alias = format!("_r{}", i);
        union_parts.push(format!(
            "SELECT ROW({}::text, {}.\"id\") AS result FROM \"{}\" AS {}",
            sql_str(&imp.type_name),
            qi(&r_alias),
            cte_name,
            qi(&r_alias),
        ));
    }

    let sql = format!("WITH\n{}\n{}", cte_parts.join(",\n"), union_parts.join("\nUNION ALL\n"),);

    let (shape, _) = emit_returning_shape(&del.target, &del.returning, true);
    SqlOutput {
        sql,
        shape,
        inference_plan: None,
    }
}

// ── RETURNING helper ─────────────────────────────────────────────────────────

/// Builds the RETURNING clause and ShapeDescriptor for DML.
/// `with_alias`: UPDATE/DELETE can use the table alias; INSERT cannot.
fn emit_returning_shape(
    target: &IrSource,
    returning: &[IrShapePointer],
    with_alias: bool,
) -> (ShapeDescriptor, Option<String>) {
    if returning.is_empty() {
        return (
            ShapeDescriptor {
                root: ShapeNode::Scalar {
                    name: String::new(),
                    position: 0,
                },
            },
            None,
        );
    }

    let alias = if with_alias { target.alias.as_str() } else { "" };
    let (pointer_exprs, shape_pointers) = build_shape(returning, alias);

    let mut parts = vec![type_disc(&target.type_name)];
    parts.extend(pointer_exprs);
    let tuple = parts.join(",\n    ");
    let sql = format!("\nRETURNING (\n    {}\n) AS result", tuple);

    let root_pointers = prepend_type(shape_pointers);
    let shape = ShapeDescriptor {
        root: ShapeNode::Object {
            name: String::new(),
            type_name: Some(target.type_name.clone()),
            position: 0,
            cardinality: Cardinality::Required,
            pointers: root_pointers,
            has_implicit_id: shape_has_implicit_id(returning),
        },
    };
    (shape, Some(sql))
}

// ── Shape emission ───────────────────────────────────────────────────────────

fn emit_scalar_set(f: &IrScalarSetPointer, pos: usize) -> (String, ShapeNode) {
    let from_sql = if !f.poly_implementors.is_empty() {
        format!(
            "(\n{}\n) AS {}",
            emit_poly_union(&f.poly_implementors, &f.poly_columns),
            qi(&f.source.alias)
        )
    } else {
        format!("{} AS {}", source_ref(&f.source), qi(&f.source.alias))
    };
    let sql = format!(
        "(SELECT COALESCE(array_agg(ROW({})::record), ARRAY[]::record[]) FROM {})",
        emit_expr(&f.bool_expr),
        from_sql,
    );
    let node = ShapeNode::Array {
        name: f.alias.clone(),
        position: pos,
        element: Box::new(ShapeNode::Scalar {
            name: String::new(),
            position: 0,
        }),
    };
    (sql, node)
}

/// Build SQL expressions and ShapeNodes for `pointers`, starting at position 1
/// (position 0 is always the type discriminator, added by the caller).
/// True when the compiler put an `id` at the front of this shape that the
/// query never asked for — see `IrScalarPointer::implicit_id`.
fn shape_has_implicit_id(pointers: &[IrShapePointer]) -> bool {
    matches!(pointers.first(), Some(IrShapePointer::Scalar(p)) if p.implicit_id)
}

fn build_shape(pointers: &[IrShapePointer], table_alias: &str) -> (Vec<String>, Vec<ShapeNode>) {
    let mut exprs = Vec::new();
    let mut nodes = Vec::new();

    for (i, pointer) in pointers.iter().enumerate() {
        let (sql, node) = emit_shape_pointer(pointer, table_alias, i + 1);
        exprs.push(sql);
        nodes.push(node);
    }

    (exprs, nodes)
}

/// A `$1`, `$2`, … template with each placeholder replaced by its argument's
/// SQL, in one pass over the template alone. Replacing one number at a time
/// over the growing string would rewrite a query parameter an argument
/// carries (`$5` inside the first argument, replaced again as the fifth), and
/// `$1` also matches the start of `$10`.
fn fill_template(template: &str, args: &[String]) -> String {
    let mut out = String::with_capacity(template.len());
    let mut chars = template.char_indices().peekable();
    while let Some((start, c)) = chars.next() {
        if c != '$' {
            out.push(c);
            continue;
        }
        let mut end = start + 1;
        while let Some((i, d)) = chars.peek().copied()
            && d.is_ascii_digit()
        {
            end = i + d.len_utf8();
            chars.next();
        }
        match template[start + 1..end].parse::<usize>() {
            Ok(n) if (1..=args.len()).contains(&n) => out.push_str(&args[n - 1]),
            _ => out.push_str(&template[start..end]),
        }
    }
    out
}

/// `, message` for an assert given `message := …`, else nothing.
fn assert_message_arg(message: &Option<IrExpr>) -> String {
    message
        .as_ref()
        .map(|m| format!(", ({})::text", emit_expr(m)))
        .unwrap_or_default()
}

fn emit_shape_pointer(pointer: &IrShapePointer, table_alias: &str, pos: usize) -> (String, ShapeNode) {
    match pointer {
        IrShapePointer::Scalar(f) => emit_scalar(f, table_alias, pos),
        IrShapePointer::SingleLink(f) => emit_single_link(f, table_alias, pos),
        IrShapePointer::MultiLink(f) => emit_multi_link(f, table_alias, pos),
        IrShapePointer::Computed(f) => (emit_expr(&f.expr), expr_shape_node(&f.alias, pos, &f.expr)),
        IrShapePointer::ScalarSet(f) => emit_scalar_set(f, pos),
        IrShapePointer::Asserted(a) => {
            let (sql, node) = emit_shape_pointer(&a.inner, table_alias, pos);
            // The inner pointer aggregates rows to `record[]`, which a PL/pgSQL
            // function cannot take as a parameter at all -- hence the cast to
            // `text[]`, whose rendering compares element-for-element the same
            // way. Bound in a sub-select so the pointer's own subquery is
            // evaluated once rather than in both the check and the result, and
            // read through `cardinality` because the assert hands back the
            // array it was given.
            // A pointer landing on one object reads as a lone record, not an
            // array, so its set is re-read through `ARRAY(…)`; the unlimited
            // check form would otherwise fail on more than one row.
            let checked_set = match &a.check {
                Some(check) if emits_one_object(check) => {
                    format!("ARRAY{}", emit_shape_pointer(check, table_alias, pos).0)
                }
                Some(check) => emit_shape_pointer(check, table_alias, pos).0,
                None if emits_one_object(&a.inner) => format!("ARRAY{sql}"),
                None => "\"_a\".\"v\"".to_string(),
            };
            let checked = format!(
                "(SELECT \"_a\".\"v\" FROM (SELECT {sql} AS \"v\") AS \"_a\"\n                     WHERE cardinality(\"_pylon\".{}({checked_set}::text[]{})) >= 0)",
                qi(&a.fn_name),
                assert_message_arg(&a.message),
            );
            (checked, node)
        }
    }
}

/// True when the pointer is emitted as a single `(SELECT …)` record rather
/// than an aggregated array of them.
fn emits_one_object(pointer: &IrShapePointer) -> bool {
    match pointer {
        IrShapePointer::SingleLink(_) => true,
        IrShapePointer::Computed(c) => matches!(
            c.expr,
            IrExpr::ObjectPathSubquery(_) | IrExpr::ObjectPathUnion { multi: false, .. }
        ),
        IrShapePointer::Asserted(a) => emits_one_object(&a.inner),
        _ => false,
    }
}

/// Convert a PostgreSQL schema-qualified type name (`"module"."TypeName"`) to
/// a Pylon-qualified name (`module::TypeName`).
fn pg_quoted_to_pylon(pg_type: &str) -> String {
    let inner = pg_type.trim_start_matches('"');
    if let Some(idx) = inner.find(r#""."#) {
        let module = &inner[..idx];
        let type_name = inner[idx + 3..].trim_end_matches('"');
        format!("{}::{}", module, type_name)
    } else {
        pg_type.to_string()
    }
}

/// A schema-qualified custom type (an enum or a domain) named in a property's
/// `pg_type`, together with whether the column holds an array of it.
///
/// Both forms are emitted as text -- `::text` for a scalar, `::text[]` for an
/// array -- so a value's shape never depends on an OID the database assigns
/// when the migration creating the type runs. The array case has to be told
/// apart from the scalar one: `"access"."AuthenticationMethod"[]` starts with a
/// quote just as `"account"."InterfaceAppearance"` does, and reading it as a
/// scalar enum casts a whole array to a single `text` (`'{MagicLink}'`) and
/// shapes it as one enum value.
struct QualifiedPgType {
    /// Pylon-qualified name of the element type, e.g. `access::AuthenticationMethod`.
    name: String,
    is_array: bool,
}

impl QualifiedPgType {
    fn of(pg_type: &str) -> Option<Self> {
        let (element, is_array) = match pg_type.strip_suffix("[]") {
            Some(element) => (element, true),
            None => (pg_type, false),
        };
        element.starts_with('"').then(|| Self {
            name: pg_quoted_to_pylon(element),
            is_array,
        })
    }

    /// The cast that carries this value back as text.
    fn text_cast(&self) -> &'static str {
        if self.is_array { "::text[]" } else { "::text" }
    }

    /// How the decoded value is shaped: a single enum, or a list of them.
    fn shape_node(&self, name: String, position: usize) -> ShapeNode {
        let element = ShapeNode::Enum {
            name: name.clone(),
            position,
            enum_type: self.name.clone(),
        };
        if !self.is_array {
            return element;
        }
        ShapeNode::Array {
            name,
            // Each element arrives as the label itself rather than as a field
            // of a record, which the decoders read at position 0.
            element: Box::new(ShapeNode::Enum {
                name: String::new(),
                position: 0,
                enum_type: self.name.clone(),
            }),
            position,
        }
    }
}

fn emit_scalar(f: &IrScalarPointer, table_alias: &str, pos: usize) -> (String, ShapeNode) {
    if let Some(nt_name) = f.pg_type.strip_prefix("__nt__:") {
        let sql = if table_alias.is_empty() {
            format!("{}::jsonb", qi(&f.column))
        } else {
            format!("{}.{}::jsonb", qi(table_alias), qi(&f.column))
        };
        return (
            sql,
            ShapeNode::NamedTuple {
                name: f.alias.clone(),
                position: pos,
                type_name: Some(nt_name.to_string()),
                members: f.tuple_shape.as_ref().map(|s| s.members.clone()),
                is_free_object: false,
            },
        );
    }
    // Schema-qualified custom types (enums, domains) have runtime OIDs unknown to a static
    // anonymous_record_decode. Cast to text — the string label is all the decoder needs.
    if let Some(qualified) = QualifiedPgType::of(&f.pg_type) {
        let cast = qualified.text_cast();
        let sql = if table_alias.is_empty() {
            format!("{}{}", qi(&f.column), cast)
        } else {
            format!("{}.{}{}", qi(table_alias), qi(&f.column), cast)
        };
        return (sql, qualified.shape_node(f.alias.clone(), pos));
    }
    // A structural pylon.Tuple[...]-typed property — same jsonb column shape
    // as the nominal `__nt__:` case above, just with no registered dataclass
    // to hydrate (type_name stays None).
    if let Some(shape) = &f.tuple_shape {
        let sql = if table_alias.is_empty() {
            format!("{}::jsonb", qi(&f.column))
        } else {
            format!("{}.{}::jsonb", qi(table_alias), qi(&f.column))
        };
        return (
            sql,
            ShapeNode::NamedTuple {
                name: f.alias.clone(),
                position: pos,
                type_name: shape.type_name.clone(),
                members: Some(shape.members.clone()),
                is_free_object: false,
            },
        );
    }
    let sql = if table_alias.is_empty() {
        format!("{}::{}", qi(&f.column), f.pg_type)
    } else {
        format!("{}.{}::{}", qi(table_alias), qi(&f.column), f.pg_type)
    };
    (
        sql,
        ShapeNode::Scalar {
            name: f.alias.clone(),
            position: pos,
        },
    )
}

fn emit_single_link(f: &IrSingleLinkPointer, parent_alias: &str, pos: usize) -> (String, ShapeNode) {
    let sub = &f.subquery;
    let [IrRowSource::Bound { source, shape }] = sub.rows.as_slice() else {
        unreachable!("single-link subquery is always schema-bound")
    };
    let sub_alias = &source.alias;

    let (sub_exprs, mut sub_nodes) = build_shape(shape, sub_alias);
    let mut parts = vec![source_type_disc(source)];
    parts.extend(sub_exprs);

    // Link properties: read from the junction table alias "jt" — only ever
    // populated for a junction-backed link (see `IrSingleLinkPointer`'s doc
    // comment), which is exactly when `f.correlation` below puts "jt" in
    // scope.
    for lp in &f.link_properties {
        parts.push(format!("\"jt\".{}", qi(&lp.name)));
        let pos = sub_nodes.len() + 1;
        sub_nodes.push(ShapeNode::Scalar {
            name: format!("@{}", lp.name),
            position: pos,
        });
    }

    let tuple = parts.join(",\n        ");

    let (from_sql, mut where_parts) = match &f.correlation {
        IrSingleLinkCorrelation::Fk { fk_column, target_pk } => {
            let from = format!("FROM {} AS {}", source_ref(source), qi(sub_alias));
            let cond = format!(
                "{}.{} = {}.{}",
                qi(parent_alias),
                qi(fk_column),
                qi(sub_alias),
                qi(target_pk),
            );
            (from, vec![cond])
        }
        // Junction-backed — same join shape a multi-link's correlated
        // subquery uses; cardinality-one is a schema-level invariant on the
        // junction table, not something this query needs to enforce itself.
        IrSingleLinkCorrelation::Junction { join, target_pk: _ } => match join {
            IrMultiLinkJoin::Standard { junction_table, module } => {
                let from = format!(
                    "FROM {} AS \"jt\"\n    INNER JOIN {} AS {}\n    ON {}.id = \"jt\".target",
                    junction_ref(module, junction_table),
                    source_ref(source),
                    qi(sub_alias),
                    qi(sub_alias),
                );
                let cond = format!("\"jt\".source = {}.id", qi(parent_alias));
                (from, vec![cond])
            }
            IrMultiLinkJoin::Through {
                junction_table,
                module,
                source_col,
                target_col,
            } => {
                let from = format!(
                    "FROM {} AS \"jt\"\n    INNER JOIN {} AS {}\n    ON {}.id = \"jt\".{}",
                    junction_ref(module, junction_table),
                    source_ref(source),
                    qi(sub_alias),
                    qi(sub_alias),
                    qi(target_col),
                );
                let cond = format!("\"jt\".{} = {}.id", qi(source_col), qi(parent_alias));
                (from, vec![cond])
            }
            IrMultiLinkJoin::BacklinkFk { .. } | IrMultiLinkJoin::BacklinkJunction { .. } => {
                unreachable!("a junction-backed single link's own forward join is always Standard or Through")
            }
        },
    };
    if let Some(filter) = &sub.filter {
        where_parts.push(emit_expr(filter));
    }

    let mut sql = format!(
        "(SELECT (\n        {}\n    )\n    {}\n    WHERE {}",
        tuple,
        from_sql,
        where_parts.join(" AND "),
    );
    if !sub.order_by.is_empty() {
        let s: Vec<_> = sub.order_by.iter().map(emit_sort_clause).collect();
        sql.push_str(&format!("\n    ORDER BY {}", s.join(", ")));
    }
    sql.push(')');

    let node = ShapeNode::Object {
        name: f.alias.clone(),
        type_name: Some(source.type_name.clone()),
        position: pos,
        cardinality: Cardinality::Optional,
        pointers: prepend_type(sub_nodes),
        has_implicit_id: shape_has_implicit_id(shape),
    };
    (sql, node)
}

fn emit_multi_link(f: &IrMultiLinkPointer, parent_alias: &str, pos: usize) -> (String, ShapeNode) {
    let sub = &f.subquery;
    let [IrRowSource::Bound { source, shape }] = sub.rows.as_slice() else {
        unreachable!("multi-link subquery is always schema-bound")
    };
    let sub_alias = &source.alias;

    let (sub_exprs, mut sub_nodes) = build_shape(shape, sub_alias);
    let mut row_parts = vec![source_type_disc(source)];
    row_parts.extend(sub_exprs);

    // Link properties: read from the junction table alias "jt".
    // The ShapeNode name carries the `@` prefix so hydration stores it as
    // `@prop` in the object's __dict__ and the REPL displays it with `@`.
    for lp in &f.link_properties {
        row_parts.push(format!("\"jt\".{}", qi(&lp.name)));
        // Positions are 1-based (0 = type discriminator). sub_nodes.len() gives
        // the count of already-assigned positions, so the next position is len+1.
        let pos = sub_nodes.len() + 1;
        sub_nodes.push(ShapeNode::Scalar {
            name: format!("@{}", lp.name),
            position: pos,
        });
    }

    let row = row_parts.join(",\n            ");

    // ORDER BY inside array_agg
    let order_sql = if !sub.order_by.is_empty() {
        let s: Vec<_> = sub.order_by.iter().map(emit_sort_clause).collect();
        format!(" ORDER BY {}", s.join(", "))
    } else {
        String::new()
    };

    let (from_sql, source_cond) = match &f.join {
        IrMultiLinkJoin::Standard { junction_table, module } => {
            let from = format!(
                "FROM {} AS \"jt\"\n    INNER JOIN {} AS {}\n    ON {}.id = \"jt\".target",
                junction_ref(module, junction_table),
                source_ref(source),
                qi(sub_alias),
                qi(sub_alias),
            );
            let cond = format!("\"jt\".source = {}.id", qi(parent_alias));
            (from, cond)
        }
        IrMultiLinkJoin::Through {
            junction_table,
            module,
            source_col,
            target_col,
        } => {
            let from = format!(
                "FROM {} AS \"jt\"\n    INNER JOIN {} AS {}\n    ON {}.id = \"jt\".{}",
                junction_ref(module, junction_table),
                source_ref(source),
                qi(sub_alias),
                qi(sub_alias),
                qi(target_col),
            );
            let cond = format!("\"jt\".{} = {}.id", qi(source_col), qi(parent_alias));
            (from, cond)
        }
        // Reverse of a single link: no junction table, the owner
        // (sub-select) rows are correlated directly by their own FK column.
        IrMultiLinkJoin::BacklinkFk { fk_col } => {
            let from = format!("FROM {} AS {}", source_ref(source), qi(sub_alias));
            let cond = format!("{}.{} = {}.id", qi(sub_alias), qi(fk_col), qi(parent_alias));
            (from, cond)
        }
        // Reverse of a multi-link: same junction table shape as `Standard`/
        // `Through`, with the owner/current column roles swapped — the
        // sub-select's own rows join via `owner_col`, the outer (current)
        // row correlates via `current_col`.
        IrMultiLinkJoin::BacklinkJunction {
            junction_table,
            module,
            owner_col,
            current_col,
        } => {
            let from = format!(
                "FROM {} AS \"jt\"\n    INNER JOIN {} AS {}\n    ON {}.id = \"jt\".{}",
                junction_ref(module, junction_table),
                source_ref(source),
                qi(sub_alias),
                qi(sub_alias),
                qi(owner_col),
            );
            let cond = format!("\"jt\".{} = {}.id", qi(current_col), qi(parent_alias));
            (from, cond)
        }
    };

    let mut where_parts = vec![source_cond];
    if let Some(filter) = &sub.filter {
        where_parts.push(emit_expr(filter));
    }

    if f.single {
        let mut sql = format!(
            "(SELECT (\n            {}\n        )\n    {}\n    WHERE {}",
            row,
            from_sql,
            where_parts.join(" AND "),
        );
        append_order_by(&mut sql, &sub.order_by);
        append_offset_limit(&mut sql, &sub.offset, &sub.limit);
        sql.push(')');
        let node = ShapeNode::Object {
            name: f.alias.clone(),
            type_name: Some(source.type_name.clone()),
            position: pos,
            cardinality: Cardinality::Optional,
            pointers: prepend_type(sub_nodes),
            has_implicit_id: shape_has_implicit_id(shape),
        };
        return (sql, node);
    }

    if f.single {
        let mut sql = format!(
            "(SELECT (\n            {}\n        )\n    {}\n    WHERE {}",
            row,
            from_sql,
            where_parts.join(" AND "),
        );
        append_order_by(&mut sql, &sub.order_by);
        append_offset_limit(&mut sql, &sub.offset, &sub.limit);
        sql.push(')');
        let node = ShapeNode::Object {
            name: f.alias.clone(),
            type_name: Some(source.type_name.clone()),
            position: pos,
            cardinality: Cardinality::Optional,
            pointers: prepend_type(sub_nodes),
            has_implicit_id: shape_has_implicit_id(shape),
        };
        return (sql, node);
    }

    let sql = if sub.limit.is_some() || sub.offset.is_some() {
        // OFFSET/LIMIT cut the *rows* that go into the array, so they can't
        // sit next to the aggregate — the aggregate collapses them to one
        // row first. Select the row's parts (and any sort keys) in a derived
        // table that carries the modifiers, then aggregate over that.
        let mut cols: Vec<String> = row_parts
            .iter()
            .enumerate()
            .map(|(i, part)| format!("{} AS \"c{}\"", part, i))
            .collect();
        let agg_row: Vec<String> = (0..row_parts.len()).map(|i| format!("\"__lim\".\"c{}\"", i)).collect();
        let mut agg_order: Vec<String> = vec![];
        for (i, s) in sub.order_by.iter().enumerate() {
            cols.push(format!("{} AS \"s{}\"", emit_expr(&s.expr), i));
            agg_order.push(emit_sort_clause(&IrSort {
                expr: IrExpr::ColumnRef {
                    alias: "__lim".to_string(),
                    column: format!("s{}", i),
                    pg_type: String::new(),
                },
                direction: s.direction.clone(),
                nulls: s.nulls.clone(),
            }));
        }
        let mut inner = format!(
            "SELECT {}\n    {}\n    WHERE {}",
            cols.join(",\n        "),
            from_sql,
            where_parts.join(" AND "),
        );
        append_order_by(&mut inner, &sub.order_by);
        append_offset_limit(&mut inner, &sub.offset, &sub.limit);
        let agg_order_sql = if agg_order.is_empty() {
            String::new()
        } else {
            format!(" ORDER BY {}", agg_order.join(", "))
        };
        format!(
            "(SELECT COALESCE(\n        array_agg(ROW(\n            {}\n        )::record{}),\n        ARRAY[]::record[]\n    )\n    FROM ({}) AS \"__lim\")",
            agg_row.join(",\n            "),
            agg_order_sql,
            inner,
        )
    } else {
        format!(
            "(SELECT COALESCE(\n        array_agg(ROW(\n            {}\n        )::record{}),\n        ARRAY[]::record[]\n    )\n    {}\n    WHERE {})",
            row,
            order_sql,
            from_sql,
            where_parts.join(" AND "),
        )
    };

    let node = ShapeNode::Array {
        name: f.alias.clone(),
        position: pos,
        element: Box::new(ShapeNode::Object {
            name: String::new(),
            type_name: Some(source.type_name.clone()),
            position: 0,
            cardinality: Cardinality::Required,
            pointers: prepend_type(sub_nodes),
            has_implicit_id: shape_has_implicit_id(shape),
        }),
    };
    (sql, node)
}

/// Prepend `ShapeNode::Scalar { name: "__type__", position: 0 }` and shift
/// existing nodes' positions by 1.
fn prepend_type(nodes: Vec<ShapeNode>) -> Vec<ShapeNode> {
    let mut out = vec![ShapeNode::Scalar {
        name: "__type__".into(),
        position: 0,
    }];
    out.extend(nodes);
    out
}

// ── SQL clause helpers ──────────────────────────────────────────────────────

fn append_filter(sql: &mut String, filter: &Option<IrExpr>) {
    if let Some(f) = filter {
        sql.push_str(&format!("\nWHERE {}", emit_expr(f)));
    }
}

fn append_order_by(sql: &mut String, order_by: &[IrSort]) {
    if !order_by.is_empty() {
        let s: Vec<_> = order_by.iter().map(emit_sort_clause).collect();
        sql.push_str(&format!("\nORDER BY {}", s.join(", ")));
    }
}

fn append_offset_limit(sql: &mut String, offset: &Option<IrExpr>, limit: &Option<IrExpr>) {
    if let Some(o) = offset {
        sql.push_str(&format!("\nOFFSET {}", emit_expr(o)));
    }
    if let Some(l) = limit {
        sql.push_str(&format!("\nLIMIT {}", emit_expr(l)));
    }
}

/// `FOR UPDATE`/`FOR SHARE`/... — Postgres's own grammar places this last,
/// after `ORDER BY`/`LIMIT`/`OFFSET`, so this must be the final thing
/// appended to a bound SELECT's SQL. No `OF table_name` — the outer
/// SELECT's own aliased table is always the only lockable target (nested
/// shape pointers are correlated subqueries, never part of the outer FROM),
/// so a bare `FOR UPDATE` already targets exactly the right row(s).
fn append_lock_clause(sql: &mut String, lock: &Option<IrLockClause>) {
    let Some(lock) = lock else { return };
    let strength = match lock.strength {
        IrLockStrength::Update => "UPDATE",
        IrLockStrength::NoKeyUpdate => "NO KEY UPDATE",
        IrLockStrength::Share => "SHARE",
        IrLockStrength::KeyShare => "KEY SHARE",
    };
    sql.push_str(&format!("\nFOR {}", strength));
    match lock.wait {
        IrLockWait::Block => {}
        IrLockWait::NoWait => sql.push_str(" NOWAIT"),
        IrLockWait::SkipLocked => sql.push_str(" SKIP LOCKED"),
    }
}

fn emit_sort_clause(s: &IrSort) -> String {
    let dir = match s.direction {
        IrSortDir::Asc => "ASC",
        IrSortDir::Desc => "DESC",
    };
    let nulls = match s.nulls {
        IrNulls::First => "NULLS FIRST",
        IrNulls::Last => "NULLS LAST",
    };
    format!("{} {} {}", emit_expr(&s.expr), dir, nulls)
}

// ── Expression emission ─────────────────────────────────────────────────────

/// One side of `a intersect b` as the rows of the set it stands for: an
/// array's elements, a binding's rows, else the single value.
fn set_operand(expr: &IrExpr) -> String {
    match expr {
        IrExpr::ArrayFromSelect(_) => format!("SELECT unnest({})", emit_expr(expr)),
        IrExpr::CteRef { name, scalar, .. } => {
            format!("SELECT {} FROM {}", if *scalar { "\"v\"" } else { "\"id\"" }, qi(name))
        }
        _ => format!("SELECT {}", emit_expr(expr)),
    }
}

/// `emit_expr`, reading the types with subtypes through `fanouts` the way
/// `emit` does — see `IrOutput::subtype_fanouts`.
pub fn emit_expr_with_fanouts(expr: &IrExpr, fanouts: &HashMap<(String, String), IrPolyFanout>) -> String {
    with_subtype_fanouts(fanouts, || emit_expr(expr))
}

pub fn emit_expr(expr: &IrExpr) -> String {
    match expr {
        IrExpr::ColumnRef { alias, column, .. } => {
            if alias.is_empty() {
                qi(column)
            } else if alias == "NEW" || alias == "OLD" {
                // plpgsql's trigger row variables (`ir::compile_trigger_handler`'s
                // `__new__`/`__old__` anchors) are recognized by the plpgsql
                // parser as bare, unquoted identifiers — quoting them like an
                // ordinary table alias (`"NEW"."col"`) turns this into an
                // *ordinary* SQL column reference instead, which Postgres then
                // rejects with "missing FROM-clause entry for table NEW" the
                // moment it appears inside a nested SQL command (e.g. the
                // VALUES list of an `INSERT` the trigger handler runs) — there's
                // no real `NEW` table to resolve it against there, only
                // plpgsql's own row-variable substitution recognizes it.
                format!("{alias}.{}", qi(column))
            } else {
                format!("{}.{}", qi(alias), qi(column))
            }
        }
        IrExpr::Param { index } => format!("${}", index + 1),
        IrExpr::Literal(lit) => emit_literal(lit),
        IrExpr::BinOp(op) => {
            let l = emit_expr(&op.left);
            let r = emit_expr(&op.right);
            match op.op {
                BinOpKind::Add => format!("({} + {})", l, r),
                BinOpKind::Sub => format!("({} - {})", l, r),
                BinOpKind::Mul => format!("({} * {})", l, r),
                BinOpKind::Div => format!("({} / {})", l, r),
                BinOpKind::FloorDiv => {
                    if is_integer_expr(&op.left) && is_integer_expr(&op.right) {
                        format!("({} / {})", l, r)
                    } else {
                        format!("floor(({}) / ({}))", l, r)
                    }
                }
                BinOpKind::Mod => format!("({} % {})", l, r),
                BinOpKind::Pow => format!("power({}, {})", l, r),
                BinOpKind::Eq => format!("({} = {})", l, r),
                BinOpKind::Ne => format!("({} <> {})", l, r),
                BinOpKind::Lt => format!("({} < {})", l, r),
                BinOpKind::Le => format!("({} <= {})", l, r),
                BinOpKind::Gt => format!("({} > {})", l, r),
                BinOpKind::Ge => format!("({} >= {})", l, r),
                BinOpKind::And => format!("({} AND {})", l, r),
                BinOpKind::Or => format!("({} OR {})", l, r),
                BinOpKind::Like => format!("({} LIKE {})", l, r),
                BinOpKind::Ilike => format!("({} ILIKE {})", l, r),
                BinOpKind::NotLike => format!("({} NOT LIKE {})", l, r),
                BinOpKind::NotIlike => format!("({} NOT ILIKE {})", l, r),
                // `ANY`/`ALL` take the array directly, so unwrap a
                // right-hand `std::array_unpack` rather than emitting its
                // `unnest()` — a set-returning function is not allowed in
                // `WHERE`, and PostgreSQL rejects it at execution time even
                // though the PyQL and the SQL both look fine.
                BinOpKind::In => format!("({} = ANY({}))", l, unwrap_unnest_for_any(&op.right, &r)),
                BinOpKind::NotIn => format!("({} <> ALL({}))", l, unwrap_unnest_for_any(&op.right, &r)),
                BinOpKind::Coalesce => format!("COALESCE({}, {})", l, r),
                BinOpKind::CoalesceEq => format!("({} IS NOT DISTINCT FROM {})", l, r),
                BinOpKind::CoalesceNe => format!("({} IS DISTINCT FROM {})", l, r),
                BinOpKind::Concat => format!("({} || {})", l, r),
            }
        }
        IrExpr::UnaryOp(op) => {
            let inner = emit_expr(&op.operand);
            match op.op {
                UnaryOpKind::Not => format!("(NOT {})", inner),
                UnaryOpKind::Minus => format!("(-{})", inner),
                UnaryOpKind::Exists => format!("EXISTS({})", inner),
                // A value that reached expression position is a single one,
                // and a single value is already distinct — `DISTINCT x` is not
                // an SQL expression. Set-valued operands carry their own
                // DISTINCT, applied where the set is built.
                UnaryOpKind::Distinct => inner,
            }
        }
        IrExpr::FunctionCall(f) => {
            let args: Vec<_> = f.args.iter().map(emit_expr).collect();
            if let Some(tmpl) = &f.sql_template {
                return fill_template(tmpl, &args);
            }
            let name = match &f.schema {
                Some(s) => format!("{}.{}", pg_schema(s), qi(&f.name)),
                None => f.name.clone(),
            };
            format!("{}({})", name, args.join(", "))
        }
        IrExpr::TypeCast(c) => {
            // PostgreSQL doesn't support arbitrary_type::jsonb; to_jsonb() accepts any input.
            // String literals have type "unknown" in PG, so cast to text first.
            if c.pg_type == "jsonb" {
                match &c.expr {
                    // A bare `$N` parameter has no PG type at all until told
                    // otherwise. to_jsonb($N) can't resolve it — it's a
                    // polymorphic function, which needs a *concrete* input
                    // type to dispatch on, and an unknown-typed placeholder
                    // gives it nothing to work with ("could not determine
                    // polymorphic type because input has type unknown").
                    // A direct ($N)::jsonb cast works because PG specially
                    // resolves an unknown-typed parameter against an
                    // explicit cast — the same reason `$1::uuid` works fine
                    // elsewhere in this file.
                    IrExpr::Param { .. } => format!("({})::jsonb", emit_expr(&c.expr)),
                    IrExpr::Literal(IrLiteral::Str(_)) => {
                        format!("to_jsonb({}::text)", emit_expr(&c.expr))
                    }
                    _ => format!("to_jsonb({})", emit_expr(&c.expr)),
                }
            } else if yields_jsonb(&c.expr) {
                // Casting json to a scalar yields the value, not its JSON
                // spelling — `<str>` of a json string is the string, with no
                // quotes around it. `::text` would keep them.
                format!("(({}) #>> '{{}}')::{}", emit_expr(&c.expr), c.pg_type)
            } else {
                format!("({})::{}", emit_expr(&c.expr), c.pg_type)
            }
        }
        IrExpr::IfElse(ie) => format!(
            "CASE WHEN {} THEN {} ELSE {} END",
            emit_expr(&ie.condition),
            emit_expr(&ie.if_),
            emit_expr(&ie.else_),
        ),
        IrExpr::Array(elems) => {
            if elems.is_empty() {
                "ARRAY[]::text[]".to_string()
            } else {
                let parts: Vec<String> = elems.iter().map(emit_expr).collect();
                format!("ARRAY[{}]", parts.join(", "))
            }
        }
        IrExpr::Null => "NULL".to_string(),
        IrExpr::AggOverSet {
            fn_name,
            schema: _,
            elems,
        } => {
            let union_all = elems
                .iter()
                .map(|e| format!("SELECT {}", emit_expr(e)))
                .collect::<Vec<_>>()
                .join(" UNION ALL ");
            format!("(SELECT {}(v) FROM ({}) AS _set(v))", fn_name, union_all)
        }
        IrExpr::AggOverCte { fn_name, cte, column } => format!(
            "(SELECT {}({}) FROM {})",
            fn_name,
            column.as_deref().map(qi).unwrap_or_else(|| "*".to_string()),
            qi(cte),
        ),
        IrExpr::ExistsOverCte { cte, column } => match column {
            Some(column) => format!("EXISTS(SELECT 1 FROM {} WHERE {} IS NOT NULL)", qi(cte), qi(column)),
            None => format!("EXISTS(SELECT 1 FROM {})", qi(cte)),
        },
        IrExpr::AggOverQuery { fn_name, inner } => {
            let inner_sql = emit_select_stmt(inner, &[]).sql;
            format!("(SELECT {}(*) FROM ({}) _agg)", fn_name, inner_sql)
        }
        IrExpr::ArrayFromSelect(src) => emit_array_source(src),
        IrExpr::ScalarSubquery(sel) => {
            format!(
                "(SELECT \"v\" FROM (\n{}\n) AS \"_scalar_sub\")",
                emit_select_stmt(sel, &[]).sql
            )
        }

        IrExpr::SetOp { op, left, right, mode } => {
            let set = format!("({}) {} ({})", set_operand(left), op.sql(), set_operand(right));
            match mode {
                crate::ir::SetOpMode::Exists => format!("EXISTS({set})"),
                crate::ir::SetOpMode::Array => format!("ARRAY({set})"),
                crate::ir::SetOpMode::Aggregate(f) => {
                    format!("(SELECT {f}(\"_s\".\"v\") FROM ({set}) AS \"_s\"(\"v\"))")
                }
            }
        }

        IrExpr::CteRef { name, scalar, .. } => {
            // scalar CTEs emit `ROW(expr) AS result, expr AS v`; use `v` for
            // expression context so we get the plain scalar type, not record.
            let col = if *scalar { "v" } else { "id" };
            format!("(SELECT \"{}\" FROM \"{}\")", col, name)
        }

        IrExpr::CteFieldRef { name, field, .. } => {
            format!("(SELECT {} FROM {})", qi(field), qi(name))
        }

        IrExpr::ForVar { name, .. } => format!("\"_for_{}\".\"v\"", name),

        IrExpr::EnumLiteral { pg_type, variant } => {
            format!("'{}'::{}", variant.replace('\'', "''"), pg_type)
        }

        IrExpr::GlobalParam { index, pg_type } => {
            format!("(${}::{})", index + 1, pg_type)
        }

        IrExpr::GlobalRef { cte_name } => {
            format!("(SELECT \"value\" FROM \"{}\")", cte_name)
        }

        IrExpr::NamedTuple { fields, .. } => {
            let pairs: Vec<String> = fields
                .iter()
                .flat_map(|(k, v)| [format!("'{}'", k.replace('\'', "''")), emit_expr(v)])
                .collect();
            format!("jsonb_build_object({})", pairs.join(", "))
        }

        IrExpr::Tuple(elems) => {
            let items: Vec<String> = elems.iter().map(emit_expr).collect();
            format!("jsonb_build_array({})", items.join(", "))
        }

        IrExpr::Subscript { expr, index, is_array } => {
            let e = emit_expr(expr);
            let i = emit_expr(index);
            if *is_array {
                format!("_pylon.array_subscript({}, ({})::bigint)", e, i)
            } else {
                format!("_pylon.str_subscript({}, ({})::bigint)", e, i)
            }
        }

        IrExpr::Slice {
            expr,
            lower,
            upper,
            is_array,
        } => {
            let e = emit_expr(expr);
            if *is_array {
                let lo = lower
                    .as_deref()
                    .map(|x| format!("({}) + 1", emit_expr(x)))
                    .unwrap_or_else(|| "1".to_string());
                let hi = upper.as_deref().map(emit_expr).unwrap_or_default();
                if hi.is_empty() {
                    format!("({})[{}:]", e, lo)
                } else {
                    format!("({})[{}:{}]", e, lo, hi)
                }
            } else {
                // substr(expr, start, length) for text/bytea.
                let start = lower
                    .as_deref()
                    .map(|x| format!("({}) + 1", emit_expr(x)))
                    .unwrap_or_else(|| "1".to_string());
                match upper.as_deref() {
                    Some(hi_expr) => {
                        let lo_val = lower.as_deref().map(emit_expr).unwrap_or_else(|| "0".to_string());
                        // GREATEST(0, ...) so reversed bounds yield '' instead of a PG error.
                        format!(
                            "substr({}, {}, GREATEST(0, ({}) - ({})))",
                            e,
                            start,
                            emit_expr(hi_expr),
                            lo_val
                        )
                    }
                    None => format!("substr({}, {})", e, start),
                }
            }
        }

        IrExpr::JsonbField { expr, field } => {
            format!("({}->{})", emit_expr(expr), sql_str(field))
        }

        IrExpr::JsonbIndex { expr, index } => {
            format!("({}->{})", emit_expr(expr), index)
        }

        IrExpr::FnParam { name, .. } => qi(name),

        IrExpr::FnSubquery(fs) => {
            let scalar = match fs.shape.as_slice() {
                [IrShapePointer::Computed(c)] => emit_expr(&c.expr),
                _ => unreachable!("a function scalar subquery always projects exactly one computed pointer"),
            };
            let args_sql = fs.fn_args.iter().map(emit_expr).collect::<Vec<_>>().join(", ");
            let mut sql = format!(
                "(SELECT {}{}\nFROM {}.{}({}) AS {}",
                if fs.distinct { "DISTINCT " } else { "" },
                scalar,
                pg_schema(&fs.fn_module),
                qi(&fs.fn_name),
                args_sql,
                qi(&fs.alias),
            );
            append_filter(&mut sql, &fs.filter);
            append_order_by(&mut sql, &fs.order_by);
            append_offset_limit(&mut sql, &fs.offset, &fs.limit);
            sql.push(')');
            sql
        }

        IrExpr::PathSubquery(ps) => {
            let scalar = match &ps.result {
                IrPathResult::Scalar(e, _) => emit_expr(e),
                IrPathResult::Object { alias, .. } => format!("{}.\"id\"", qi(alias)),
            };
            let from_sql = emit_path_joins(&ps.root, &ps.joins);
            let mut sql = format!(
                "(SELECT {}{}\nFROM {}",
                if ps.distinct { "DISTINCT " } else { "" },
                scalar,
                from_sql
            );
            append_filter(&mut sql, &ps.filter);
            append_order_by(&mut sql, &ps.order_by);
            append_offset_limit(&mut sql, &ps.offset, &ps.limit);
            sql.push(')');
            sql
        }

        IrExpr::ObjectPathUnion { branches, limit, multi } => {
            let arms: Vec<String> = branches
                .iter()
                .map(|ps| {
                    let IrPathResult::Object {
                        alias,
                        type_name,
                        shape,
                    } = &ps.result
                    else {
                        unreachable!("an object path union's branches always land on objects")
                    };
                    let (pointer_exprs, _) = build_shape(shape, alias);
                    let mut parts = vec![result_type_disc(ps, alias, type_name)];
                    parts.extend(pointer_exprs);
                    let mut sql = format!(
                        "SELECT (\n    {}\n) AS \"r\"\nFROM {}",
                        parts.join(",\n    "),
                        emit_path_joins(&ps.root, &ps.joins),
                    );
                    append_filter(&mut sql, &ps.filter);
                    sql
                })
                .collect();
            let mut sql = format!(
                "{}(SELECT \"r\" FROM (\n{}\n) AS \"_u\"",
                if *multi { "ARRAY" } else { "" },
                arms.join("\nUNION ALL\n")
            );
            if let Some(limit) = limit {
                sql.push_str(&format!("\nLIMIT {}", emit_expr(limit)));
            }
            sql.push(')');
            sql
        }

        IrExpr::ObjectPathSubquery(ps) => {
            let IrPathResult::Object {
                alias,
                type_name,
                shape,
            } = &ps.result
            else {
                unreachable!("an object path subquery always lands on an object")
            };
            let (pointer_exprs, _) = build_shape(shape, alias);
            let mut parts = vec![result_type_disc(ps, alias, type_name)];
            parts.extend(pointer_exprs);
            let mut sql = format!(
                "(SELECT (\n    {}\n)\nFROM {}",
                parts.join(",\n    "),
                emit_path_joins(&ps.root, &ps.joins),
            );
            append_filter(&mut sql, &ps.filter);
            append_order_by(&mut sql, &ps.order_by);
            append_offset_limit(&mut sql, &ps.offset, &ps.limit);
            sql.push(')');
            sql
        }

        IrExpr::ObjectSubquery(sel) => {
            let [IrRowSource::Bound { source, shape }] = sel.rows.as_slice() else {
                unreachable!("an object subquery is always schema-bound")
            };
            let alias = &source.alias;
            let (sub_exprs, _) = build_shape(shape, alias);
            let mut row_parts = vec![source_type_disc(source)];
            row_parts.extend(sub_exprs);
            let mut sql = format!(
                "(SELECT (\n        {}\n    )\n    FROM {} AS {}",
                row_parts.join(",\n        "),
                source_ref(source),
                qi(alias),
            );
            append_filter(&mut sql, &sel.filter);
            append_order_by(&mut sql, &sel.order_by);
            append_offset_limit(&mut sql, &sel.offset, &sel.limit);
            sql.push(')');
            sql
        }

        IrExpr::Subquery(sel) => {
            let [IrRowSource::Bound { source, shape }] = sel.rows.as_slice() else {
                unreachable!("scalar/exists subquery is always schema-bound")
            };
            let alias = &source.alias;
            let mut sql = if shape.is_empty() {
                // EXISTS inner: SELECT 1 FROM …
                format!("(SELECT 1\nFROM {} AS {}", source_ref(source), qi(alias))
            } else if let Some(c) = shape.iter().find_map(|f| {
                if let IrShapePointer::Computed(c) = f {
                    Some(c)
                } else {
                    None
                }
            }) {
                // Computed subquery (e.g. a type-intersection splat's
                // computed pointer): select the compiled expression itself,
                // not a bare column — the expression already references
                // `alias`'s own columns via ColumnRef.
                format!(
                    "(SELECT {}\nFROM {} AS {}",
                    emit_expr(&c.expr),
                    source_ref(source),
                    qi(alias)
                )
            } else {
                // Scalar subquery: SELECT alias.col FROM …
                let pk_col = shape
                    .iter()
                    .find_map(|f| {
                        if let IrShapePointer::Scalar(s) = f {
                            Some(s.column.as_str())
                        } else {
                            None
                        }
                    })
                    .unwrap_or("id");
                format!(
                    "(SELECT {}.{}\nFROM {} AS {}",
                    qi(alias),
                    qi(pk_col),
                    source_ref(source),
                    qi(alias),
                )
            };
            append_filter(&mut sql, &sel.filter);
            append_order_by(&mut sql, &sel.order_by);
            append_offset_limit(&mut sql, &sel.offset, &sel.limit);
            sql.push(')');
            sql
        }
        IrExpr::RawSql(s) => format!("({s})"),
    }
}

// ── Vector search ────────────────────────────────────────────────────────────

fn emit_vector_search(vs: &IrVectorSearch) -> SqlOutput {
    let alias = &vs.source.alias;
    let dist_sql = format!(
        "{}.{} {} {}",
        qi(alias),
        qi(&vs.vector_col),
        vs.distance_op,
        emit_expr(&vs.query_expr),
    );

    // Build the object sub-tuple.  If object_shape is empty, include all properties
    // (type disc + id implicitly come from build_shape when no elements provided;
    //  with no shape elements the shape is empty, so we fall back to "just id").
    // Use build_shape when there are explicit shape pointers; otherwise emit a minimal tuple.
    let (obj_tuple, object_shape_nodes) = if vs.object_shape.is_empty() {
        // No explicit shape: produce (type_disc, id) as minimum.
        let type_expr = type_disc(&vs.source.type_name);
        let id_expr = format!("{}.\"id\"", qi(alias));
        let tuple = format!("{},\n    {}", type_expr, id_expr);
        let id_node = ShapeNode::Scalar {
            name: "id".to_string(),
            position: 1,
        };
        (tuple, vec![id_node])
    } else {
        let (pointer_exprs, shape_pointers) = build_shape(&vs.object_shape, alias);
        let mut parts = vec![type_disc(&vs.source.type_name)];
        parts.extend(pointer_exprs);
        (parts.join(",\n    "), prepend_type(shape_pointers))
    };

    // Outer tuple: NULL (type slot), object sub-tuple at pos 1, distance at pos 2.
    let outer = format!(
        "NULL::text,\n    ROW(\n    {}\n    )::record,\n    {}",
        obj_tuple, dist_sql,
    );
    let mut sql = format!(
        "SELECT (\n    {}\n) AS result\nFROM {} AS {}",
        outer,
        source_ref(&vs.source),
        qi(alias),
    );
    append_filter(&mut sql, &vs.filter);

    // ORDER BY distance if requested.
    if let Some(dir) = &vs.order_by_distance {
        let dir_sql = match dir {
            IrSortDir::Asc => "ASC",
            IrSortDir::Desc => "DESC",
        };
        sql.push_str(&format!("\nORDER BY {} {}", dist_sql, dir_sql));
    }
    append_offset_limit(&mut sql, &vs.offset, &vs.limit);

    let object_node = ShapeNode::Object {
        name: "object".to_string(),
        type_name: Some(vs.source.type_name.clone()),
        position: 1,
        cardinality: Cardinality::Many,
        pointers: object_shape_nodes,
        has_implicit_id: shape_has_implicit_id(&vs.object_shape),
    };
    let shape = ShapeDescriptor {
        root: ShapeNode::VectorSearch {
            object_position: 1,
            distance_position: 2,
            object_node: Box::new(object_node),
        },
    };
    let inference_plan = vs.inference_model.as_ref().map(|model_name| InferencePlan::Embedding {
        model_name: model_name.clone(),
        type_name: vs.inference_type_name.clone().unwrap_or_default(),
        index_name: vs.inference_index_name.clone().unwrap_or(None),
        query_param_name: vs.inference_query_param_name.clone().unwrap_or_default(),
        query_literal: vs.inference_query_literal.clone(),
    });
    SqlOutput {
        sql,
        shape,
        inference_plan,
    }
}

// ── FTS search ───────────────────────────────────────────────────────────────

fn emit_fts_search(fs: &IrFtsSearch) -> SqlOutput {
    use crate::schema::SearchBackend;
    if fs.backend != SearchBackend::Postgres {
        return emit_fts_search_deferred(fs);
    }

    let alias = &fs.source.alias;
    let search_col = format!("{}.{}", qi(alias), qi(&fs.search_col));
    let query_sql = emit_expr(&fs.query_expr);
    let tsquery = format!("{}('english', {})", fs.tsquery_fn, query_sql);
    let rank_sql = format!("ts_rank({}, {})", search_col, tsquery);

    let (obj_tuple, object_shape_nodes) = if fs.object_shape.is_empty() {
        let type_expr = type_disc(&fs.source.type_name);
        let id_expr = format!("{}.\"id\"", qi(alias));
        let tuple = format!("{},\n    {}", type_expr, id_expr);
        let id_node = ShapeNode::Scalar {
            name: "id".to_string(),
            position: 1,
        };
        (tuple, vec![id_node])
    } else {
        let (pointer_exprs, shape_pointers) = build_shape(&fs.object_shape, alias);
        let mut parts = vec![type_disc(&fs.source.type_name)];
        parts.extend(pointer_exprs);
        (parts.join(",\n    "), prepend_type(shape_pointers))
    };

    let outer = format!(
        "NULL::text,\n    ROW(\n    {}\n    )::record,\n    {}",
        obj_tuple, rank_sql,
    );
    let mut sql = format!(
        "SELECT (\n    {}\n) AS result\nFROM {} AS {}\nWHERE {} @@ {}",
        outer,
        source_ref(&fs.source),
        qi(alias),
        search_col,
        tsquery,
    );
    if let Some(f) = &fs.filter {
        sql.push_str(&format!(" AND ({})", emit_expr(f)));
    }
    if let Some(dir) = &fs.order_by_rank {
        let dir_sql = match dir {
            IrSortDir::Asc => "ASC",
            IrSortDir::Desc => "DESC",
        };
        sql.push_str(&format!("\nORDER BY {} {}", rank_sql, dir_sql));
    }
    append_offset_limit(&mut sql, &fs.offset, &fs.limit);

    let object_node = ShapeNode::Object {
        name: "object".to_string(),
        type_name: Some(fs.source.type_name.clone()),
        position: 1,
        cardinality: Cardinality::Many,
        pointers: object_shape_nodes,
        has_implicit_id: shape_has_implicit_id(&fs.object_shape),
    };
    let shape = ShapeDescriptor {
        root: ShapeNode::FtsSearch {
            object_position: 1,
            rank_position: 2,
            object_node: Box::new(object_node),
        },
    };
    SqlOutput {
        sql,
        shape,
        inference_plan: None,
    }
}

fn emit_fts_search_deferred(fs: &IrFtsSearch) -> SqlOutput {
    let alias = &fs.source.alias;
    let ids_idx = fs
        .deferred_ids_param
        .expect("deferred_ids_param must be set for deferred backend");
    let scores_idx = fs
        .deferred_scores_param
        .expect("deferred_scores_param must be set for deferred backend");
    let ids_param = format!("${}", ids_idx + 1);
    let scores_param = format!("${}", scores_idx + 1);

    let (obj_tuple, object_shape_nodes) = if fs.object_shape.is_empty() {
        let type_expr = type_disc(&fs.source.type_name);
        let id_expr = format!("{}.\"id\"", qi(alias));
        let tuple = format!("{},\n    {}", type_expr, id_expr);
        let id_node = ShapeNode::Scalar {
            name: "id".to_string(),
            position: 1,
        };
        (tuple, vec![id_node])
    } else {
        let (pointer_exprs, shape_pointers) = build_shape(&fs.object_shape, alias);
        let mut parts = vec![type_disc(&fs.source.type_name)];
        parts.extend(pointer_exprs);
        (parts.join(",\n    "), prepend_type(shape_pointers))
    };

    let outer = format!(
        "NULL::text,\n    ROW(\n    {}\n    )::record,\n    \"_os\".\"score\"",
        obj_tuple,
    );
    let mut sql = format!(
        concat!(
            "SELECT (\n    {}\n) AS result\n",
            "FROM {} AS {}\n",
            "JOIN UNNEST({}::uuid[], {}::float8[]) AS \"_os\"(\"id\", \"score\")\n",
            "    ON \"_os\".\"id\" = {}.\"id\"",
        ),
        outer,
        source_ref(&fs.source),
        qi(alias),
        ids_param,
        scores_param,
        qi(alias),
    );
    if let Some(f) = &fs.filter {
        sql.push_str(&format!("\nWHERE ({})", emit_expr(f)));
    }
    if let Some(dir) = &fs.order_by_rank {
        let dir_sql = match dir {
            IrSortDir::Asc => "ASC",
            IrSortDir::Desc => "DESC",
        };
        sql.push_str(&format!("\nORDER BY \"_os\".\"score\" {}", dir_sql));
    }
    // limit/offset are passed to OpenSearch as size/from, not emitted in Postgres SQL.
    let size = fs.limit.as_ref().and_then(|lim| {
        if let IrExpr::Literal(IrLiteral::Int(n)) = lim {
            Some(*n as usize)
        } else {
            None
        }
    });

    let object_node = ShapeNode::Object {
        name: "object".to_string(),
        type_name: Some(fs.source.type_name.clone()),
        position: 1,
        cardinality: Cardinality::Many,
        pointers: object_shape_nodes,
        has_implicit_id: shape_has_implicit_id(&fs.object_shape),
    };
    let shape = ShapeDescriptor {
        root: ShapeNode::FtsSearch {
            object_position: 1,
            rank_position: 2,
            object_node: Box::new(object_node),
        },
    };
    let backend_str = match fs.backend {
        crate::schema::SearchBackend::Meilisearch => "meilisearch",
        _ => "opensearch",
    };
    let inference_plan = Some(InferencePlan::Search {
        backend: backend_str.to_string(),
        index_name: fs.deferred_index_name.clone().unwrap_or_default(),
        query_param_name: fs.deferred_query_param_name.clone().unwrap_or_default(),
        query_literal: fs.deferred_query_literal.clone(),
        size,
    });
    SqlOutput {
        sql,
        shape,
        inference_plan,
    }
}

// ── Function select ──────────────────────────────────────────────────────────

fn emit_function_select(sel: &IrFunctionSelect) -> SqlOutput {
    let alias = &sel.alias;
    let (pointer_exprs, shape_pointers) = build_shape(&sel.shape, alias);

    let type_expr = if sel.polymorphic {
        format!("{}.\"__type__\"", qi(alias))
    } else {
        type_disc(&sel.type_name)
    };
    let mut parts = vec![type_expr];
    parts.extend(pointer_exprs);
    let tuple = parts.join(",\n    ");
    let distinct = if sel.distinct { "DISTINCT " } else { "" };

    let args_sql = sel.fn_args.iter().map(emit_expr).collect::<Vec<_>>().join(", ");
    let fn_call = format!("{}.{}({})", pg_schema(&sel.fn_module), qi(&sel.fn_name), args_sql);

    let from_clause = if sel.polymorphic {
        // Polymorphic: we can't peek inside the function — assume it returns
        // a __type__ column since we control the DDL.  Use the fn call directly.
        format!("{} AS {}", fn_call, qi(alias))
    } else {
        format!("{} AS {}", fn_call, qi(alias))
    };

    let mut sql = format!("SELECT {}(\n    {}\n) AS result\nFROM {}", distinct, tuple, from_clause,);
    append_filter(&mut sql, &sel.filter);
    append_order_by(&mut sql, &sel.order_by);
    append_offset_limit(&mut sql, &sel.offset, &sel.limit);

    let root_pointers = prepend_type(shape_pointers);
    SqlOutput {
        sql,
        shape: ShapeDescriptor {
            root: ShapeNode::Object {
                name: String::new(),
                type_name: Some(sel.type_name.clone()),
                position: 0,
                cardinality: Cardinality::Many,
                pointers: root_pointers,
                has_implicit_id: shape_has_implicit_id(&sel.shape),
            },
        },
        inference_plan: None,
    }
}

/// Emit the SQL body expression for a user-defined function DDL.
///
/// For a bare scalar free select emits just the expression — e.g. `"a" + "b"`.
/// For everything else (object functions, non-scalar free selects) emits a
/// full `SELECT … FROM …` via `emit_dml_as_cte_source`.
pub fn emit_fn_body(ir: &crate::ir::IrOutput) -> String {
    with_subtype_fanouts(&ir.subtype_fanouts, || emit_output_fn_body(ir))
}

fn emit_output_fn_body(ir: &crate::ir::IrOutput) -> String {
    let body = match &ir.stmt {
        IrStmt::Select(sel) if matches!(sel.rows.as_slice(), [IrRowSource::Free(IrFreeExpr::Scalar(_))]) => {
            let IrRowSource::Free(IrFreeExpr::Scalar(e)) = &sel.rows[0] else {
                unreachable!()
            };
            format!("SELECT {}", emit_expr(e))
        }
        other => emit_dml_as_cte_source(other),
    };
    if ir.ctes.is_empty() {
        body
    } else {
        let cte_prefix = emit_cte_prefix(&ir.ctes);
        format!("{}{}", cte_prefix, body)
    }
}

fn emit_literal(lit: &IrLiteral) -> String {
    match lit {
        IrLiteral::Str(s) => sql_str(s),
        IrLiteral::Int(i) => i.to_string(),
        IrLiteral::Float(f) => {
            // Explicit ::float8 cast — a bare untyped numeral like `1.0`
            // defaults to Postgres `numeric`, but an un-cast PyQL float
            // literal means `float64`; only an explicit `<decimal>`/`123n`
            // literal should ever produce a real decimal. Without this,
            // `numeric`'s different wire OID
            // also broke decoding inside ROW() composites (see
            // _pg_decode_numeric in pylon/client.py, still needed for
            // genuine decimal casts).
            let s = f.to_string();
            let s = if s.contains('.') || s.contains('e') {
                s
            } else {
                format!("{}.0", s)
            };
            format!("({}::float8)", s)
        }
        IrLiteral::Bool(b) => {
            if *b {
                "TRUE".into()
            } else {
                "FALSE".into()
            }
        }
    }
}

// ── Tests ────────────────────────────────────────────────────────────────────

#[cfg(test)]
mod tests {
    use super::*;
    use crate::ir;
    use crate::parse;
    use crate::schema::{
        FunctionDescriptor, FunctionParamDescriptor, GlobalDescriptor, LinkDescriptor, MultiLinkDescriptor,
        NamedTupleDescriptor, PropertyDescriptor, SchemaDescriptor, TypeDescriptor,
    };

    #[test]
    fn a_template_argument_keeps_the_parameters_it_carries() {
        // The first argument carries query parameter `$2`; filled one number
        // at a time, it would be rewritten into the second argument.
        let filled = fill_template("f($1, $2, $10)", &["($2)::int".to_string(), "b".to_string()]);
        assert_eq!(filled, "f(($2)::int, b, $10)");
    }

    fn make_schema() -> SchemaDescriptor {
        SchemaDescriptor {
            types: vec![
                TypeDescriptor {
                    name: "Person".into(),
                    module: "default".into(),
                    table: "Person".into(),
                    abstract_: false,
                    materialized: false,
                    description: None,
                    parents: vec![],
                    interfaces: vec![],
                    bases: vec![],
                    properties: vec![
                        PropertyDescriptor {
                            name: "id".into(),
                            pg_type: "uuid".into(),
                            nullable: false,
                            default_sql: Some("uuidv7()".into()),
                            default_pyql: None,
                            description: None,
                            check_constraints: vec![],
                            is_exclusive: true,
                            is_pk: true,
                            is_readonly: true,
                            rewrites: vec![],
                            tuple_members: None,
                            column_type: None,
                        },
                        PropertyDescriptor {
                            name: "name".into(),
                            pg_type: "text".into(),
                            nullable: false,
                            default_sql: None,
                            default_pyql: None,
                            description: None,
                            check_constraints: vec![],
                            is_exclusive: false,
                            is_pk: false,
                            is_readonly: false,
                            rewrites: vec![],
                            tuple_members: None,
                            column_type: None,
                        },
                        PropertyDescriptor {
                            name: "age".into(),
                            pg_type: "int8".into(),
                            nullable: true,
                            default_sql: None,
                            default_pyql: None,
                            description: None,
                            check_constraints: vec![],
                            is_exclusive: false,
                            is_pk: false,
                            is_readonly: false,
                            rewrites: vec![],
                            tuple_members: None,
                            column_type: None,
                        },
                    ],
                    links: vec![LinkDescriptor {
                        name: "company".into(),
                        target: "default::Company".into(),
                        nullable: true,
                        through: None,
                        description: None,
                        default_pyql: None,
                        is_exclusive: false,
                        is_readonly: false,
                        rewrites: vec![],
                        on_delete: vec![],
                    }],
                    multilinks: vec![MultiLinkDescriptor {
                        name: "posts".into(),
                        target: "default::Post".into(),
                        through: None,
                        nullable: false,
                        description: None,
                        default_pyql: None,
                        on_delete: vec![],
                        is_exclusive: false,
                    }],
                    computed: vec![],
                    constraints: vec![],
                    indexes: vec![],
                    partition: None,
                    vector_indexes: vec![],
                    search_indexes: vec![],
                    triggers: vec![],
                    junction: false,
                    signals: vec![],
                },
                TypeDescriptor {
                    name: "Company".into(),
                    module: "default".into(),
                    table: "Company".into(),
                    abstract_: false,
                    materialized: false,
                    description: None,
                    parents: vec![],
                    interfaces: vec![],
                    bases: vec![],
                    properties: vec![PropertyDescriptor {
                        name: "name".into(),
                        pg_type: "text".into(),
                        nullable: false,
                        default_sql: None,
                        default_pyql: None,
                        description: None,
                        check_constraints: vec![],
                        is_exclusive: false,
                        is_pk: false,
                        is_readonly: false,
                        rewrites: vec![],
                        tuple_members: None,
                        column_type: None,
                    }],
                    links: vec![],
                    multilinks: vec![],
                    computed: vec![],
                    constraints: vec![],
                    indexes: vec![],
                    partition: None,
                    vector_indexes: vec![],
                    search_indexes: vec![],
                    triggers: vec![],
                    junction: false,
                    signals: vec![],
                },
                TypeDescriptor {
                    name: "Post".into(),
                    module: "default".into(),
                    table: "Post".into(),
                    abstract_: false,
                    materialized: false,
                    description: None,
                    parents: vec![],
                    interfaces: vec![],
                    bases: vec![],
                    properties: vec![PropertyDescriptor {
                        name: "title".into(),
                        pg_type: "text".into(),
                        nullable: false,
                        default_sql: None,
                        default_pyql: None,
                        description: None,
                        check_constraints: vec![],
                        is_exclusive: false,
                        is_pk: false,
                        is_readonly: false,
                        rewrites: vec![],
                        tuple_members: None,
                        column_type: None,
                    }],
                    links: vec![],
                    multilinks: vec![],
                    computed: vec![],
                    constraints: vec![],
                    indexes: vec![],
                    partition: None,
                    vector_indexes: vec![],
                    search_indexes: vec![],
                    triggers: vec![],
                    junction: false,
                    signals: vec![],
                },
            ],
            scalars: vec![],
            enums: vec![],
            named_tuples: vec![],
            globals: vec![],
            functions: vec![],
            aliases: vec![],
            channels: vec![],
            ..Default::default()
        }
    }

    fn compile_and_emit(query: &str) -> SqlOutput {
        let schema = make_schema();
        compile_and_emit_with(query, &schema)
    }

    fn compile_and_emit_with(query: &str, schema: &SchemaDescriptor) -> SqlOutput {
        let ast = parse::parse(query).expect("parse failed");
        let ir = ir::compile(&ast, schema).expect("IR compile failed");
        emit(&ir)
    }

    #[test]
    fn a_nested_read_inside_a_with_opens_no_second_with() {
        let out = compile_and_emit_with(
            "WITH a := (SELECT Person FILTER .age > 1), b := (SELECT Person LIMIT 1) SELECT (SELECT (a UNION b))",
            &make_schema(),
        );
        assert_eq!(out.sql.matches("WITH").count(), 1, "{}", out.sql);
    }

    #[test]
    fn a_comparison_against_a_multi_row_with_binding_tests_membership() {
        // Read as a scalar subquery, the binding aborts the query on its
        // second row: "more than one row returned by a subquery used as an
        // expression".
        let out = compile_and_emit_with(
            "WITH companies := (SELECT Company FILTER .name = 'x') SELECT Person FILTER .company = companies",
            &make_schema(),
        );
        assert!(
            out.sql.contains("= ANY((SELECT \"id\" FROM \"companies\"))"),
            "{}",
            out.sql
        );

        let out = compile_and_emit_with(
            "WITH companies := (SELECT Company) SELECT Person FILTER any(.company = companies)",
            &make_schema(),
        );
        assert!(
            out.sql.contains("= ANY((SELECT \"id\" FROM \"companies\"))"),
            "{}",
            out.sql
        );

        let out = compile_and_emit_with(
            "WITH companies := (SELECT Company) SELECT Person FILTER .company != companies",
            &make_schema(),
        );
        assert!(out.sql.contains("NOT (\"t1\".\"company_id\" = ANY("), "{}", out.sql);
    }

    #[test]
    fn a_comparison_against_a_single_row_with_binding_stays_a_scalar_read() {
        let out = compile_and_emit_with(
            "WITH company := (SELECT Company LIMIT 1) SELECT Person FILTER .company = company",
            &make_schema(),
        );
        assert!(out.sql.contains("= (SELECT \"id\" FROM \"company\")"), "{}", out.sql);
    }

    #[test]
    fn an_aggregate_over_a_backlink_counts_inside_a_subquery() {
        let out = compile_and_emit_with("SELECT Post { n := count(.<posts) }", &make_schema());
        assert!(
            !out.sql.contains("count((SELECT"),
            "an aggregate over a scalar subquery:\n{}",
            out.sql
        );
        assert!(out.sql.contains("count("), "{}", out.sql);
    }

    #[test]
    fn an_aggregate_over_a_relative_subselect_counts_inside_a_subquery() {
        let out = compile_and_emit_with(
            "SELECT Post { n := count((SELECT .<posts FILTER .age > 1)) }",
            &make_schema(),
        );
        assert!(
            !out.sql.contains("count((SELECT"),
            "an aggregate over a scalar subquery:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains(r#"(SELECT count("_s"."v") FROM unnest(ARRAY("#),
            "{}",
            out.sql
        );
    }

    #[test]
    fn any_over_a_link_type_check_tests_each_element() {
        let out = compile_and_emit_with(
            "SELECT Company { name } FILTER any(.<company.posts IS Post)",
            &make_schema(),
        );
        assert!(
            !out.sql.contains("bool_or(ARRAY"),
            "an aggregate over outer rows:\n{}",
            out.sql
        );
        assert!(
            out.sql
                .contains("FROM unnest(ARRAY(SELECT ('default::Post' = 'default::Post')"),
            "{}",
            out.sql
        );
    }

    #[test]
    fn a_link_type_check_reads_the_linked_object() {
        let out = compile_and_emit_with("SELECT Person { name } FILTER .company IS Company", &make_schema());
        assert!(!out.sql.contains("WHERE FALSE"), "{}", out.sql);
        assert!(out.sql.contains("'default::Company'"), "{}", out.sql);
    }

    /// `make_schema` with an `Admin` extending `Person`, in a table of its own.
    fn schema_with_a_subtype() -> SchemaDescriptor {
        let mut schema = make_schema();
        let mut admin = schema.types[0].clone();
        admin.name = "Admin".into();
        admin.table = "Admin".into();
        admin.bases = vec!["default::Person".into()];
        schema.types.push(admin);
        schema
    }

    #[test]
    fn a_type_with_subtypes_reads_their_rows_too() {
        let out = compile_and_emit_with("SELECT Person { name }", &schema_with_a_subtype());
        assert!(
            out.sql.contains("'default::Admin'::text AS \"__type__\"") && out.sql.contains("FROM \"public\".\"Admin\""),
            "expected the subtype's table in the read:\n{}",
            out.sql
        );
    }

    #[test]
    fn a_path_from_a_type_with_subtypes_reads_their_rows_too() {
        let out = compile_and_emit_with("SELECT Person.name", &schema_with_a_subtype());
        assert!(out.sql.contains("FROM \"public\".\"Admin\""), "{}", out.sql);
    }

    #[test]
    fn a_multilink_of_a_type_with_subtypes_reads_their_junctions_too() {
        let out = compile_and_emit_with("SELECT Person { posts: { title } }", &schema_with_a_subtype());
        assert!(
            out.sql.contains(
                "(SELECT \"source\", \"target\" FROM \"public\".\"Person.posts\" UNION ALL \
                 SELECT \"source\", \"target\" FROM \"public\".\"Admin.posts\")"
            ),
            "expected the junctions unioned:\n{}",
            out.sql
        );
    }

    #[test]
    fn is_a_supertype_matches_its_subtypes() {
        let out = compile_and_emit_with(
            "SELECT Person { name } FILTER Person IS Admin",
            &schema_with_a_subtype(),
        );
        assert!(out.sql.contains("\"__type__\" = 'default::Admin'"), "{}", out.sql);
    }

    #[test]
    fn a_write_to_a_type_with_subtypes_reaches_their_tables() {
        let out = compile_and_emit_with("UPDATE Person SET { age := 1 }", &schema_with_a_subtype());
        assert!(
            out.sql.contains("UPDATE \"public\".\"Person\"") && out.sql.contains("UPDATE \"public\".\"Admin\""),
            "expected both tables updated:\n{}",
            out.sql
        );
        assert!(
            !out.sql.contains("UPDATE (\n"),
            "a write never targets the union:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_comparing_a_multilink_itself_to_an_object() {
        // `filter any(.posts = post)` — the multi-link with no tail step. It
        // used to reach `compile_path`, which only knows properties, links
        // and computeds, and reported "has no link or property 'posts'".
        let out = compile_and_emit_with(
            "WITH p := (SELECT Post LIMIT 1) SELECT Person { id } FILTER any(.posts = p)",
            &make_schema(),
        );
        assert!(
            out.sql.contains("EXISTS((SELECT 1\nFROM \"public\".\"Person.posts\""),
            "expected a junction EXISTS:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("\"target\" = (SELECT \"id\" FROM \"p\")"),
            "expected the junction target compared to the bound object's id:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_comparing_a_multilink_itself_warns_without_any() {
        let ast = parse::parse("WITH p := (SELECT Post LIMIT 1) SELECT Person { id } FILTER .posts = p").unwrap();
        let ir = ir::compile(&ast, &make_schema()).unwrap();
        assert!(
            ir.warnings.iter().any(|w| w.contains("multi-link '.posts'")),
            "expected the set-valued FILTER warning, got: {:?}",
            ir.warnings
        );
    }

    #[test]
    fn test_detached_names_the_type_of_a_nested_select() {
        // `detached T` as an insert's link value: the prefix used to hide the
        // type name from the nested select's subject resolution, which then
        // reported "expected a type name as SELECT subject".
        let out = compile_and_emit_with(
            "INSERT Person { name := $n, company := (SELECT detached Company FILTER .name = $c LIMIT 1) }",
            &make_schema(),
        );
        assert!(
            out.sql.contains("INSERT INTO \"public\".\"Person\""),
            "expected the insert to compile:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("FROM \"public\".\"Company\""),
            "expected the detached select to read the Company table:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_select_over_an_update_on_a_with_binding() {
        // A bare `with` binding as an update's subject names rows, not a
        // type. The top-level form already knew that; wrapping it in a select
        // asked what type the statement yields and got the binding's name,
        // reported as "unknown type 'p'".
        let out = compile_and_emit_with(
            "WITH p := (SELECT Person FILTER .id = $a LIMIT 1) SELECT (UPDATE p SET { name := $n }) { id }",
            &make_schema(),
        );
        assert!(
            out.sql.contains("UPDATE \"public\".\"Person\""),
            "expected the update to resolve to the bound type's table:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_mutation_as_a_free_object_field() {
        // Already allowed as an item of a free *set*; a named field of a free
        // object is the same question and used to answer "an update cannot
        // stand in for a value".
        let out = compile_and_emit_with(
            "SELECT { renamed := (UPDATE Person FILTER .id = $a SET { name := $n }), n := 1 }",
            &make_schema(),
        );
        assert!(
            out.sql.contains("UPDATE \"public\".\"Person\""),
            "expected the mutation to become a data-modifying CTE:\n{}",
            out.sql
        );
        assert!(
            out.sql
                .contains("(SELECT \"t1\".\"id\"\nFROM \"_nested_dml_0\" AS \"t1\") AS \"_f0\""),
            "expected the field to carry the rows the mutation touched:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_sub_selects_filter_stays_on_its_own_subject_when_a_field_chain_projects_off_it() {
        // `(select Person filter .id = $a).company.name` filters People, not
        // Companies. The merge that splices the field chain onto the inner
        // select used to leave the filter to bind against the last step.
        let out = compile_and_emit_with("SELECT (SELECT Person FILTER .name = $n).company.name", &make_schema());
        assert!(
            out.sql.contains("WHERE (\"t0\".\"name\" = $1)"),
            "the inner filter belongs to the Person alias:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_reading_back_a_nested_insert_through_its_own_statement() {
        let out = compile_and_emit_with(
            "SELECT (INSERT Person { name := $n, posts := (INSERT Post { title := $t }) }) { id, posts: { title } }",
            &make_schema(),
        );
        assert!(out.sql.contains("__ml_add_0"), "got:\n{}", out.sql);
    }

    #[test]
    fn test_reading_back_a_nested_insert_within_its_own_statement() {
        // Postgres does not show one statement's CTE writes to the rest of
        // that statement, so reading the link back off the base tables found
        // nothing and it hydrated as empty/None (confirmed live).
        let out = compile_and_emit_with(
            "SELECT (INSERT Person { name := $n, posts := (INSERT Post { title := $t }) }) \
             { id, posts: { title } }",
            &make_schema(),
        );
        assert!(
            out.sql.contains("FROM \"_dml__ml_add_0\" AS \"jt\""),
            "the junction rows must come from the CTE that wrote them:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("\"_nested_dml_0\""),
            "and the targets from the nested insert's own CTE:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_nested_single_link_insert_reads_back_from_its_cte() {
        let out = compile_and_emit_with(
            "SELECT (INSERT Person { name := $n, company := (INSERT Company { name := $c }) }) \
             { id, company: { name } }",
            &make_schema(),
        );
        assert!(
            out.sql.contains("FROM \"_nested_dml_0\""),
            "the linked row must be read from the CTE that inserted it:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_free_object_field_holding_an_object_keeps_it_an_object() {
        // Gel compiles a free shape into a real object type whose fields are
        // real pointers, so an object field stays an object. Pylon used to
        // build the free object as jsonb, which flattened it to the bare id.
        let out = compile_and_emit_with(
            "WITH c := (SELECT Company LIMIT 1) SELECT { co := c { name }, n := 1 }",
            &make_schema(),
        );
        assert!(
            out.sql.contains("'default::Company'::text"),
            "the field must carry the object's own row:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_projecting_an_object_off_a_free_object() {
        // `.co` is a path step through a pointer, not jsonb extraction, so
        // what comes back is the object with its shape.
        let out = compile_and_emit_with(
            "WITH c := (SELECT Company LIMIT 1) SELECT { co := c { name }, n := 1 }.co",
            &make_schema(),
        );
        assert!(
            !out.sql.contains("jsonb_build_object"),
            "projecting a pointer must not go through jsonb:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("SELECT v AS result, v FROM"),
            "the object row is already the result row:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_shape_on_a_single_valued_object_path() {
        // `c := .company { name }` — an object route existed only for walks
        // that crossed a multi step, so a single-valued one fell through to
        // expression position and reported "shapes and set literals are not
        // valid in expression context".
        let out = compile_and_emit_with("SELECT Person { c := .company { name } }", &make_schema());
        assert!(
            out.sql.contains("'default::Company'::text"),
            "the pointer must carry the object's own row:\n{}",
            out.sql
        );
        assert!(
            !out.sql.contains("ARRAY(SELECT"),
            "a single-valued walk is one object, not an array of one:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_shapeless_single_link_still_reads_as_its_value() {
        // The object route is taken only when a shape says an object was
        // meant; without one this stays what it always was.
        let out = compile_and_emit_with("SELECT Person { c := .company }", &make_schema());
        assert!(
            !out.sql.contains("'default::Company'::text"),
            "a bare link reference must not grow an object row:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_tuple_element_holding_a_shaped_object() {
        // `select (c { name }, 1)` — an element is the same question a free
        // object's field asks, and used to report "shapes and set literals
        // are not valid in expression context".
        let out = compile_and_emit_with(
            "WITH c := (SELECT Company LIMIT 1) SELECT (c { name }, 1)",
            &make_schema(),
        );
        assert!(
            out.sql.contains("'default::Company'::text"),
            "the element must carry the object's own row:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_an_object_returning_function_as_a_link_value() {
        // `created_by := account_of_transaction()` — the function names the
        // row to link to, so its key is what gets stored. The schema itself
        // writes this as a link default; written out in a query it used to be
        // refused as "part of a larger expression".
        let mut schema = make_schema();
        // make_schema's Company carries only a name; a link's value is its key.
        let company = schema
            .types
            .iter_mut()
            .find(|t| t.name == "Company")
            .expect("make_schema declares Company");
        company.properties.push(PropertyDescriptor {
            name: "id".into(),
            pg_type: "uuid".into(),
            nullable: false,
            default_sql: None,
            default_pyql: None,
            description: None,
            check_constraints: vec![],
            is_exclusive: true,
            is_pk: true,
            is_readonly: true,
            rewrites: vec![],
            tuple_members: None,
            column_type: None,
        });
        schema.functions.push(crate::schema::FunctionDescriptor {
            name: "current_company".into(),
            module: "default".into(),
            params: vec![],
            return_pg_type: "default::Company".into(),
            return_is_object: true,
            return_is_set: false,
            return_is_polymorphic: false,
            volatility: "stable".into(),
            body: "select Company limit 1".into(),
        });
        let out = compile_and_emit_with("INSERT Person { name := $n, company := current_company() }", &schema);
        assert!(
            out.sql.contains("FROM \"public\".\"current_company\"()"),
            "the function must supply the foreign key:\n{}",
            out.sql
        );
    }

    /// `make_schema` plus a `Company`-returning function and a `Person.company`
    /// link defaulting to it — jaldis's `Creatable.created_by :=
    /// account_of_transaction()` in miniature.
    fn make_schema_with_an_object_returning_default() -> SchemaDescriptor {
        let mut schema = make_schema();
        let company = schema
            .types
            .iter_mut()
            .find(|t| t.name == "Company")
            .expect("make_schema declares Company");
        company.properties.push(PropertyDescriptor {
            name: "id".into(),
            pg_type: "uuid".into(),
            nullable: false,
            default_sql: None,
            default_pyql: None,
            description: None,
            check_constraints: vec![],
            is_exclusive: true,
            is_pk: true,
            is_readonly: true,
            rewrites: vec![],
            tuple_members: None,
            column_type: None,
        });
        schema.functions.push(crate::schema::FunctionDescriptor {
            name: "current_company".into(),
            module: "default".into(),
            params: vec![],
            return_pg_type: "default::Company".into(),
            return_is_object: true,
            return_is_set: false,
            return_is_polymorphic: false,
            volatility: "stable".into(),
            body: "select Company limit 1".into(),
        });
        let person = schema.types.iter_mut().find(|t| t.name == "Person").unwrap();
        let company_link = person.links.iter_mut().find(|l| l.name == "company").unwrap();
        company_link.default_pyql = Some("current_company()".into());
        schema
    }

    #[test]
    fn an_object_returning_link_default_is_applied_by_the_insert() {
        // A column DEFAULT cannot hold this one — PostgreSQL evaluates it with
        // no query in scope — so the insert has to carry it, the way Gel's
        // `_gen_pointers_from_defaults` expands every default into the shape.
        let schema = make_schema_with_an_object_returning_default();
        let out = compile_and_emit_with("INSERT Person { name := $n }", &schema);
        assert!(
            out.sql.contains("\"company_id\"") && out.sql.contains("current_company"),
            "the default must supply the foreign key:\n{}",
            out.sql
        );
    }

    #[test]
    fn an_explicit_value_wins_over_an_inlined_default() {
        let schema = make_schema_with_an_object_returning_default();
        let out = compile_and_emit_with("INSERT Person { name := $n, company := {} }", &schema);
        assert!(
            !out.sql.contains("current_company"),
            "the shape named the pointer, so the default does not apply:\n{}",
            out.sql
        );
    }

    #[test]
    fn an_object_returning_link_default_gets_no_column_default() {
        // Emitting one produces DDL PostgreSQL refuses to run; it used to be
        // dropped silently here, leaving the pointer with no default at all.
        let schema = make_schema_with_an_object_returning_default();
        let ddl = crate::export::export_schema(&schema).unwrap();
        assert!(
            !ddl.contains("\"company_id\" uuid DEFAULT"),
            "the column must carry no default:\n{ddl}"
        );
        crate::validate::validate_schema_types(&schema).expect("an inlined default is not an error");
    }

    #[test]
    fn a_column_expressible_default_stays_in_the_ddl() {
        // Only what a column DEFAULT cannot hold moves into the insert.
        let mut schema = make_schema();
        let person = schema.types.iter_mut().find(|t| t.name == "Person").unwrap();
        let age = person.properties.iter_mut().find(|p| p.name == "age").unwrap();
        age.default_pyql = Some("21".into());
        assert!(
            crate::ir::inlined_pointer_defaults(schema.types.iter().find(|t| t.name == "Person").unwrap(), &schema)
                .is_empty()
        );
        let ddl = crate::export::export_schema(&schema).unwrap();
        assert!(ddl.contains("DEFAULT 21"), "{ddl}");
        let out = compile_and_emit_with("INSERT Person { name := $n }", &schema);
        assert!(
            !out.sql.contains("21"),
            "the column DEFAULT still applies it:\n{}",
            out.sql
        );
    }

    /// `make_schema` plus an `Employee` extending `Person`, where `Person`
    /// declares a computed naming itself — jaldis's
    /// `BrandAddon.bundle := (BrandAddon is BrandAddonBundle)` in miniature.
    fn make_schema_with_an_inherited_self_naming_computed() -> SchemaDescriptor {
        let mut schema = make_schema();
        let person = schema.types.iter_mut().find(|t| t.name == "Person").unwrap();
        person.computed.push(crate::schema::ComputedDescriptor {
            name: "own_name".into(),
            expression: "default::Person.name".into(),
            return_type: Some("text".into()),
            link_target: None,
            link_multi: false,
        });
        let mut employee = person.clone();
        employee.name = "Employee".into();
        employee.table = "Employee".into();
        employee.bases = vec!["default::Person".into()];
        schema.types.push(employee);
        schema
    }

    #[test]
    fn an_inherited_computed_naming_its_declaring_type_still_means_the_row() {
        // Gel compiles a computed once, against the type that declares it, so
        // the name reaches the subject on every subtype that inherits it.
        // Pylon compiles it again per subtype, where the name matched nothing
        // and the reference read the whole table instead.
        let schema = make_schema_with_an_inherited_self_naming_computed();
        let out = compile_and_emit_with("SELECT Employee { own_name }", &schema);
        assert!(
            !out.sql.contains("\"Person\""),
            "the computed must read the Employee row, not Person's table:\n{}",
            out.sql
        );
        crate::validate::validate_schema_types(&schema).expect("an inherited computed is single-valued");
    }

    #[test]
    fn an_inherited_type_check_computed_tests_the_row_it_is_read_on() {
        // jaldis's own: `BrandAddon.bundle := (BrandAddon is BrandAddonBundle)`,
        // read on `BrandAddonBundle`. `is` had a same-scope check of its own
        // that only knew the two spellings of the type being compiled, so on
        // the subtype it read `BrandAddon` as a separate set — many values
        // behind a declared `bool`.
        let mut schema = make_schema();
        let person = schema.types.iter_mut().find(|t| t.name == "Person").unwrap();
        person.computed.push(crate::schema::ComputedDescriptor {
            name: "employed".into(),
            expression: "(default::Person is default::Employee)".into(),
            return_type: Some("boolean".into()),
            link_target: None,
            link_multi: false,
        });
        let mut employee = person.clone();
        employee.name = "Employee".into();
        employee.table = "Employee".into();
        employee.bases = vec!["default::Person".into()];
        schema.types.push(employee);

        crate::validate::validate_schema_types(&schema).expect("the check is single-valued on both types");
        let out = compile_and_emit_with("SELECT Employee { employed }", &schema);
        assert!(
            !out.sql.contains("\"Person\""),
            "the check must read the Employee row's own type:\n{}",
            out.sql
        );
    }

    #[test]
    fn a_shape_writing_the_computed_out_itself_does_not_get_the_anchor() {
        // The rule Gel actually has is narrow: only a *declared* computed is
        // compiled against its declaring type. Naming a supertype anywhere
        // else is that supertype's own set — confirmed against Gel, where
        // `select BrandAddonBundle { n := count(BrandAddon) }` is the full
        // count while the same shape naming `BrandAddonBundle` is 1. So the
        // same expression written out in a shape keeps resolving on its own,
        // even where it shadows the inherited computed's name.
        let schema = make_schema_with_an_inherited_self_naming_computed();
        let ast = parse::parse("SELECT Employee { own_name := default::Person.name }").expect("parse failed");
        let Err(error) = ir::compile(&ast, &schema) else {
            panic!("an absolute path has nothing to anchor to here")
        };
        assert!(
            format!("{error:?}").contains("absolute paths are not valid"),
            "{error:?}"
        );
    }

    #[test]
    fn a_default_that_compiles_nowhere_is_still_an_error() {
        // jaldis carried `sequence_next(INTROSPECT marketplace::OrderNo)`, Gel
        // syntax Pylon cannot parse. Inlining must not turn that into silence.
        let mut schema = make_schema();
        let person = schema.types.iter_mut().find(|t| t.name == "Person").unwrap();
        let age = person.properties.iter_mut().find(|p| p.name == "age").unwrap();
        age.default_pyql = Some("sequence_next(INTROSPECT default::AgeSeq)".into());
        let errs = crate::validate::validate_schema_types(&schema).unwrap_err();
        let (_, msg, _) = errs[0].class_name_message_position();
        assert!(msg.contains("default"), "{msg}");
    }

    #[test]
    fn test_comparing_a_link_with_an_object_returning_function() {
        // Comparing objects compares identity, so the call stands for the id
        // of the row it returns. Left whole it is a function call with nowhere
        // to go but the subject of a select.
        let mut schema = make_schema();
        // make_schema's Company carries only a name; identity needs its key.
        let company = schema
            .types
            .iter_mut()
            .find(|t| t.name == "Company")
            .expect("make_schema declares Company");
        company.properties.push(PropertyDescriptor {
            name: "id".into(),
            pg_type: "uuid".into(),
            nullable: false,
            default_sql: None,
            default_pyql: None,
            description: None,
            check_constraints: vec![],
            is_exclusive: true,
            is_pk: true,
            is_readonly: true,
            rewrites: vec![],
            tuple_members: None,
            column_type: None,
        });
        schema.functions.push(crate::schema::FunctionDescriptor {
            name: "current_company".into(),
            module: "default".into(),
            params: vec![],
            return_pg_type: "default::Company".into(),
            return_is_object: true,
            return_is_set: false,
            return_is_polymorphic: false,
            volatility: "stable".into(),
            body: "select Company limit 1".into(),
        });
        let out = compile_and_emit_with("SELECT Person FILTER .company = current_company()", &schema);
        assert!(
            out.sql.contains("\"company_id\"") && out.sql.contains("current_company"),
            "the link's key against the function's row:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_choosing_between_two_object_sets() {
        // `A if cond else B` over objects is not the scalar CASE an if/else
        // over values gives: it stands for A's rows when the condition holds
        // and B's when it does not.
        let out = compile_and_emit_with(
            "WITH a := (SELECT Person FILTER .name = $n), b := (SELECT Person) \
             SELECT (a IF EXISTS a ELSE b) { name }",
            &make_schema(),
        );
        assert!(
            out.sql.contains("UNION ALL"),
            "expected one branch per side:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_union_operand_written_inline() {
        // A union reads one relation per branch, so an operand that names no
        // relation yet -- a sub-select, or an insert in a get-or-create -- is
        // hoisted into a CTE of its own first.
        let out = compile_and_emit_with(
            "SELECT ((SELECT Person FILTER .name = $n) UNION (SELECT Person)) { name }",
            &make_schema(),
        );
        assert!(
            out.sql.starts_with("WITH"),
            "expected the operands hoisted:\n{}",
            out.sql
        );
        assert!(out.sql.contains("UNION ALL"), "{}", out.sql);
    }

    #[test]
    fn test_a_for_loop_variable_over_objects() {
        // The variable holds the row's key, so a path off it reads that row
        // back. It used to be treated as a scalar throughout, and any path
        // rooted at one reported "unknown type".
        let out = compile_and_emit_with("FOR p IN (SELECT Person) UNION (SELECT p.name)", &make_schema());
        assert!(
            out.sql.contains("FROM \"public\".\"Person\"") && out.sql.contains("\"_for_p\".\"v\""),
            "the body must read the row the variable names:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_shape_on_a_for_loop_variable() {
        let out = compile_and_emit_with("FOR p IN (SELECT Person) UNION (SELECT p { name })", &make_schema());
        assert!(
            out.sql.contains("'default::Person'::text"),
            "a shape on the variable yields its object:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_for_loop_variable_over_values_is_still_a_scalar() {
        let out = compile_and_emit_with("FOR n IN {1, 2} UNION (SELECT n)", &make_schema());
        assert!(out.sql.contains("VALUES"), "{}", out.sql);
    }

    #[test]
    fn test_exists_over_a_sub_select_on_a_relative_path() {
        // `exists ((select .posts filter …))` — the sub-select is relative to
        // the enclosing object, so it needs the same rooting and correlation
        // `count((select .posts))` gets. Compiled without them it resolved in
        // free context and called the pointer unknown.
        let out = compile_and_emit_with(
            "SELECT Person { id } FILTER EXISTS ((SELECT .posts FILTER .title = $t))",
            &make_schema(),
        );
        assert!(out.sql.contains("EXISTS("), "{}", out.sql);
        assert!(
            out.sql.contains("\"Person.posts\""),
            "the junction join must survive into the EXISTS:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("\"t0\".\"id\""),
            "and it must correlate to the enclosing row:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_counting_the_rows_a_mutation_touched() {
        // `count((delete X filter …))` — the mutation runs as its own
        // data-modifying CTE and the aggregate reads it back, rather than
        // being refused as "a delete cannot stand in for a value".
        let out = compile_and_emit_with("SELECT count((DELETE Person FILTER .name = $n))", &make_schema());
        assert!(out.sql.starts_with("WITH"), "{}", out.sql);
        assert!(out.sql.contains("DELETE FROM \"public\".\"Person\""), "{}", out.sql);
        assert!(out.sql.contains("count(*)"), "{}", out.sql);
    }

    #[test]
    fn test_a_conflict_target_naming_several_columns() {
        // `unless conflict on (.a, .b)` names columns for Postgres to infer
        // the arbiter index from; emitted as one composite value it would
        // look for an expression index on the tuple instead.
        let out = compile_and_emit_with(
            "INSERT Person { name := $n } UNLESS CONFLICT ON (.name, .id) ELSE (UPDATE Person SET { name := $n })",
            &make_schema(),
        );
        assert!(
            out.sql.contains("ON CONFLICT (\"name\", \"id\")"),
            "expected a column list:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_for_loop_over_a_path() {
        // `for post in p.posts union (…)` — a walk off a binding names a set
        // to iterate just as a sub-select does, only written without the
        // parentheses.
        let out = compile_and_emit_with(
            "WITH p := (SELECT Person LIMIT 1) FOR post IN p.posts UNION (SELECT post.title)",
            &make_schema(),
        );
        assert!(out.sql.contains("CROSS JOIN LATERAL"), "{}", out.sql);
    }

    #[test]
    fn test_a_relative_sub_select_keeps_its_enclosing_object() {
        // `assert_exists((select .posts { title }))` and `posts := (select
        // .posts filter …)` in an update both compiled their sub-select with
        // no enclosing object, so the relative path resolved in free context
        // and the pointer came back unknown.
        let out = compile_and_emit_with(
            "SELECT Person { ps := assert_exists((SELECT .posts { title })) }",
            &make_schema(),
        );
        assert!(out.sql.contains("assert_exists"), "{}", out.sql);
        assert!(
            out.sql.contains("\"Person.posts\""),
            "the walk must reach the junction:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_relative_sub_select_as_a_multi_link_value() {
        let out = compile_and_emit_with(
            "UPDATE Person FILTER .name = $n SET { posts := (SELECT .posts FILTER .title = $t) }",
            &make_schema(),
        );
        assert!(
            out.sql.contains("\"Person.posts\""),
            "the value must read the row's own posts:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_shape_written_after_a_sub_select() {
        // `(select Post filter …) { title }` means the same as putting the
        // shape inside the parentheses, so the shape is pushed onto the inner
        // statement's own result rather than refused.
        let out = compile_and_emit_with(
            "SELECT Person { ps := (SELECT Post FILTER .title = $t) { title } }",
            &make_schema(),
        );
        assert!(
            out.sql.contains("'default::Post'::text"),
            "the pointer must carry the object rows:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_shape_after_a_sub_select_with_a_with_block() {
        let out = compile_and_emit_with(
            "SELECT Person { ps := (WITH t := $t SELECT Post FILTER .title = t) { title } }",
            &make_schema(),
        );
        assert!(out.sql.contains("'default::Post'::text"), "{}", out.sql);
    }

    #[test]
    fn test_a_computed_declared_as_an_object_returning_call() {
        // `recent := recent()` — the rows the function yields are the
        // pointer's objects. Read as a plain expression the call was "part of
        // a larger expression", which such a function refuses, so a splat over
        // a type carrying one could not compile at all.
        let mut schema = make_schema();
        schema.functions.push(crate::schema::FunctionDescriptor {
            name: "recent".into(),
            module: "default".into(),
            params: vec![],
            return_pg_type: "default::Post".into(),
            return_is_object: true,
            return_is_set: true,
            return_is_polymorphic: false,
            volatility: "stable".into(),
            body: "select Post".into(),
        });
        schema.types[0].computed.push(crate::schema::ComputedDescriptor {
            name: "recent".into(),
            expression: "default::recent()".into(),
            return_type: None,
            link_target: None,
            link_multi: false,
        });
        let out = compile_and_emit_with("SELECT Person { recent }", &schema);
        assert!(
            out.sql.contains("ARRAY(SELECT") && out.sql.contains("'default::Post'::text"),
            "the pointer must carry the function's rows:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_one_branch_of_an_object_if_else_is_empty() {
        // `{}` contributes no rows, so Gel's own rewrite (`SELECT A WHERE
        // Cond UNION ALL SELECT B WHERE NOT Cond`) reduces to the other
        // branch under its guard.
        let out = compile_and_emit_with(
            "WITH p := (SELECT Person LIMIT 1) SELECT p { name } IF EXISTS p ELSE {}",
            &make_schema(),
        );
        assert!(!out.sql.contains("UNION ALL"), "nothing to union with:\n{}", out.sql);
        assert!(out.sql.contains("'default::Person'::text"), "{}", out.sql);
    }

    #[test]
    fn test_a_conditional_insert_is_guarded_by_its_condition() {
        // `(insert …) if cond else {}` — guarding a reader would not do: the
        // insert is a data-modifying CTE that Postgres runs regardless, so
        // the condition goes into the insert itself and it writes nothing
        // when false. Verified live: running it twice leaves one row.
        let out = compile_and_emit_with("SELECT (INSERT Person { name := $n }) IF FALSE ELSE {}", &make_schema());
        assert!(
            out.sql.contains("SELECT") && out.sql.contains("WHERE"),
            "a guarded insert selects its row rather than listing values:\n{}",
            out.sql
        );
        assert!(!out.sql.contains("VALUES"), "VALUES always yields a row:\n{}", out.sql);
    }

    #[test]
    fn test_a_conditional_update_carries_its_condition() {
        // Guarding the branch only filters what is read back -- the mutation
        // is its own CTE and Postgres runs it regardless. An update has a
        // WHERE of its own to fold the condition into, which is what keeps it
        // from running when the condition is false.
        let out = compile_and_emit(
            "SELECT (UPDATE Person FILTER .name = $n SET { name := $m }) IF EXISTS (SELECT Company) ELSE {}",
        );
        let update = out.sql.find("UPDATE").expect("an update is emitted");
        let where_clause = out.sql[update..].find("WHERE").expect("the update is filtered");
        assert!(
            out.sql[update + where_clause..].contains("EXISTS"),
            "the condition narrows the rows the update touches:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_conditional_delete_carries_its_condition() {
        // Guarding the branch would only filter what is read back — Postgres
        // runs a data-modifying CTE regardless — so the condition has to be
        // ANDed into the DELETE's own WHERE, as it already is for an update.
        let out = compile_and_emit("SELECT (DELETE Person FILTER .name = $n) IF FALSE ELSE {}");
        let delete = out
            .sql
            .split("DELETE FROM")
            .nth(1)
            .expect("the delete should still be emitted");
        let where_clause = delete.split("WHERE").nth(1).expect("the delete should be filtered");
        assert!(
            where_clause.contains("FALSE") || where_clause.contains("false"),
            "the condition must narrow the delete itself:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_pointer_a_binding_declared_in_its_own_shape() {
        // `p := (select Person { boss := … })` then `p { boss }` — the
        // pointer is on no type, so reading it back has to find it through
        // the binding it was written on, and compile it against that row.
        let out = compile_and_emit_with(
            "WITH p := (SELECT Person { co := .company { name } } LIMIT 1) SELECT p { name, co: { name } }",
            &make_schema(),
        );
        assert!(
            out.sql.contains("'default::Company'::text"),
            "the declared pointer must be inlined:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_binding_declared_pointer_read_from_a_tuple_element() {
        let out = compile_and_emit_with(
            "WITH p := (SELECT Person { co := .company { name } } LIMIT 1) SELECT (p { co: { name } }, 1)",
            &make_schema(),
        );
        assert!(out.sql.contains("'default::Company'::text"), "{}", out.sql);
    }

    #[test]
    fn test_a_for_loop_variable_shaped_inside_a_tuple() {
        // The variable holds one row's key, so the element has to be narrowed
        // to it. Unfiltered it read the whole table and answered with rows
        // the loop never named -- wrong, and silently so.
        let out = compile_and_emit_with(
            "WITH p := (SELECT Person LIMIT 1) FOR q IN p.posts UNION (SELECT (a := q { title }, b := 1))",
            &make_schema(),
        );
        assert!(
            out.sql.contains("\"_for_q\".\"v\""),
            "the element must be narrowed to the loop's row:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_choosing_between_two_walks() {
        // `(select p.posts …) if cond else (select p.posts …)` — a walk names
        // no type of its own, so neither branch could be resolved.
        let out = compile_and_emit_with(
            "WITH p := (SELECT Person LIMIT 1), q := (SELECT p.posts LIMIT 1) IF TRUE ELSE (SELECT p.posts) \
             SELECT q { title }",
            &make_schema(),
        );
        assert!(out.sql.contains("UNION ALL"), "one branch per side:\n{}", out.sql);
    }

    #[test]
    fn test_a_bare_coalesce_of_correlated_walks_reaching_many_objects() {
        // Without a shape the walks stand for the ids they reach, which is
        // still a set -- SQL's COALESCE takes one value per side.
        let out = compile_and_emit_with(
            "SELECT Post { owners := .<posts[is Person] ?? .<posts[is Person] }",
            &make_schema(),
        );
        assert!(
            !out.sql.contains("COALESCE("),
            "not a value-by-value choice:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("ARRAY(SELECT \"r\" FROM ("),
            "the arms are aggregated:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_union_of_correlated_walks_reaching_many() {
        // Without the LIMIT the union stands for every object both walks
        // reach, so its arms have to be aggregated. Read as a lone record it
        // compiled fine and failed in PostgreSQL on the second row, with
        // "more than one row returned by a subquery used as an expression".
        let out = compile_and_emit_with(
            "SELECT Post { owners := (.<posts[is Person] UNION .<posts[is Person]) { name } }",
            &make_schema(),
        );
        assert!(
            out.sql.contains("ARRAY(SELECT \"r\" FROM ("),
            "the arms are aggregated:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_coalesce_of_correlated_walks_reaching_many() {
        // `([is A].links ?? [is B].links) { … }` — the same set a union of
        // them gives, so it aggregates the same way.
        let out = compile_and_emit_with(
            "SELECT Post { owners := (.<posts[is Person] ?? .<posts[is Person]) { name } }",
            &make_schema(),
        );
        assert!(
            out.sql.contains("ARRAY(SELECT \"r\" FROM ("),
            "the arms are aggregated:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_union_of_correlated_walks() {
        // `(.<a[is T] union .<b[is T]) { id }` — each operand hangs off the
        // enclosing row, so none can be hoisted into a CTE the way a
        // standalone union's operands are; they are read as one set instead.
        let out = compile_and_emit_with(
            "SELECT Post { owners := (SELECT (.<posts[is Person] UNION .<posts[is Person]) { name } LIMIT 1) }",
            &make_schema(),
        );
        assert!(out.sql.contains("UNION ALL"), "one arm per operand:\n{}", out.sql);
        assert!(
            out.sql.contains("AS \"_u\""),
            "the arms are read as one set:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_for_loop_whose_body_updates() {
        // Driven from the iteration with `UPDATE … FROM`: the select-bodied
        // form uses a LATERAL, which cannot hold DML. Verified live -- three
        // rows iterated, three rows updated.
        let out = compile_and_emit_with(
            "FOR p IN (SELECT Person) UNION (UPDATE Person FILTER .id = p.id SET { name := $n })",
            &make_schema(),
        );
        assert!(out.sql.contains("UPDATE \"public\".\"Person\""), "{}", out.sql);
        assert!(
            out.sql.contains("FROM \"_for_p\""),
            "driven from the iteration:\n{}",
            out.sql
        );
        assert!(
            !out.sql.contains("LATERAL"),
            "DML cannot sit in a LATERAL:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_for_loop_body_appends_the_row_it_is_iterating() {
        let out = compile_and_emit(
            "WITH t := (SELECT Post) FOR q IN t UNION (UPDATE Person FILTER .id = $i SET { posts += q })",
        );
        // One junction row per iteration, taken from the iteration itself --
        // not every Post, which an uncorrelated value subquery would append.
        assert!(
            out.sql.contains("\"_iter\"") && out.sql.contains("FROM \"_ids\""),
            "junction rows driven from the iteration:\n{}",
            out.sql
        );
        assert!(
            !out.sql.contains("LATERAL"),
            "DML cannot sit in a LATERAL:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_select_on_a_for_loop_variable_reads_only_its_row() {
        let out = compile_and_emit("FOR p IN (SELECT Person) UNION (SELECT (SELECT p) { name })");
        assert!(
            out.sql.contains("= \"_for_p\".\"v\""),
            "narrowed to the row the variable holds:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_walk_off_a_for_loop_variable_reads_the_binding() {
        // The rows a sibling CTE just wrote are not in the base table yet, so
        // a walk that re-reads the table finds nothing.
        let out = compile_and_emit("WITH ps := (SELECT Post) FOR q IN ps UNION (SELECT Person FILTER .name = q.title)");
        assert!(
            out.sql.contains("FROM \"ps\""),
            "the walk reads the binding:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_mutating_for_loop_binding_becomes_sibling_ctes() {
        let out = compile_and_emit(
            "WITH made := (FOR n IN {'a', 'b'} UNION (INSERT Post { title := n })) SELECT count(made)",
        );
        assert!(
            out.sql.contains("\"made__ids\" AS (\nINSERT INTO"),
            "the insert is its own top-level CTE:\n{}",
            out.sql
        );
        assert!(
            !out.sql.contains("LATERAL"),
            "DML cannot sit in a LATERAL:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_count_of_a_binding_counts_its_rows() {
        // Read as an ordinary expression a binding becomes a scalar subquery,
        // which aborts on the second row.
        let out = compile_and_emit("WITH people := (SELECT Person) SELECT count(people)");
        assert!(out.sql.contains("(SELECT count(*) FROM \"people\")"), "{}", out.sql);
    }

    #[test]
    fn test_dividing_integers_yields_a_float() {
        for (query, expected) in [
            ("SELECT 100 / 133", "((100)::float8 / 133)"),
            ("SELECT Person { a := .age / 2 }", "::float8 / 2)"),
            ("SELECT Person.age / 2", "::float8 / 2)"),
            (
                "WITH total := max({133}), current := 38 - 1 SELECT 100 / total * current",
                "(((100)::float8 / (SELECT \"v\" FROM \"total\")) * (SELECT \"v\" FROM \"current\"))",
            ),
            (
                "WITH p := (SELECT Person LIMIT 1), index := p.age - 1, total := max(Person.age) SELECT 100 / total * index",
                "((100)::float8 / (SELECT \"v\" FROM \"total\"))",
            ),
        ] {
            let out = compile_and_emit(query);
            assert!(out.sql.contains(expected), "{query}:\n{}", out.sql);
        }
    }

    #[test]
    fn test_dividing_anything_but_two_integers_is_left_to_postgres() {
        for query in [
            "SELECT 100 // 133",
            "SELECT 100 / 1.5",
            "SELECT <decimal>100 / <decimal>3",
            "SELECT 100 / <decimal>3",
        ] {
            let out = compile_and_emit(query);
            assert!(!out.sql.contains("::float8 /"), "{query}:\n{}", out.sql);
        }
    }

    #[test]
    fn test_a_free_object_under_a_condition_stays_an_object() {
        for query in [
            "SELECT { a := 1 } IF true ELSE {}",
            "SELECT {} IF true ELSE { a := 1 }",
            "SELECT { a := 1 } IF true ELSE {} LIMIT 1",
        ] {
            let out = compile_and_emit(query);
            assert!(
                matches!(&out.shape.root, ShapeNode::Object { type_name: None, pointers, .. } if pointers.len() == 1),
                "{query}: {:?}\n{}",
                out.shape.root,
                out.sql
            );
        }
    }

    #[test]
    fn test_exists_of_a_binding_asks_for_any_row() {
        let out = compile_and_emit("WITH people := (SELECT Person) SELECT exists people");
        assert!(out.sql.contains("EXISTS(SELECT 1 FROM \"people\")"), "{}", out.sql);

        let out = compile_and_emit("WITH names := {'a', 'b'} SELECT exists names");
        assert!(
            out.sql
                .contains("EXISTS(SELECT 1 FROM \"names\" WHERE \"v\" IS NOT NULL)"),
            "{}",
            out.sql
        );
    }

    #[test]
    fn test_a_conditional_insert_beside_a_read_branch_unions_both() {
        let out = compile_and_emit(
            "WITH existing := (SELECT Person FILTER .name = $n LIMIT 1) \
             SELECT (existing IF EXISTS existing ELSE (INSERT Person { name := $n })) { name }",
        );
        assert!(out.sql.contains("UNION ALL"), "{}", out.sql);
        assert!(
            out.sql.contains("INSERT INTO") && out.sql.contains("WHERE (NOT EXISTS(SELECT 1 FROM \"existing\"))"),
            "the insert carries the condition itself:\n{}",
            out.sql
        );
    }

    /// `@pylon.abstract` — unlike an interface it is materialised as nothing
    /// at all, so nothing can read a relation under its name. Modelled on
    /// `Person`, the one test type carrying the `id` every implementor of a
    /// real abstract has.
    fn make_schema_with_a_plain_abstract() -> SchemaDescriptor {
        let mut schema = make_schema();
        let mut archivable = schema
            .types
            .iter()
            .find(|t| t.name == "Person")
            .expect("Person is in the test schema")
            .clone();
        archivable.name = "Archivable".into();
        archivable.table = "Archivable".into();
        archivable.abstract_ = true;
        archivable.materialized = false;
        archivable.links = vec![];
        archivable.multilinks = vec![];
        schema.types.push(archivable);
        let person = schema
            .types
            .iter_mut()
            .find(|t| t.name == "Person")
            .expect("just found");
        person.parents.push("default::Archivable".into());
        schema
    }

    #[test]
    fn test_an_intersection_with_a_plain_abstract_fans_out() {
        let schema = make_schema_with_a_plain_abstract();
        let out = compile_and_emit_with("SELECT Post { who := [is default::Archivable].name }", &schema);
        assert!(
            !out.sql.contains("\"Archivable\""),
            "a plain abstract backs no relation:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("\"public\".\"Person\""),
            "read from the type that carries the column:\n{}",
            out.sql
        );
    }

    /// The same plain abstract, but keeping `Person`'s `company` link so a
    /// backlink can be narrowed to it the way jaldis narrows one to
    /// `default::Creatable`.
    fn make_schema_with_a_linking_plain_abstract() -> SchemaDescriptor {
        let mut schema = make_schema_with_a_plain_abstract();
        let person_links = schema
            .types
            .iter()
            .find(|t| t.name == "Person")
            .expect("Person is in the test schema")
            .links
            .clone();
        let archivable = schema
            .types
            .iter_mut()
            .find(|t| t.name == "Archivable")
            .expect("just added");
        archivable.links = person_links;
        schema
    }

    #[test]
    fn test_a_plain_abstract_in_a_subquery_fans_out() {
        let schema = make_schema_with_a_plain_abstract();
        let out = compile_and_emit_with(
            "SELECT Person FILTER .id IN (SELECT default::Archivable FILTER .age > 1).id",
            &schema,
        );
        assert!(
            !out.sql.contains("\"Archivable\""),
            "a plain abstract backs no relation:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_backlink_narrowed_to_a_plain_abstract_fans_out() {
        let schema = make_schema_with_a_linking_plain_abstract();
        let out = compile_and_emit_with(
            "SELECT Company FILTER .<company[is default::Archivable].name = 'a'",
            &schema,
        );
        assert!(
            !out.sql.contains("\"Archivable\""),
            "a plain abstract backs no relation:\n{}",
            out.sql
        );
    }

    /// `Account` as a real interface (a view), with a plain abstract only its
    /// implementor carries — jaldis reads `[is default::Auditable].*` off
    /// `marketplace::IdentifiableListing` exactly this way.
    fn make_interface_schema_with_a_plain_abstract() -> SchemaDescriptor {
        let mut schema = make_interface_schema();
        let mut archivable = schema
            .types
            .iter()
            .find(|t| t.name == "Individual")
            .expect("Individual is in the interface schema")
            .clone();
        archivable.name = "Archivable".into();
        archivable.table = "Archivable".into();
        archivable.abstract_ = true;
        archivable.materialized = false;
        archivable.interfaces = vec![];
        archivable.links = vec![];
        archivable.multilinks = vec![];
        archivable.computed = vec![];
        schema.types.push(archivable);
        for t in schema.types.iter_mut() {
            match t.name.as_str() {
                // A view to read, so the splat cannot fall back to it.
                "Account" => t.materialized = true,
                "Individual" => t.parents.push("default::Archivable".into()),
                _ => {}
            }
        }
        schema
    }

    #[test]
    fn test_a_splat_over_a_plain_abstract_fans_out() {
        let schema = make_interface_schema_with_a_plain_abstract();
        let out = compile_and_emit_with("SELECT Account { id, [is default::Archivable].* }", &schema);
        assert!(
            !out.sql.contains("\"Archivable\""),
            "a plain abstract backs no relation:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("first_name"),
            "the splat still reaches the mixin's own pointers:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_an_aggregate_over_a_relative_walk_runs_inside_the_subquery() {
        // Compiled as an expression the walk stands for the array of its
        // elements, leaving the aggregate with an array argument.
        let out = compile_and_emit("SELECT Person { latest := max(.posts.title) }");
        assert!(out.sql.contains("(SELECT max("), "{}", out.sql);
        assert!(!out.sql.contains("max(ARRAY("), "{}", out.sql);
    }

    /// `Company.staff` — a computed standing for a narrowed backlink, the
    /// shape jaldis reaches for wherever one type owns rows of another.
    fn make_schema_with_a_computed_backlink() -> SchemaDescriptor {
        use crate::schema::ComputedDescriptor;
        let mut schema = make_schema();
        let company = schema
            .types
            .iter_mut()
            .find(|t| t.name == "Company")
            .expect("Company is in the test schema");
        company.computed = vec![ComputedDescriptor {
            name: "staff".into(),
            expression: ".<company[is default::Person]".into(),
            return_type: None,
            link_target: None,
            link_multi: false,
        }];
        schema
    }

    #[test]
    fn test_a_computed_backlink_reads_the_objects_it_lands_on() {
        let schema = make_schema_with_a_computed_backlink();
        let out = compile_and_emit_with("SELECT Company { staff: { name } }", &schema);
        assert!(
            out.sql.contains("\"public\".\"Person\""),
            "the walk reaches Person:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_computed_backlinks_modifiers_belong_to_what_it_lands_on() {
        // Bound to the declaring type instead, `.name` is a field Company has
        // not got and the query is rejected outright.
        let schema = make_schema_with_a_computed_backlink();
        let out = compile_and_emit_with("SELECT Company.staff { name } FILTER .name = 'ada'", &schema);
        assert!(out.sql.contains("'ada'"), "{}", out.sql);
        assert!(
            out.sql.contains("\"public\".\"Person\""),
            "filtered on the Person the walk lands on:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_an_unreserved_keyword_names_a_free_shape_field() {
        // `last`, `first` and `order` are unreserved in EdgeQL, so a shape can
        // carry them. Read as a set literal, the `:=` has nowhere to go.
        let out = compile_and_emit("SELECT { last := 2, first := 1, order := 3 }");
        assert!(out.sql.contains("\"last\""), "{}", out.sql);
        assert!(out.sql.contains("\"first\""), "{}", out.sql);
        assert!(out.sql.contains("\"order\""), "{}", out.sql);
    }

    #[test]
    fn test_a_shaped_sub_select_over_a_type_is_an_object() {
        // Shapeless it stands for the row's key; with a shape it is the object
        // the shape reads, which used to be refused outright.
        let out = compile_and_emit("SELECT Person { latest := (SELECT Post { title } LIMIT 1) }");
        assert!(out.sql.contains("\"title\""), "{}", out.sql);
    }

    #[test]
    fn test_a_walk_inside_a_path_selects_shape_finds_its_row() {
        // The shape is compiled against the type the walk landed on, so a
        // relative path inside it has that row to resolve against.
        let out = compile_and_emit("SELECT Person.posts { mine := (SELECT Post FILTER .title = .title LIMIT 1) }");
        assert!(out.sql.contains("\"public\".\"Post\""), "{}", out.sql);
    }

    #[test]
    fn test_a_walk_off_a_binding_in_a_free_select() {
        // A free select has no row of its own, so the walk is read back as one
        // subquery; resolved as a type name the binding is simply unknown.
        let out = compile_and_emit("WITH p := (SELECT Person) SELECT { n := p.company.name }");
        assert!(out.sql.contains("\"p\""), "{}", out.sql);
        assert!(out.sql.contains("\"public\".\"Company\""), "{}", out.sql);
    }

    #[test]
    fn test_a_free_selects_filter_can_root_at_a_binding() {
        let out = compile_and_emit("WITH p := (SELECT Person) SELECT { n := 1 } FILTER p.company.name = 'Acme'");
        assert!(out.sql.contains("'Acme'"), "{}", out.sql);
    }

    #[test]
    fn test_a_trailing_shape_on_a_field_access_select() {
        let out = compile_and_emit("SELECT (SELECT Person).company { name }");
        assert!(out.sql.contains("\"name\""), "{}", out.sql);
        assert!(out.sql.contains("\"public\".\"Company\""), "{}", out.sql);
    }

    #[test]
    fn test_a_union_of_walks_names_each_branch() {
        // A walk names no relation, and a union is emitted one relation per
        // branch — so each is hoisted into a CTE of its own first.
        let out = compile_and_emit(
            "WITH p := (SELECT Person FILTER .name = $a), q := (SELECT Person FILTER .name = $b), \
             t := (SELECT p.posts UNION q.posts) SELECT t { title }",
        );
        assert!(out.sql.contains("UNION ALL"), "{}", out.sql);
        assert!(out.sql.contains("\"title\""), "{}", out.sql);
    }

    #[test]
    fn test_sibling_for_loops_each_get_their_own_iterator() {
        // One WITH clause cannot hold `_for_p` twice, and two loops written
        // over the same variable name are ordinary.
        let out = compile_and_emit(
            "WITH a := (SELECT Post), b := (SELECT Post), \
             x := (FOR p IN a UNION (INSERT Person { name := p.title, age := 1 })), \
             y := (FOR p IN b UNION (INSERT Person { name := p.title, age := 2 })) \
             SELECT x",
        );
        assert!(out.sql.contains("\"_for_p\" AS"), "{}", out.sql);
        assert!(out.sql.contains("\"_for_p__1\" AS"), "{}", out.sql);
    }

    #[test]
    fn test_a_nested_for_carries_the_outer_loops_key() {
        // The inner iterator has the outer one in its FROM, and the insert
        // joins the two back on the key it carries — Gel's own iterator bond.
        let out = compile_and_emit(
            "WITH made := (FOR p IN (SELECT Person) UNION ( \
               FOR q IN p.posts UNION (INSERT Company { name := q.title }) \
             )) SELECT count(made)",
        );
        assert!(
            out.sql.contains("\"_outer\""),
            "the inner iterator carries the bond:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("JOIN \"_for_p\" ON"),
            "the insert reads both loop variables:\n{}",
            out.sql
        );
        assert!(
            !out.sql.contains("LATERAL (\nINSERT"),
            "DML cannot sit in a LATERAL:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_link_value_can_be_a_walk_off_a_binding() {
        // `(select placed.line_items limit 1)` — the walk names no type of its
        // own, so reading one off it used to report the path as no type at all.
        let mut schema = make_schema();
        let company = schema
            .types
            .iter_mut()
            .find(|t| t.name == "Company")
            .expect("make_schema declares Company");
        company.properties.push(PropertyDescriptor {
            name: "id".into(),
            pg_type: "uuid".into(),
            nullable: false,
            default_sql: None,
            default_pyql: None,
            description: None,
            check_constraints: vec![],
            is_exclusive: true,
            is_pk: true,
            is_readonly: true,
            rewrites: vec![],
            tuple_members: None,
            column_type: None,
        });
        let out = compile_and_emit_with(
            "WITH others := (SELECT Person) \
             SELECT (INSERT Person { name := $n, age := 1, company := (SELECT others.company LIMIT 1) })",
            &schema,
        );
        assert!(out.sql.contains("\"company_id\""), "{}", out.sql);
    }

    #[test]
    fn test_a_shape_over_a_coalesce_of_walks() {
        // Each operand is a walk off the enclosing row, and every one after the
        // first only stands in when the ones before it are empty.
        let schema = make_schema_with_computed_links();
        let out = compile_and_emit_with("SELECT Person { x := (.published ?? .plain) { title } }", &schema);
        assert!(out.sql.contains("UNION ALL"), "{}", out.sql);
        assert!(
            out.sql.contains("NOT EXISTS"),
            "the later branch only stands in when the first is empty:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_an_aggregate_over_a_walk_onto_objects_counts_its_rows() {
        // Falling through instead put the aggregate back through the very
        // expression route that sent it here, and the recursion took the
        // process down with it.
        let out = compile_and_emit("WITH c := (SELECT Company) SELECT c { n := count(.<company[is default::Person]) }");
        assert!(out.sql.contains("count("), "{}", out.sql);
        assert!(out.sql.contains("\"public\".\"Person\""), "{}", out.sql);
    }

    #[test]
    fn an_aggregate_over_a_computed_set_counts_inside_a_subquery() {
        // `count(.plain)` is one step, so the junction-only count a stored
        // multi-link gets cannot apply — it has no junction to read. Taking
        // the walk as a value instead made the whole select an aggregate
        // query, which Postgres rejects ("column must appear in the GROUP BY
        // clause").
        let schema = make_schema_with_computed_links();
        let out = compile_and_emit_with("SELECT Person { n := count(.plain) }", &schema);
        assert!(
            !out.sql.contains("count((SELECT"),
            "the aggregate belongs inside the walk's subquery, not around it:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("(SELECT count(") && out.sql.contains("\"public\".\"Post\""),
            "it counts the rows the computed's own path lands on:\n{}",
            out.sql
        );
    }

    /// The specialized count for a stored multi-link reads its junction table
    /// alone — the computed case above must not have pulled it off that route.
    #[test]
    fn an_aggregate_over_a_stored_multi_link_still_counts_its_junction_rows() {
        let schema = make_schema_with_computed_links();
        let out = compile_and_emit_with("SELECT Person { n := count(.posts) }", &schema);
        assert!(out.sql.contains("\"public\".\"Person.posts\""), "{}", out.sql);
        assert!(
            !out.sql.contains("JOIN \"public\".\"Post\""),
            "counting junction rows needs no join onto the targets:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_select_can_name_its_own_result() {
        // `SELECT OptionallyAliasedExpr` in EdgeQL: the alias names the result
        // and the select's own clauses may read it by that name.
        let out = compile_and_emit("SELECT oldest := max(Person.age)");
        assert!(out.sql.contains("max("), "{}", out.sql);
        let filtered = compile_and_emit("SELECT n := Person.age FILTER n > 18");
        assert!(
            filtered.sql.contains("18"),
            "the alias resolves to the result:\n{}",
            filtered.sql
        );
    }

    #[test]
    fn test_exists_sees_a_computed_pointer() {
        // The suggester could see the name the resolver had just reported as
        // unknown — `has no link or property 'published'. Did you mean
        // 'published'?`
        let schema = make_schema_with_computed_links();
        let out = compile_and_emit_with("SELECT Person FILTER EXISTS .published", &schema);
        assert!(out.sql.contains("EXISTS"), "{}", out.sql);
        assert!(
            out.sql.contains("\"public\".\"Post\""),
            "the computed's own path:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_walking_through_a_computed_field_access() {
        use crate::schema::ComputedDescriptor;
        // `parent := (select .parents … limit 1).parent` — the modifiers belong
        // to the step before the field access, not to what it lands on.
        let mut schema = make_schema_with_computed_links();
        let person = schema.types.iter_mut().find(|t| t.name == "Person").expect("Person");
        person.computed.push(ComputedDescriptor {
            name: "latest_title".into(),
            expression: "(select .posts order by .title desc limit 1).title".into(),
            return_type: None,
            link_target: None,
            link_multi: false,
        });
        let out = compile_and_emit_with("SELECT Person { x := .latest_title }", &schema);
        assert!(out.sql.contains("\"title\""), "{}", out.sql);
        assert!(
            out.sql.contains("LIMIT"),
            "the computed's own limit survives:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_computed_field_access_onto_an_object_takes_a_shape() {
        use crate::schema::ComputedDescriptor;
        // Compiled as an expression, the pointer came back as the bare id of
        // the object it lands on, whatever shape was written on it.
        let mut schema = make_schema_with_computed_links();
        let post = schema.types.iter_mut().find(|t| t.name == "Post").expect("Post");
        post.links.push(LinkDescriptor {
            name: "author".into(),
            target: "default::Person".into(),
            nullable: true,
            through: None,
            description: None,
            default_pyql: None,
            is_exclusive: false,
            is_readonly: false,
            rewrites: vec![],
            on_delete: vec![],
        });
        let person = schema.types.iter_mut().find(|t| t.name == "Person").expect("Person");
        person.computed.retain(|c| c.name != "looper");
        person.computed.push(ComputedDescriptor {
            name: "latest_author".into(),
            expression: "((select .posts order by .title desc limit 1)).author".into(),
            return_type: None,
            link_target: None,
            link_multi: false,
        });
        let out = compile_and_emit_with("SELECT Person { latest_author: { name } }", &schema);
        assert!(out.sql.contains("\"name\""), "{}", out.sql);
        assert!(
            out.sql.contains("\"title\" DESC"),
            "the order stays on the posts:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("LIMIT"),
            "the computed's own limit survives:\n{}",
            out.sql
        );
        let splat = compile_and_emit_with("SELECT Person { * }", &schema);
        assert!(
            !splat.sql.contains("latest_author"),
            "`*` leaves links out:\n{}",
            splat.sql
        );
    }

    #[test]
    fn test_a_narrowed_pointer_can_carry_a_shape() {
        // `[is T].posts: { title }` — the parser stopped at the `:`, and the
        // route behind it resolved a stored property and nothing else, so a
        // link reported itself as the name it had just been asked for.
        let out = compile_and_emit("SELECT Company { x := 1, [is default::Person].posts: { title } }");
        assert!(out.sql.contains("\"title\""), "{}", out.sql);
        assert!(out.sql.contains("\"public\".\"Post\""), "{}", out.sql);
    }

    #[test]
    fn test_select_type_name_as_a_path_step() {
        let out = compile_and_emit("SELECT Person.__type__");
        assert!(out.sql.contains("ROW('default::Person')"), "{}", out.sql);
        assert!(out.sql.contains("\"Person\" AS \"t0\""), "{}", out.sql);
    }

    #[test]
    fn test_aggregate_over_a_with_bound_object_set() {
        let out = compile_and_emit("WITH people := (SELECT Person) SELECT max(people.age)");
        assert!(out.sql.contains("\"people\" AS ("));
        assert!(out.sql.contains("max("));
        // Reads the binding, not the base table again.
        assert!(!out.sql.contains("FROM \"default\".\"Person\" AS \"t1\""));
    }

    #[test]
    fn test_select_union_of_two_object_bindings() {
        let out = compile_and_emit("WITH a := (SELECT Person LIMIT 1), b := (SELECT Person) SELECT (a UNION b)");
        assert!(out.sql.contains("SELECT * FROM \"a\""));
        assert!(out.sql.contains("UNION ALL"));
        assert!(out.sql.contains("SELECT * FROM \"b\""));
    }

    #[test]
    fn test_select_union_of_different_object_types_is_rejected() {
        let schema = make_schema();
        let ast = parse::parse("WITH a := (SELECT Person), b := (SELECT Company) SELECT (a UNION b)").unwrap();
        let Err(err) = ir::compile(&ast, &schema) else {
            panic!("union of unrelated types must be rejected")
        };
        assert!(format!("{err}").contains("UNION"));
    }

    #[test]
    fn test_for_union_as_a_function_body() {
        let schema = make_schema();
        let descriptor = FunctionDescriptor {
            name: "recent".into(),
            module: "default".into(),
            params: vec![],
            return_pg_type: "default::Person".into(),
            return_is_object: true,
            return_is_set: true,
            return_is_polymorphic: false,
            volatility: "stable".into(),
            body: "for n in {1, 2} union (select Person filter .age = n)".into(),
        };
        let ir = ir::compile_fn_body(&descriptor, &schema).expect("function body must compile");
        let body = emit_fn_body(&ir);
        assert!(body.contains("VALUES"), "{body}");
        assert!(body.contains("CROSS JOIN LATERAL"), "{body}");
        assert!(body.contains("\"_body\".*"), "{body}");
    }

    #[test]
    fn test_with_binding_in_a_computed_reads_the_enclosing_object() {
        let out = compile_and_emit("SELECT Person { n := (WITH own := .name SELECT own) }");
        assert!(out.sql.contains("\"name\""), "{}", out.sql);
    }

    #[test]
    fn test_subject_path_update_names_its_own_alias() {
        let mut schema = make_schema();
        let company = schema.types.iter_mut().find(|t| t.name == "Company").unwrap();
        company.properties.insert(
            0,
            PropertyDescriptor {
                name: "id".into(),
                pg_type: "uuid".into(),
                nullable: false,
                default_sql: Some("uuidv7()".into()),
                default_pyql: None,
                description: None,
                check_constraints: vec![],
                is_exclusive: true,
                is_pk: true,
                is_readonly: true,
                rewrites: vec![],
                tuple_members: None,
                column_type: None,
            },
        );
        // The row set narrowing the update compares against *that update's*
        // alias; a bare `id` would be ambiguous once a nested statement's CTE
        // joins the FROM.
        let out = compile_and_emit_with("SELECT (UPDATE Person.company SET { name := 'x' }) { name }", &schema);
        assert!(out.sql.contains("\"t1\".\"id\" = ANY("), "{}", out.sql);
    }

    #[test]
    fn test_an_update_past_a_from_returns_only_its_own_columns() {
        // `UPDATE … FROM cte … RETURNING *` returns the joined relation's
        // columns as well, so a CTE built that way carries two `id`s and any
        // later reference to one is ambiguous.
        let out = compile_and_emit(
            "SELECT { (UPDATE Person FILTER .name = 'a' SET { company := (INSERT Company { name := 'c' }) }) }",
        );
        assert!(out.sql.contains("FROM \"_nested_dml_0\""), "{}", out.sql);
        assert!(
            out.sql.contains("RETURNING \"t0\".*"),
            "the update names its own target:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_nested_insert_in_a_mutation_written_as_a_free_set() {
        let out = compile_and_emit(
            "SELECT { (UPDATE Person FILTER .name = 'a' SET { company := (INSERT Company { name := 'c' }) }) }",
        );
        assert!(
            out.sql.contains("INSERT INTO \"public\".\"Company\""),
            "the hoisted insert is defined, not just referenced:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_nested_insert_as_a_link_value_in_an_update() {
        // The hoisted insert sits beside the update rather than inside it, so
        // the update needs a FROM to read its id, and its own `id` has to name
        // the update's alias — a bare one could mean either relation.
        let out = compile_and_emit(
            "SELECT (UPDATE Person FILTER .name = 'a' SET { company := (INSERT Company { name := 'c' }) }) { name }",
        );
        assert!(out.sql.contains("INSERT INTO \"public\".\"Company\""), "{}", out.sql);
        assert!(out.sql.contains("_nested_dml_0"), "{}", out.sql);
        assert!(
            out.sql.contains("FROM \"_nested_dml_0\""),
            "the update reads the hoisted CTE:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_nested_insert_as_a_link_value() {
        let out = compile_and_emit(
            "SELECT (INSERT Person { name := 'a', company := (INSERT Company { name := 'c' }) }) { name }",
        );
        assert!(out.sql.contains("_nested_dml_0"), "{}", out.sql);
        assert!(out.sql.contains("INSERT INTO \"public\".\"Company\""), "{}", out.sql);
    }

    #[test]
    fn test_nested_insert_in_a_one_element_set_is_that_insert() {
        let out = compile_and_emit(
            "SELECT (INSERT Person { name := 'a', company := { (INSERT Company { name := 'c' }) } }) { name }",
        );
        assert!(out.sql.contains("INSERT INTO \"public\".\"Company\""), "{}", out.sql);
    }

    #[test]
    fn test_insert_with_no_assignments_uses_default_values() {
        let out = compile_and_emit("SELECT (INSERT Person { name := 'a', company := (INSERT Company {}) }) { name }");
        assert!(out.sql.contains("DEFAULT VALUES"), "{}", out.sql);
    }

    #[test]
    fn test_backlink_narrowed_to_an_interface_its_implementor_declares() {
        // `notes` is declared on Individual; `[is Account]` is satisfied by an
        // Individual row, so the backlink resolves through the implementor.
        use crate::schema::{LinkDescriptor, TypeDescriptor};
        let mut schema = make_interface_schema();
        let individual = schema.types.iter().find(|t| t.name == "Individual").unwrap().clone();
        schema.types.push(TypeDescriptor {
            name: "Note".into(),
            module: "default".into(),
            table: "Note".into(),
            abstract_: false,
            materialized: true,
            description: None,
            parents: vec![],
            interfaces: vec![],
            bases: vec![],
            properties: individual.properties[..1].to_vec(),
            links: vec![],
            multilinks: vec![],
            computed: vec![],
            constraints: vec![],
            indexes: vec![],
            partition: None,
            vector_indexes: vec![],
            search_indexes: vec![],
            triggers: vec![],
            junction: false,
            signals: vec![],
        });
        schema
            .types
            .iter_mut()
            .find(|t| t.name == "Individual")
            .unwrap()
            .links
            .push(LinkDescriptor {
                name: "note".into(),
                target: "default::Note".into(),
                nullable: true,
                description: None,
                default_pyql: None,
                is_exclusive: false,
                is_readonly: false,
                rewrites: vec![],
                on_delete: vec![],
                through: None,
            });
        let out = compile_and_emit_with("SELECT Note FILTER EXISTS .<note[is Account]", &schema);
        assert!(out.sql.contains("\"Individual\""), "{}", out.sql);
    }

    #[test]
    fn test_subject_reads_the_row_a_constraint_checks() {
        let schema = make_schema();
        let sql = ir::compile_constraint_expr("__subject__.age > 18", "default::Person", &schema)
            .expect("__subject__ names the row under check");
        assert!(sql.contains("\"age\""), "{sql}");
    }

    #[test]
    fn test_free_select_filter_gates_the_result_and_warns() {
        let schema = make_schema();
        let ast = parse::parse("SELECT count(Person) FILTER (Person.age > 18)").unwrap();
        let ir = ir::compile(&ast, &schema).unwrap();
        let out = emit(&ir);
        assert!(
            out.sql.contains("EXISTS("),
            "the condition gates the count:\n{}",
            out.sql
        );
        assert!(
            ir.warnings.iter().any(|w| w.contains("FILTER clause")),
            "a set-valued filter warns: {:?}",
            ir.warnings
        );
    }

    #[test]
    fn test_free_sub_select_with_modifiers_reads_inline() {
        // The inner statement reads the enclosing row, so it cannot become a
        // CTE ahead of the FROM clause that defines it.
        let out = compile_and_emit(
            "SELECT Person { n := (WITH mine := .name SELECT count(Post) FILTER (Post.title = mine)) }",
        );
        assert!(out.sql.contains("_scalar_sub"), "{}", out.sql);
        assert!(!out.sql.contains("WITH"), "nothing is hoisted:\n{}", out.sql);
    }

    #[test]
    fn test_correlated_with_binding_is_not_hoisted() {
        let out = compile_and_emit("SELECT Person { n := (WITH own := .name SELECT own) }");
        assert!(
            !out.sql.contains("\"own\" AS ("),
            "the binding is inlined:\n{}",
            out.sql
        );
        assert!(out.sql.contains("\"name\""), "{}", out.sql);
    }

    #[test]
    fn test_array_literal_binding_is_readable_and_typed() {
        let out = compile_and_emit("WITH order := ['a', 'b'] SELECT std::find(order, 'b')");
        assert!(out.sql.contains("array_position"), "its type is known:\n{}", out.sql);
        assert!(
            out.sql.contains("AS v"),
            "the binding exposes a value column:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_any_over_a_single_boolean_is_that_boolean() {
        let out = compile_and_emit("SELECT Person FILTER std::any(.age > 18)");
        assert!(!out.sql.contains("bool_or"), "no aggregate in WHERE:\n{}", out.sql);
        assert!(out.sql.contains("> 18"), "{}", out.sql);
    }

    #[test]
    fn test_comparing_a_value_to_a_set_is_membership() {
        let schema = make_schema_with_computed_links();
        let out = compile_and_emit_with("SELECT Person FILTER .name = Person.posts.title", &schema);
        assert!(out.sql.contains("= ANY("), "{}", out.sql);
    }

    #[test]
    fn test_for_over_a_with_binding_iterates_every_row() {
        let out = compile_and_emit(
            "WITH names := (SELECT Person.name) FOR n IN names UNION (SELECT Person FILTER .name = n)",
        );
        assert!(out.sql.contains("FROM \"names\""), "{}", out.sql);
        assert!(!out.sql.contains("VALUES"), "the whole set is iterated:\n{}", out.sql);
    }

    #[test]
    fn test_for_body_bindings_stay_inside_the_body() {
        // `matching` reads the loop variable, so it cannot be hoisted to the
        // statement's own WITH — nothing there can see `_for_n`.
        let out = compile_and_emit(
            "WITH names := (SELECT Person.name) \
             FOR n IN names UNION (WITH matching := (SELECT Person FILTER .name = n) SELECT matching)",
        );
        let lateral = out.sql.find("CROSS JOIN LATERAL").expect("a lateral body");
        let binding = out.sql.find("\"matching\" AS (").expect("the body's binding");
        assert!(binding > lateral, "the binding sits inside the body:\n{}", out.sql);
    }

    #[test]
    fn test_array_valued_binding_picks_the_array_overload() {
        // The binding carries `text[]`, so the call resolves to the array
        // overload even though the value itself is no longer an array literal.
        let out = compile_and_emit("WITH l := <array<std::str>>$0 SELECT std::find(l, 'a')");
        assert!(out.sql.contains("array_position"), "{}", out.sql);
        assert!(!out.sql.contains("strpos"), "{}", out.sql);
    }

    #[test]
    fn test_aggregate_over_a_path_in_a_filter_is_a_subquery() {
        let out = compile_and_emit("SELECT Person FILTER .age = std::max(Person.age)");
        assert!(
            out.sql.contains("(SELECT max("),
            "an aggregate in WHERE needs its own query:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_distinct_over_a_single_value_is_that_value() {
        let out = compile_and_emit("SELECT Person { n := (DISTINCT .name) }");
        assert!(out.sql.contains("\"name\""), "{}", out.sql);
        assert!(!out.sql.contains("DISTINCT \""), "{}", out.sql);
    }

    #[test]
    fn test_narrowing_an_interface_joins_the_implementor_table() {
        // `first_name` lives on Individual's own table; the Account view has
        // only the columns every implementor shares.
        let mut schema = make_interface_schema();
        let account = schema.types.iter_mut().find(|t| t.name == "Account").unwrap();
        account.materialized = true;
        let out = compile_and_emit_with("SELECT Account[is Individual].first_name", &schema);
        assert!(
            out.sql.contains("JOIN \"public\".\"Individual\""),
            "the implementor's own table is joined:\n{}",
            out.sql
        );
        assert!(out.sql.contains("\"first_name\""), "{}", out.sql);
    }

    #[test]
    fn test_sub_select_shape_declares_names_its_filter_and_projection_read() {
        let out = compile_and_emit("SELECT Person { n := (SELECT .posts { t := .title } FILTER .t = 'x' LIMIT 1).t }");
        assert!(out.sql.contains("'x'"), "{}", out.sql);
        assert!(out.sql.contains("\"title\""), "{}", out.sql);
    }

    #[test]
    fn test_assert_over_an_object_set_returns_rows() {
        let out = compile_and_emit("SELECT std::assert_distinct((SELECT Person))");
        assert!(
            out.sql.contains("\"assert_distinct\"(ARRAY(SELECT \"id\""),
            "{}",
            out.sql
        );
        assert!(out.sql.contains("= ANY("), "{}", out.sql);
        assert!(out.sql.contains("\"Person\" AS \"t1\""), "{}", out.sql);
        assert!(!out.sql.contains("unnest("), "{}", out.sql);
    }

    #[test]
    fn test_assert_over_a_for_union_reads_its_rows() {
        let out =
            compile_and_emit("SELECT std::assert_distinct((FOR n IN {1, 2} UNION (SELECT Person FILTER .age = n)))");
        assert!(out.sql.contains("CROSS JOIN LATERAL"), "{}", out.sql);
        assert!(out.sql.contains("= ANY("), "{}", out.sql);
    }

    #[test]
    fn test_declared_computed_with_binding_reads_its_own_object() {
        // Validation compiles a declared computed on its own, with no
        // enclosing SELECT — `.name` in the binding still means this object's.
        use crate::schema::ComputedDescriptor;
        let schema = make_schema();
        let cd = ComputedDescriptor {
            name: "shouted".into(),
            expression: "(with own := .name select std::str_upper(own))".into(),
            return_type: None,
            link_target: None,
            link_multi: false,
        };
        let ir = ir::compile_computed_in_type(&cd, "default::Person", &schema)
            .expect("a computed may read the object it is declared on");
        assert!(ir.is_some());
    }

    #[test]
    fn test_sub_select_modifiers_scope_to_its_own_subject() {
        // `.title` belongs to Post, the sub-select's subject — not to Person,
        // where the projection `.author.name` lands.
        let mut schema = make_schema();
        let post = schema
            .types
            .iter_mut()
            .find(|t| t.name == "Post")
            .expect("Post is in the test schema");
        post.links.push(LinkDescriptor {
            name: "author".into(),
            target: "default::Person".into(),
            nullable: true,
            description: None,
            default_pyql: None,
            is_exclusive: false,
            is_readonly: false,
            rewrites: vec![],
            on_delete: vec![],
            through: None,
        });
        let out = compile_and_emit_with(
            "SELECT Person { a := (SELECT .posts FILTER .title = 'x' LIMIT 1).author.name }",
            &schema,
        );
        assert!(out.sql.contains("'x'"), "{}", out.sql);
    }

    #[test]
    fn test_free_select_set_literal() {
        let schema = make_schema();
        let ast = parse::parse("SELECT {1, 2, 3}").unwrap();
        let ir = ir::compile(&ast, &schema).unwrap();
        let out = emit(&ir);
        // Three UNION ALL branches
        assert_eq!(out.sql.matches("UNION ALL").count(), 2);
        assert!(out.sql.contains("1 AS v"));
        assert!(out.sql.contains("2 AS v"));
        assert!(out.sql.contains("3 AS v"));
        assert!(out.sql.contains("ROW(v) AS result"));
        assert!(matches!(out.shape.root, crate::query::ShapeNode::Scalar { .. }));
    }

    #[test]
    fn test_free_select_free_object() {
        let schema = make_schema();
        let ast = parse::parse("SELECT { foo := 'bar', n := 42 }").unwrap();
        let ir = ir::compile(&ast, &schema).unwrap();
        let out = emit(&ir);
        assert!(out.sql.contains("'bar'"));
        assert!(out.sql.contains("42"));
        assert!(out.sql.contains("AS result"));
        // Shape should describe an object with pointers foo and n
        let crate::query::ShapeNode::Object {
            pointers, type_name, ..
        } = &out.shape.root
        else {
            panic!("expected Object shape")
        };
        assert!(type_name.is_none());
        assert_eq!(pointers.len(), 2);
        assert!(matches!(&pointers[0], crate::query::ShapeNode::Scalar { name, position: 0 } if name == "foo"));
        assert!(matches!(&pointers[1], crate::query::ShapeNode::Scalar { name, position: 1 } if name == "n"));
    }

    #[test]
    fn test_free_select_object_with_enum_field_casts_to_text_and_tags_shape() {
        // Regression: `select { gender := default::Gender.Male }` failed at
        // runtime with "no decoder for composite type element ... " —
        // A custom enum's OID is database-assigned, so inside an anonymous ROW()
        // composite. free_item_shape/emit_free_select always treated every
        // free-object field as a plain untyped Scalar, never casting an
        // enum-valued field to ::text (unlike a schema object's emit_scalar,
        // which already does this for a real enum-typed property column).
        let mut schema = make_schema();
        schema.enums.push(crate::schema::EnumDescriptor {
            name: "Gender".into(),
            module: "default".into(),
            members: vec!["Male".into(), "Female".into()],
        });
        let out = compile_and_emit_with("select { gender := default::Gender.Male }", &schema);
        // The enum value is computed once (`'Male'::"public"."Gender" AS
        // "_f0"`) and reused for both the `result` composite (cast to
        // `::text` there, since an enum's OID is database-assigned inside an
        // anonymous ROW()) and the plain per-field column exposed for
        // `IrExpr::CteFieldRef` access — so the `::text` cast now applies
        // to that computed column, not inline on the enum literal itself.
        assert!(
            out.sql.contains("'Male'::\"public\".\"Gender\""),
            "expected the enum literal, got:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("ROW(\"_f0\"::text) AS result"),
            "expected the ROW composite to cast the enum field to text, got:\n{}",
            out.sql
        );
        let crate::query::ShapeNode::Object { pointers, .. } = &out.shape.root else {
            panic!("expected Object shape")
        };
        assert_eq!(pointers.len(), 1);
        assert!(
            matches!(
                &pointers[0],
                crate::query::ShapeNode::Enum { name, position: 0, enum_type }
                    if name == "gender" && enum_type == "public::Gender"
            ),
            "expected Enum-tagged shape, got: {:?}",
            pointers[0]
        );
    }

    #[test]
    fn test_free_select_bare_enum_literal_casts_to_text_inside_row() {
        // Same bug, bare scalar form: `select default::Gender.Male;` (no
        // shape/object wrapper) also wraps the value in ROW() for top-level
        // top-level decoding.
        let mut schema = make_schema();
        schema.enums.push(crate::schema::EnumDescriptor {
            name: "Gender".into(),
            module: "default".into(),
            members: vec!["Male".into(), "Female".into()],
        });
        let out = compile_and_emit_with("select default::Gender.Male", &schema);
        assert!(
            out.sql.contains("ROW(v::text) AS result"),
            "expected ROW(v::text), got:\n{}",
            out.sql
        );
        assert!(
            matches!(
                &out.shape.root,
                crate::query::ShapeNode::Enum { enum_type, .. } if enum_type == "public::Gender"
            ),
            "expected Enum-tagged shape, got: {:?}",
            out.shape.root
        );
    }

    // ── FOR UPDATE / FOR SHARE row-locking clause ──────────────────────────────

    #[test]
    fn test_for_update_defaults_to_blocking() {
        let out = compile_and_emit("SELECT Person FOR UPDATE");
        assert!(out.sql.trim_end().ends_with("FOR UPDATE"), "got:\n{}", out.sql);
    }

    #[test]
    fn test_for_update_skip_locked() {
        let out = compile_and_emit("SELECT Person FOR UPDATE SKIP LOCKED");
        assert!(
            out.sql.trim_end().ends_with("FOR UPDATE SKIP LOCKED"),
            "got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_for_update_nowait() {
        let out = compile_and_emit("SELECT Person FOR UPDATE NOWAIT");
        assert!(out.sql.trim_end().ends_with("FOR UPDATE NOWAIT"), "got:\n{}", out.sql);
    }

    #[test]
    fn test_for_no_key_update_skip_locked() {
        let out = compile_and_emit("SELECT Person FOR NO KEY UPDATE SKIP LOCKED");
        assert!(
            out.sql.trim_end().ends_with("FOR NO KEY UPDATE SKIP LOCKED"),
            "got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_for_share() {
        let out = compile_and_emit("SELECT Person FOR SHARE");
        assert!(out.sql.trim_end().ends_with("FOR SHARE"), "got:\n{}", out.sql);
    }

    #[test]
    fn test_for_key_share_nowait() {
        let out = compile_and_emit("SELECT Person FOR KEY SHARE NOWAIT");
        assert!(
            out.sql.trim_end().ends_with("FOR KEY SHARE NOWAIT"),
            "got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_for_update_comes_after_order_by_limit_offset_in_emitted_sql() {
        // Postgres's own grammar places the locking clause last — confirm
        // the emitter matches, not just that all the pieces are present.
        let out = compile_and_emit("SELECT Person { name } ORDER BY .name OFFSET 1 LIMIT 5 FOR UPDATE SKIP LOCKED");
        let order_pos = out.sql.find("ORDER BY").unwrap();
        let offset_pos = out.sql.find("OFFSET").unwrap();
        let limit_pos = out.sql.find("LIMIT").unwrap();
        let for_pos = out.sql.find("FOR UPDATE").unwrap();
        assert!(
            order_pos < offset_pos && offset_pos < limit_pos && limit_pos < for_pos,
            "got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_for_update_combined_with_distinct_is_rejected() {
        let ast = parse::parse("SELECT DISTINCT Person FOR UPDATE").expect("parse failed");
        let err = ir::compile(&ast, &make_schema())
            .err()
            .expect("expected a compile error");
        assert!(err.to_string().contains("DISTINCT"), "unexpected: {err}");
    }

    #[test]
    fn test_for_update_combined_with_select_over_insert_is_rejected() {
        let ast = parse::parse("SELECT (INSERT Person { name := 'Alice' }) { name } FOR UPDATE").expect("parse failed");
        let err = ir::compile(&ast, &make_schema())
            .err()
            .expect("expected a compile error");
        assert!(err.to_string().contains("INSERT"), "unexpected: {err}");
    }

    #[test]
    fn test_for_update_on_an_interface_type_is_rejected() {
        fn id_prop() -> PropertyDescriptor {
            PropertyDescriptor {
                name: "id".into(),
                pg_type: "uuid".into(),
                nullable: false,
                default_sql: Some("uuidv7()".into()),
                default_pyql: None,
                description: None,
                check_constraints: vec![],
                is_exclusive: true,
                is_pk: true,
                is_readonly: true,
                rewrites: vec![],
                tuple_members: None,
                column_type: None,
            }
        }
        let schema = SchemaDescriptor {
            types: vec![
                TypeDescriptor {
                    name: "Account".into(),
                    module: "default".into(),
                    table: "Account".into(),
                    abstract_: true,
                    materialized: true,
                    description: None,
                    parents: vec![],
                    interfaces: vec![],
                    bases: vec![],
                    properties: vec![id_prop()],
                    links: vec![],
                    multilinks: vec![],
                    computed: vec![],
                    constraints: vec![],
                    indexes: vec![],
                    partition: None,
                    vector_indexes: vec![],
                    search_indexes: vec![],
                    triggers: vec![],
                    junction: false,
                    signals: vec![],
                },
                TypeDescriptor {
                    name: "Individual".into(),
                    module: "default".into(),
                    table: "Individual".into(),
                    abstract_: false,
                    materialized: true,
                    description: None,
                    parents: vec![],
                    interfaces: vec!["default::Account".into()],
                    bases: vec![],
                    properties: vec![id_prop()],
                    links: vec![],
                    multilinks: vec![],
                    computed: vec![],
                    constraints: vec![],
                    indexes: vec![],
                    partition: None,
                    vector_indexes: vec![],
                    search_indexes: vec![],
                    triggers: vec![],
                    junction: false,
                    signals: vec![],
                },
            ],
            scalars: vec![],
            enums: vec![],
            named_tuples: vec![],
            globals: vec![],
            functions: vec![],
            aliases: vec![],
            channels: vec![],
            ..Default::default()
        };
        let ast = parse::parse("SELECT Account FOR UPDATE").expect("parse failed");
        let err = ir::compile(&ast, &schema).err().expect("expected a compile error");
        assert!(err.to_string().contains("interface"), "unexpected: {err}");
    }

    #[test]
    fn test_schema_select_distinct_emits_distinct_keyword() {
        // Regression: compile_stmt's Distinct/Detached unwrap and
        // compile_select's own re-derivation of the same unwrap were merged
        // into a single pass-through (Phase 2 of the compile_expr merge) —
        // confirm `select distinct` still compiles and emits DISTINCT.
        let out = compile_and_emit("SELECT DISTINCT Person { name }");
        assert!(out.sql.contains("DISTINCT"), "expected DISTINCT in SQL:\n{}", out.sql);
    }

    #[test]
    fn test_schema_select_detached_compiles_as_ordinary_select() {
        // Same merge — `detached` at the top level of a schema select is a
        // no-op (already independent); confirm it still compiles cleanly
        // instead of erroring or double-unwrapping.
        let out = compile_and_emit("SELECT DETACHED Person { name }");
        assert!(
            out.sql.contains("\"name\""),
            "expected name column in SQL:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_detached_subquery_correlates_against_the_outer_row() {
        // The anti-join idiom: "the Person with no other Person of the same
        // name that is younger". `detached` is what makes the inner `Person` a
        // different row from the outer one, so the bare `Person.name` inside
        // has to resolve to the *outer* alias. It used to be stripped as a
        // no-op, which compiled to `t1.name = t1.name` — a tautology that
        // quietly matched every row instead of erroring.
        let out = compile_and_emit(
            "SELECT Person { name } FILTER NOT EXISTS (\
             SELECT DETACHED Person FILTER .name = Person.name AND .age > Person.age)",
        );
        assert!(
            out.sql.contains("\"t1\".\"name\" = \"t0\".\"name\""),
            "the inner row should be compared against the outer one:\n{}",
            out.sql
        );
        assert!(
            !out.sql.contains("\"t1\".\"name\" = \"t1\".\"name\""),
            "self-comparison means the correlation was lost:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_an_assert_on_a_pointer_checks_its_rows() {
        // The pointer's own subquery is bound once and read back through the
        // check, rather than appearing in both — and the cast is what makes
        // the call legal at all, since PL/pgSQL cannot take a `record[]`.
        let out = compile_and_emit("SELECT Person { name, p := assert_exists(.posts { title }) }");
        assert!(
            out.sql
                .contains(r#"cardinality("_pylon"."assert_exists"("_a"."v"::text[]))"#),
            "the assert should read the aggregated rows:\n{}",
            out.sql
        );
        assert_eq!(
            out.sql.matches("array_agg(ROW(").count(),
            1,
            "the pointer's subquery should be evaluated once, not once per use:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_an_assert_on_a_scalar_stays_an_ordinary_call() {
        // Only an object pointer takes the wrapper; `assert_exists(.name)` is
        // a value expression and must keep compiling as the function call it is.
        let out = compile_and_emit("SELECT Person { n := assert_exists(.name) }");
        assert!(
            !out.sql.contains(r#""_a"."v""#),
            "a scalar assert should not be wrapped as a pointer check:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_backlink_tail_walks_past_the_first_step() {
        // `.<link[is T].multilink.prop` — the tail after a backlink was read
        // one arity at a time (property, or link-then-property), so anything
        // crossing a multi-link had no single column to read and was refused.
        // Counted against the migrated data it agreed with Gel: 4 and 4.
        let out = compile_and_emit("SELECT Post FILTER any(.<posts[is Person].posts.title = $t)");
        assert!(
            out.sql.contains(r#""public"."Person.posts""#),
            "the tail should walk the multi-link's junction:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_an_intersection_narrows_the_rest_of_a_multilink_walk() {
        // `.posts[is Post].title` — the narrowed type's own table is what the
        // remaining steps join to, so a row of any other type drops out
        // without a separate check. Agreed with Gel on real data (1 and 1).
        let out = compile_and_emit("SELECT Person FILTER any(.posts[is Post].title = $t)");
        assert!(
            out.sql.contains(r#""public"."Post""#),
            "the walk should continue from the narrowed type:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_union_of_bindings_emits_selects_not_bare_names() {
        // A CTE reference normally emits just its quoted name, which works
        // where the caller aliases it into a FROM. As a `UNION ALL` operand it
        // has to be a SELECT of its own — `("a" UNION ALL "b")` is not SQL.
        let out = compile_and_emit_with(
            "WITH a := (INSERT Post { title := $t1 }), b := (INSERT Post { title := $t2 }) \
             SELECT (INSERT Person { name := $n, posts := (a UNION b) })",
            &make_schema(),
        );
        assert!(
            out.sql.contains(r#"(SELECT "_s"."id" FROM "a" AS "_s")"#),
            "each union operand should be a select:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_bare_narrowing_reads_as_a_type_check_and_as_an_object() {
        // `exists [is Sub]` is whether the row *is* one; `x := [is Sub] { … }`
        // is that row read as it. Both used to fall off the end of the path
        // walk ("invalid path step" / "empty path traversal").
        for query in [
            "SELECT Person FILTER EXISTS [is Person]",
            "SELECT Person { p := [is Person] { name } }",
        ] {
            let ast = parse::parse(query).unwrap_or_else(|e| panic!("{query}: {e}"));
            assert!(ir::compile(&ast, &make_schema()).is_ok(), "{query} should compile");
        }
    }

    #[test]
    fn test_a_multilink_takes_a_set_wrapped_in_a_select() {
        // `translations := (select { a, b })` — a select with no clauses of its
        // own is the expression it wraps. Left wrapped, a free row source (a
        // set literal, a union) reached the generic subquery arm and was
        // refused, while the identical unwrapped `{ a, b }` compiled.
        let out = compile_and_emit_with(
            "WITH a := (INSERT Post { title := $t1 }), b := (INSERT Post { title := $t2 }) \
             SELECT (INSERT Person { name := $n, posts := (SELECT { a, b }) })",
            &make_schema(),
        );
        assert!(
            out.sql.contains(r#""public"."Person.posts""#),
            "both targets should reach the junction:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_an_aggregate_over_a_subselected_path() {
        // `array_agg((select Person.name))` — the sub-select walks to a scalar,
        // which has no type name to be a schema select's subject. It means the
        // same as `array_agg(Person.name)` with the select's own modifiers on
        // the walk, which the path-select builder already handles. Checked
        // against Gel's own answer, element for element.
        let out = compile_and_emit("SELECT array_agg((SELECT Person.name))");
        assert!(
            out.sql.contains("array_agg") && out.sql.contains(r#""public"."Person""#),
            "the aggregate should wrap the walk's column:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_distinct_aggregate_argument_reaches_the_aggregate() {
        // `array_agg(distinct x)` used to emit `array_agg(x)`: the unary
        // `distinct` emits its operand unchanged, because it is normally
        // applied where the set is built — which an aggregate argument is not.
        // Both spellings mean the same thing and both were checked against
        // Gel, which returns one element where the old output returned two.
        for query in [
            "SELECT array_agg(DISTINCT Person.name)",
            "SELECT array_agg((SELECT DISTINCT Person.name))",
        ] {
            let out = compile_and_emit(query);
            assert!(
                out.sql.contains("array_agg(DISTINCT "),
                "{query} should deduplicate inside the aggregate:\n{}",
                out.sql
            );
        }
    }

    #[test]
    fn test_is_not_negates_the_type_check() {
        // `x IS NOT T` is Gel's own `Expr IS NOT TypeExpr` — the negation of
        // the check, not a comparison against some type `not T`. Counted
        // against the migrated data it came out as 3 of 3, matching Gel.
        let positive = compile_and_emit("SELECT Person FILTER Person IS Person");
        let negative = compile_and_emit("SELECT Person FILTER Person IS NOT Person");
        assert!(
            negative.sql.contains("NOT") && !positive.sql.contains("NOT"),
            "the negation should reach the SQL:\npositive:\n{}\nnegative:\n{}",
            positive.sql,
            negative.sql
        );
    }

    #[test]
    fn test_a_cast_takes_a_prefix_operator_as_its_operand() {
        // `<bool>exists x` is valid EdgeQL — Gel's cast takes a whole `Expr` at
        // CAST precedence. Pylon's cast went straight to a postfix expression,
        // so any prefix operator after one failed to parse.
        for query in ["SELECT <bool>EXISTS (SELECT Person LIMIT 1)", "SELECT <int64>-1"] {
            let ast = parse::parse(query).unwrap_or_else(|e| panic!("{query} should parse: {e}"));
            assert!(ir::compile(&ast, &make_schema()).is_ok(), "{query} should compile");
        }
    }

    #[test]
    fn test_a_bare_multilink_reads_as_a_set() {
        // `p in .posts` — a multi-link read as a value, not traversed through.
        // The single-step resolver knew properties, links and computeds but
        // never multi-links, so the pointer reported itself as unknown while
        // the suggester, which does know them, offered the name straight back.
        let out =
            compile_and_emit("WITH p := (SELECT DETACHED Post LIMIT 1) SELECT Person { name } FILTER p IN .posts");
        assert!(
            out.sql.contains(r#""public"."Person.posts""#),
            "the membership test should read the link's junction:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_select_reads_its_own_shape_computed_in_its_clauses() {
        // `select T { x := … } order by .x` — Gel puts a shape's computeds in
        // scope for the select's own clauses (but not for a sibling pointer,
        // which it rejects, so the scope stops at the clauses).
        let out = compile_and_emit("SELECT Person { n := .name } ORDER BY .n ASC");
        assert!(
            out.sql.contains("ORDER BY"),
            "the order by should compile against the declared pointer:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_declared_pointer_can_be_walked_through() {
        // `p { boss := … }` read back as `.boss.name` — the walk builder
        // resolved a *schema* computed as a step but not one a shape in scope
        // declared, so anything past the first step reported it as unknown.
        let out = compile_and_emit_with(
            "WITH p := (SELECT Person { boss := .company } LIMIT 1) SELECT p { name, boss_name := .boss.name }",
            &make_schema(),
        );
        assert!(
            out.sql.contains(r#""public"."Company""#),
            "the walk should reach the declared pointer's own target table:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_replaced_subject_shape_stays_in_scope() {
        // `(select .company { n := .name }) { n }` — a shape at the point of
        // use replaces the subject's own, so what that one declared has to
        // stay readable or the replacement cannot see it. Before, this was an
        // outright compile error.
        let out = compile_and_emit("SELECT Person { c := (SELECT .company { n := .name }) { n } }");
        assert!(
            out.sql.contains(r#"'default::Company'::text"#) && out.sql.contains(r#""t2"."name""#),
            "`n` should resolve to the company's own column:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_a_sibling_shape_pointer_is_not_in_scope() {
        // Gel rejects one shape pointer reading another; matching that is what
        // keeps four graph queries reported as the application bugs they are.
        let schema = make_schema();
        let ast = parse::parse("SELECT Person { n := .name, copy := .n }").unwrap();
        assert!(ir::compile(&ast, &schema).is_err());
    }

    #[test]
    fn test_shape_over_a_with_binding_reads_the_binding() {
        // `metadata := metadata { … }` — a binding names a row source, so a
        // shape on it projects that source. The pointer route only took
        // relative paths, so this fell through to the expression route and was
        // rejected as "shapes are not valid in expression context".
        let out = compile_and_emit("WITH c := (SELECT Company LIMIT 1) SELECT Person { name, employer := c { name } }");
        assert!(
            out.sql.contains("\"c\""),
            "the pointer should read the binding's CTE:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_shape_over_a_bare_type_name_is_still_rejected() {
        // Unlike a binding, a type name in an enclosing scope is a factored
        // prefix, so a shape on one does not mean "every row of the table".
        let schema = make_schema();
        let ast = parse::parse("SELECT Person { name, other := Company { name } }").unwrap();
        assert!(ir::compile(&ast, &schema).is_err());
    }

    #[test]
    fn test_nested_shape_reads_a_prefix_bound_by_the_outer_select() {
        // `Person.name` inside the `posts` shape names a type the enclosing
        // select binds, so EdgeQL reads the outer row — not every Person. The
        // prefix used to be rejected outright because only the *innermost*
        // subject was checked.
        let out = compile_and_emit("SELECT Person { name, posts: { title, who := Person.name } }");
        assert!(
            out.sql.contains(r#""t0"."name""#),
            "the nested computed should read the outer Person alias:\n{}",
            out.sql
        );
        assert!(
            !out.sql.contains(r#"FROM "public"."Person" "t2""#),
            "the prefix should not open a second Person source:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_prefix_naming_a_type_out_of_scope_is_rejected() {
        // Nothing binds `Company` here, so the prefix stays an absolute path —
        // Pylon must not silently read the whole table.
        let schema = make_schema();
        let ast = parse::parse("SELECT Person { name, who := Company.name }").unwrap();
        assert!(ir::compile(&ast, &schema).is_err());
    }

    #[test]
    fn test_free_select_tuple() {
        let schema = make_schema();
        let ast = parse::parse("SELECT (1, 2)").unwrap();
        let ir = ir::compile(&ast, &schema).unwrap();
        let out = emit(&ir);
        assert!(out.sql.contains("1"));
        assert!(out.sql.contains("2"));
        assert!(out.sql.contains("AS result"));
        assert!(matches!(out.shape.root, crate::query::ShapeNode::Tuple { .. }));
    }

    #[test]
    fn test_free_select_scalar_literal() {
        let schema = make_schema();
        let ast = parse::parse("SELECT 'hello'").unwrap();
        let ir = ir::compile(&ast, &schema).unwrap();
        let out = emit(&ir);
        assert!(out.sql.contains("SELECT 'hello' AS v"));
        assert!(out.sql.contains("ROW(v) AS result"));
        assert!(matches!(out.shape.root, crate::query::ShapeNode::Scalar { .. }));
    }

    #[test]
    fn test_float_literal_casts_to_float8() {
        // Regression: an un-cast float literal like `1.0` is untyped in
        // Postgres and defaults to `numeric`, not `float64` — silently
        // changing PyQL's semantics (only an explicit `<decimal>`/`123n`
        // literal should ever produce a real decimal) and breaking decode
        // inside ROW() composites (numeric's wire format differs from
        // float8's). An explicit ::float8 cast keeps it a real float.
        let out = compile_and_emit("SELECT 1.0");
        assert!(
            out.sql.contains("(1.0::float8)"),
            "expected explicit float8 cast, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_free_select_array_literal() {
        let schema = make_schema();
        let ast = parse::parse("SELECT [1, 2, 3]").unwrap();
        let ir = ir::compile(&ast, &schema).unwrap();
        let out = emit(&ir);
        assert!(out.sql.contains("SELECT ARRAY[1, 2, 3] AS v"), "{}", out.sql);
        assert!(out.sql.contains("SELECT v AS result, v FROM"), "{}", out.sql);
        assert!(matches!(out.shape.root, crate::query::ShapeNode::RawScalar));
    }

    #[test]
    fn test_select_scalars() {
        let out = compile_and_emit("SELECT Person { name, age }");
        assert!(out.sql.contains("'default::Person'::text"));
        assert!(out.sql.contains("\"name\"::text"));
        assert!(out.sql.contains("\"age\"::int8"));
        assert!(out.sql.contains("FROM \"public\".\"Person\""));
        assert!(out.sql.contains(") AS result"));
    }

    #[test]
    fn test_select_filter_param() {
        let out = compile_and_emit("SELECT Person { name } FILTER .name = $name");
        assert!(out.sql.contains("WHERE"));
        assert!(out.sql.contains("$1"));
    }

    #[test]
    fn test_filter_in_set_literal_compiles_to_any_array() {
        let out = compile_and_emit("SELECT Person { name } FILTER .name IN {'Carol'}");
        assert!(out.sql.contains("= ANY(ARRAY['Carol'])"), "got:\n{}", out.sql);
    }

    #[test]
    fn test_filter_in_multi_element_set_literal_compiles_to_any_array() {
        let out = compile_and_emit("SELECT Person { name } FILTER .name IN {'Carol', 'Bob'}");
        assert!(out.sql.contains("= ANY(ARRAY['Carol', 'Bob'])"), "got:\n{}", out.sql);
    }

    /// `x IN std::array_unpack(arr)` is the documented way to test membership
    /// of an array parameter. It used to emit `= ANY(unnest(arr))` — a
    /// set-returning function inside `WHERE`, which PostgreSQL rejects at
    /// execution time (`set-returning functions are not allowed in WHERE`)
    /// even though both the PyQL and the SQL parse cleanly. `ANY` takes the
    /// array directly, so the `unnest` has to go.
    #[test]
    fn test_filter_in_array_unpack_passes_the_array_straight_to_any() {
        let out = compile_and_emit("SELECT Person { name } FILTER .name IN std::array_unpack(<array<str>>$names)");
        assert!(
            !out.sql.contains("ANY(unnest("),
            "unnest inside ANY is invalid in WHERE:\n{}",
            out.sql
        );
        assert!(out.sql.contains("= ANY("), "got:\n{}", out.sql);
    }

    #[test]
    fn test_filter_not_in_array_unpack_passes_the_array_straight_to_all() {
        let out = compile_and_emit("SELECT Person { name } FILTER .name NOT IN std::array_unpack(<array<str>>$names)");
        assert!(
            !out.sql.contains("ALL(unnest("),
            "unnest inside ALL is invalid in WHERE:\n{}",
            out.sql
        );
        assert!(out.sql.contains("<> ALL("), "got:\n{}", out.sql);
    }

    /// `all(array_unpack($t) in .posts.title)` quantifies over the unpacked
    /// array, not over `.posts`. Read as a condition on one `.posts` element
    /// it would both ask the opposite question and leave the `unnest()` in a
    /// `WHERE`, which PostgreSQL rejects at execution time.
    #[test]
    fn test_all_over_unpacked_array_aggregates_the_array_not_the_multilink() {
        let out = compile_and_emit(
            "SELECT Person { name } FILTER all(std::array_unpack(<array<str>>$titles) IN .posts.title)",
        );
        assert!(out.sql.contains("bool_and("), "got:\n{}", out.sql);
        assert!(out.sql.contains("_unnested"), "got:\n{}", out.sql);
    }

    /// The mirror of the above: the multilink is what is quantified over and
    /// the array is only the haystack, so `IN` still folds into `= ANY(array)`
    /// and nothing unnests.
    #[test]
    fn test_any_over_a_multilink_against_an_unpacked_array_still_folds_into_any() {
        let out = compile_and_emit(
            "SELECT Person { name } FILTER any(.posts.title IN std::array_unpack(<array<str>>$titles))",
        );
        assert!(!out.sql.contains("unnest("), "got:\n{}", out.sql);
        assert!(out.sql.contains("= ANY("), "got:\n{}", out.sql);
    }

    #[test]
    fn test_any_over_unpacked_array_aggregates_the_array_not_the_multilink() {
        let out = compile_and_emit(
            "SELECT Person { name } FILTER any(std::array_unpack(<array<str>>$titles) IN .posts.title)",
        );
        assert!(out.sql.contains("bool_or("), "got:\n{}", out.sql);
        assert!(out.sql.contains("_unnested"), "got:\n{}", out.sql);
    }

    /// A schema with a `text[]` property, which `make_schema` has none of.
    fn schema_with_array_property() -> SchemaDescriptor {
        let mut schema = make_schema();
        schema.types[0].properties.push(crate::schema::PropertyDescriptor {
            name: "perms".into(),
            pg_type: "text[]".into(),
            nullable: true,
            default_sql: None,
            default_pyql: None,
            description: None,
            check_constraints: vec![],
            is_exclusive: false,
            is_pk: false,
            is_readonly: false,
            rewrites: vec![],
            tuple_members: None,
            column_type: None,
        });
        schema
    }

    /// `array_agg(array_unpack(teams.perms))` flattens every row's array into
    /// one. Compiled as written it emitted `array_agg(unnest(...))`, which
    /// PostgreSQL rejects at execution time ("aggregate function calls cannot
    /// contain set-returning function calls"), so the unpacking has to happen
    /// in a row source the aggregate reads from outside.
    #[test]
    fn test_aggregate_over_an_unpacked_walk_unnests_in_a_row_source() {
        let out = compile_and_emit_with(
            "WITH teams := (SELECT Person { perms } FILTER .age > 1) SELECT std::array_agg(std::array_unpack(teams.perms))",
            &schema_with_array_property(),
        );
        assert!(!out.sql.contains("array_agg(unnest("), "got:\n{}", out.sql);
        assert!(
            out.sql.contains(
                r#"coalesce(array_agg("_s"."v"), '{}') FROM unnest(ARRAY(SELECT unnest("t1"."perms") FROM "teams" AS "t1"))"#
            ),
            "got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_aggregate_over_an_unpacked_type_rooted_path_unnests_in_a_row_source() {
        let out = compile_and_emit_with(
            "SELECT std::count(std::array_unpack(Person.perms))",
            &schema_with_array_property(),
        );
        assert!(!out.sql.contains("count(unnest("), "got:\n{}", out.sql);
        assert!(
            out.sql
                .contains(r#"FROM unnest(ARRAY(SELECT unnest("t0"."perms") FROM "public"."Person" AS "t0"))"#),
            "got:\n{}",
            out.sql
        );
    }

    /// One array standing on its own — a parameter, or one row's property —
    /// is already the row source, so it needs no walk gathered around it.
    #[test]
    fn test_aggregate_over_an_unpacked_parameter_unnests_the_array_itself() {
        let out = compile_and_emit("SELECT std::array_agg(std::array_unpack(<array<str>>$names))");
        assert!(!out.sql.contains("array_agg(unnest("), "got:\n{}", out.sql);
        assert!(
            out.sql.contains(r#"FROM unnest(($1)::text[]) AS "_s"("v")"#),
            "got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_aggregate_over_an_unpacked_property_unnests_that_row_s_array() {
        let out = compile_and_emit_with(
            "SELECT Person { n := std::array_agg(std::array_unpack(.perms)) }",
            &schema_with_array_property(),
        );
        assert!(!out.sql.contains("array_agg(unnest("), "got:\n{}", out.sql);
        assert!(
            out.sql.contains(r#"FROM unnest("t0"."perms") AS "_s"("v")"#),
            "got:\n{}",
            out.sql
        );
    }

    /// `array_agg` over no rows is `[]` in Gel and NULL in SQL, which reaches
    /// a caller as a missing value where a list was promised.
    #[test]
    fn test_array_agg_over_nothing_is_an_empty_array_not_null() {
        for query in [
            "SELECT std::array_agg(Person.name)",
            "SELECT std::array_agg((SELECT Person).name)",
        ] {
            let out = compile_and_emit(query);
            assert!(
                out.sql.contains("coalesce(") && out.sql.contains("'{}'"),
                "{query} got:\n{}",
                out.sql
            );
        }
    }

    /// The `coalesce` that empty array wears must not hide what the binding
    /// holds: `contains(ids, .id)` picks its overload off the binding's type,
    /// and an array read as text silently becomes a substring search.
    #[test]
    fn test_an_array_agg_binding_stays_an_array_through_its_coalesce() {
        let out = compile_and_emit(
            "WITH ids := std::array_agg(Person.id) SELECT Person { name } FILTER std::contains(ids, .id)",
        );
        assert!(
            out.sql.contains("@> ARRAY["),
            "expected the array overload, got:\n{}",
            out.sql
        );
        assert!(
            !out.sql.contains("strpos("),
            "the string overload must not win, got:\n{}",
            out.sql
        );
    }

    /// An aggregate over a walk that is *not* unpacked keeps its own row
    /// source — the rewrite above must not reach it.
    #[test]
    fn test_aggregate_over_a_plain_walk_keeps_its_row_source() {
        let out = compile_and_emit("SELECT std::array_agg(Person.name)");
        assert!(
            out.sql.contains(r#"array_agg("t0"."name")"#) && out.sql.contains(r#"FROM "public"."Person""#),
            "got:\n{}",
            out.sql
        );
    }

    /// The order check is a safety net for every emitter that assembles a
    /// WITH clause, so it has to actually catch one.
    #[test]
    fn test_the_cte_order_check_catches_a_forward_reference() {
        let bad = "WITH \"b\" AS (\n SELECT * FROM \"a\"\n),\n\"a\" AS (\n SELECT 1\n)\nSELECT 1";
        assert!(
            super::forward_cte_reference(bad).is_some(),
            "expected a forward reference"
        );
        let good = "WITH \"a\" AS (\n SELECT 1\n),\n\"b\" AS (\n SELECT * FROM \"a\"\n)\nSELECT 1";
        assert_eq!(super::forward_cte_reference(good), None, "ordered CTEs must pass");
    }

    /// `.status IN <Enum>array_unpack($1)` — the cast sits between `IN` and
    /// the unpack, and used to hide it, leaving a set-returning function in a
    /// `WHERE` that PostgreSQL rejects at execution time.
    #[test]
    fn test_a_cast_unpacked_array_still_folds_into_any() {
        let out = compile_and_emit("SELECT Person { name } FILTER .name IN <str>std::array_unpack(<array<str>>$names)");
        assert!(
            !out.sql.contains("unnest("),
            "the unpack must fold into ANY, got:\n{}",
            out.sql
        );
        assert!(out.sql.contains("= ANY("), "got:\n{}", out.sql);
    }

    /// A bare `array_unpack` outside `IN` still has to unnest — the unwrap is
    /// specific to `ANY`/`ALL`, which take an array rather than a set.
    #[test]
    fn test_array_unpack_outside_in_still_unnests() {
        let out = compile_and_emit("SELECT std::array_unpack(<array<str>>$names)");
        assert!(out.sql.contains("unnest("), "got:\n{}", out.sql);
    }

    #[test]
    fn test_filter_not_in_set_literal_compiles_to_all_array() {
        let out = compile_and_emit("SELECT Person { name } FILTER .name NOT IN {'Carol'}");
        assert!(out.sql.contains("<> ALL(ARRAY['Carol'])"), "got:\n{}", out.sql);
    }

    #[test]
    fn test_deep_splat_fetches_single_link_target_properties_not_just_id() {
        // `**` on a single link must expand to the target type's own
        // properties (one level of `*`, not `**` again — recursing into
        // the target's own links would never terminate for a cyclic link
        // graph), not just an implicit `{ id }`.
        let out = compile_and_emit("SELECT Person { ** }");
        assert!(
            out.sql.contains("\"name\""),
            "expected Company.name pulled in via .company's ** expansion, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_deep_splat_fetches_multilink_target_properties_not_just_id() {
        let out = compile_and_emit("SELECT Person { ** }");
        assert!(
            out.sql.contains("\"title\""),
            "expected Post.title pulled in via .posts' ** expansion, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_schema_type_cast_select() {
        let out = compile_and_emit("SELECT <default::Person><uuid>'019ef1bb-0d42-7a9f-8f6b-b38d028a49ba'");
        assert!(out.sql.contains("FROM \"public\".\"Person\""));
        assert!(out.sql.contains("WHERE"));
        assert!(out.sql.contains("'019ef1bb-0d42-7a9f-8f6b-b38d028a49ba'"));
    }

    #[test]
    fn test_select_single_link() {
        let out = compile_and_emit("SELECT Person { name, company { name } }");
        assert!(out.sql.contains("'default::Company'::text"));
        assert!(out.sql.contains("FROM \"public\".\"Company\""));
        // join condition: parent FK column = target PK
        assert!(out.sql.contains("\"company_id\" = "));
    }

    #[test]
    fn test_select_multi_link() {
        let out = compile_and_emit("SELECT Person { name, posts { title } }");
        assert!(out.sql.contains("array_agg(ROW("));
        assert!(out.sql.contains("ARRAY[]::record[]"));
        assert!(out.sql.contains("'default::Post'::text"));
        assert!(out.sql.contains("\"Person.posts\""));
    }

    fn compile_err(query: &str) -> String {
        let schema = make_schema();
        let ast = parse::parse(query).expect("parse failed");
        match ir::compile(&ast, &schema) {
            Ok(_) => panic!("expected a compile error"),
            Err(e) => format!("{e}"),
        }
    }

    #[test]
    fn test_multi_link_limit_cuts_rows_before_aggregating() {
        let out = compile_and_emit("SELECT Person { posts: { title } filter .title = 'x' limit 1 }");
        // The LIMIT has to sit on the row source, not next to array_agg —
        // the aggregate has already collapsed the rows by then.
        assert!(out.sql.contains("LIMIT 1"), "{}", out.sql);
        assert!(out.sql.contains("AS \"__lim\""), "{}", out.sql);
        let agg = out.sql.find("array_agg").unwrap();
        let limit = out.sql.find("LIMIT 1").unwrap();
        assert!(agg < limit, "{}", out.sql);
        assert!(out.sql.contains("\"t1\".\"title\"::text AS \"c1\""), "{}", out.sql);
    }

    #[test]
    fn test_multi_link_order_by_with_limit_orders_the_aggregate_too() {
        let out = compile_and_emit("SELECT Person { posts: { title } order by .title desc offset 1 limit 2 }");
        // The sort key rides along as its own column so the array keeps the
        // derived table's order rather than relying on aggregation order.
        assert!(out.sql.contains("\"t1\".\"title\" AS \"s0\""), "{}", out.sql);
        assert!(
            out.sql.contains("ORDER BY \"__lim\".\"s0\" DESC NULLS LAST"),
            "{}",
            out.sql
        );
        assert!(out.sql.contains("OFFSET 1"), "{}", out.sql);
        assert!(out.sql.contains("LIMIT 2"), "{}", out.sql);
    }

    #[test]
    fn test_multi_link_without_limit_keeps_the_flat_aggregate() {
        let out = compile_and_emit("SELECT Person { posts: { title } filter .title = 'x' }");
        assert!(!out.sql.contains("__lim"), "{}", out.sql);
        assert!(out.sql.contains("array_agg(ROW("), "{}", out.sql);
    }

    #[test]
    fn test_sub_select_as_computed_pointer_is_a_link_pointer() {
        let out = compile_and_emit("SELECT Person { recent := (select .posts filter .title = 'x' limit 1) { title } }");
        let ShapeNode::Object { pointers, .. } = &out.shape.root else {
            panic!()
        };
        // Named after the pointer, not after the link it selects from, and
        // one object: `limit 1` makes the sub-select at most one row.
        let ShapeNode::Object {
            name,
            pointers: elem_pointers,
            ..
        } = &pointers[2]
        else {
            panic!("{:?}", pointers[2])
        };
        assert_eq!(name, "recent");
        assert!(matches!(&elem_pointers[1], ShapeNode::Scalar { name, .. } if name == "title"));
        assert!(out.sql.contains("\"jt\".source = \"t0\".id"), "{}", out.sql);
        assert!(out.sql.contains("LIMIT 1"), "{}", out.sql);
    }

    #[test]
    fn test_sub_select_shape_inside_the_parens_is_equivalent() {
        let outer = compile_and_emit("SELECT Person { recent := (select .posts limit 1) { title } }");
        let inner = compile_and_emit("SELECT Person { recent := (select .posts { title } limit 1) }");
        assert_eq!(outer.sql, inner.sql);
    }

    #[test]
    fn test_sub_select_field_access_is_a_correlated_scalar_subquery() {
        let out = compile_and_emit("SELECT Person { t := (select .posts filter .title = 'x' limit 1).title }");
        // Rooted at the enclosing type and correlated back to its row, so
        // the subquery only ever sees this Person's posts.
        assert!(out.sql.contains("(\"t1\".\"id\" = \"t0\".\"id\")"), "{}", out.sql);
        assert!(out.sql.contains("\"t2\".\"title\""), "{}", out.sql);
        assert!(out.sql.contains("LIMIT 1"), "{}", out.sql);
        let ShapeNode::Object { pointers, .. } = &out.shape.root else {
            panic!()
        };
        assert!(matches!(&pointers[2], ShapeNode::Scalar { name, .. } if name == "t"));
    }

    #[test]
    fn test_sub_select_field_access_over_a_single_link() {
        let out = compile_and_emit("SELECT Person { c := (select .company).name }");
        assert!(out.sql.contains("\"t1\".\"company_id\" = \"t2\".\"id\""), "{}", out.sql);
        assert!(out.sql.contains("(\"t1\".\"id\" = \"t0\".\"id\")"), "{}", out.sql);
    }

    #[test]
    fn test_sub_select_field_access_in_a_filter() {
        let out = compile_and_emit("SELECT Person { name } filter (select .posts limit 1).title = 'x'");
        assert!(out.sql.contains("WHERE ((SELECT \"t2\".\"title\""), "{}", out.sql);
        assert!(out.sql.contains("LIMIT 1) = 'x')"), "{}", out.sql);
    }

    #[test]
    fn test_sub_select_over_an_absolute_path_is_not_correlated() {
        let out = compile_and_emit("SELECT Person { n := (select Company.name limit 1) }");
        assert!(out.sql.contains("FROM \"public\".\"Company\""), "{}", out.sql);
        assert!(!out.sql.contains("= \"t0\".\"id\""), "{}", out.sql);
    }

    #[test]
    fn test_backlink_sub_select_carries_its_modifiers() {
        let out = compile_and_emit("SELECT Post { authors := (select .<posts[is Person] { name } limit 2) }");
        assert!(out.sql.contains("LIMIT 2"), "{}", out.sql);
        assert!(out.sql.contains("AS \"__lim\""), "{}", out.sql);
        let ShapeNode::Object { pointers, .. } = &out.shape.root else {
            panic!()
        };
        assert!(matches!(&pointers[1], ShapeNode::Array { name, .. } if name == "authors"));
    }

    #[test]
    fn test_with_bound_scalar_is_typed_for_overload_resolution() {
        // A `with`-bound value had no inferred type, so a call over one fell
        // back to the first registered overload and needed an explicit cast.
        let out = compile_and_emit("WITH xs := (select Person.name) SELECT contains(xs, 'a')");
        assert!(
            out.sql.contains("strpos((SELECT \"v\" FROM \"xs\"), 'a')"),
            "{}",
            out.sql
        );
    }

    /// `Person` with a translation-style computed model: one computed
    /// resolves a filtered subset of a multi-link, others read through it.
    fn make_schema_with_computed_links() -> SchemaDescriptor {
        use crate::schema::ComputedDescriptor;
        let mut schema = make_schema();
        let person = schema.types.iter_mut().find(|t| t.name == "Person").unwrap();
        person.computed = vec![
            ComputedDescriptor {
                name: "published".into(),
                expression: "(select .posts filter .title != '')".into(),
                return_type: None,
                link_target: None,
                link_multi: false,
            },
            ComputedDescriptor {
                name: "plain".into(),
                expression: ".posts".into(),
                return_type: None,
                link_target: None,
                link_multi: false,
            },
            ComputedDescriptor {
                name: "featured".into(),
                expression: "(select .published filter .title != 'draft')".into(),
                return_type: None,
                link_target: None,
                link_multi: false,
            },
            ComputedDescriptor {
                name: "looper".into(),
                expression: "(select .looper)".into(),
                return_type: None,
                link_target: None,
                link_multi: false,
            },
            ComputedDescriptor {
                name: "capped".into(),
                expression: "(select .posts limit 1)".into(),
                return_type: None,
                link_target: None,
                link_multi: false,
            },
        ];
        schema
    }

    /// `Person.recent` computed from an object-returning function call, the
    /// shape jaldis's translation model uses.
    fn make_schema_with_object_fn_computed() -> SchemaDescriptor {
        use crate::schema::{ComputedDescriptor, FunctionDescriptor, FunctionParamDescriptor};
        let mut schema = make_schema();
        schema.functions = vec![FunctionDescriptor {
            name: "latest".into(),
            module: "default".into(),
            params: vec![FunctionParamDescriptor {
                name: "owner".into(),
                pg_type: "uuid".into(),
            }],
            return_pg_type: "default::Post".into(),
            return_is_object: true,
            return_is_set: true,
            return_is_polymorphic: false,
            volatility: "stable".into(),
            body: String::new(),
        }];
        let person = schema.types.iter_mut().find(|t| t.name == "Person").unwrap();
        person.computed = vec![
            ComputedDescriptor {
                name: "recent".into(),
                expression: "latest(.id)".into(),
                return_type: None,
                link_target: None,
                link_multi: false,
            },
            ComputedDescriptor {
                name: "vetted".into(),
                expression: "(select latest(.id) filter .title != '')".into(),
                return_type: None,
                link_target: None,
                link_multi: false,
            },
        ];
        schema
    }

    #[test]
    fn test_path_traverses_through_a_function_backed_computed() {
        // `recent := latest(.id)` has no path to splice in, so the call
        // itself becomes the next row source — LATERAL, because its
        // arguments read the alias the traversal has reached.
        let schema = make_schema_with_object_fn_computed();
        let out = compile_and_emit_with("SELECT Person { t := .recent.title }", &schema);
        assert!(
            out.sql
                .contains("CROSS JOIN LATERAL \"public\".\"latest\"(\"t1\".\"id\") AS \"t2\""),
            "{}",
            out.sql
        );
        // The function returns a set, so the traversal is multi-valued.
        assert!(out.sql.contains("ARRAY(SELECT \"t2\".\"title\""), "{}", out.sql);
    }

    #[test]
    fn test_sub_select_over_a_function_backed_computed() {
        // The jaldis translation model, minimal: a function-backed computed
        // resolves the set, a scalar computed picks an attribute out of it.
        let schema = make_schema_with_object_fn_computed();
        let out = compile_and_emit_with(
            "SELECT Person { t := (select .recent filter .title = 'x' limit 1).title }",
            &schema,
        );
        assert!(
            out.sql.contains("CROSS JOIN LATERAL \"public\".\"latest\""),
            "{}",
            out.sql
        );
        assert!(out.sql.contains("\"t2\".\"title\" = 'x'"), "{}", out.sql);
        assert!(out.sql.contains("LIMIT 1"), "{}", out.sql);
        assert!(!out.sql.contains("ARRAY("), "{}", out.sql);
    }

    #[test]
    fn test_function_backed_computed_carries_its_own_filter() {
        let schema = make_schema_with_object_fn_computed();
        let out = compile_and_emit_with(
            "SELECT Person { t := (select .vetted filter .title = 'x' limit 1).title }",
            &schema,
        );
        assert!(
            out.sql
                .contains("(\"t2\".\"title\" = 'x') AND (\"t2\".\"title\" <> '')"),
            "{}",
            out.sql
        );
    }

    #[test]
    fn test_function_backed_computed_checks_its_argument_count() {
        use crate::schema::ComputedDescriptor;
        let mut schema = make_schema_with_object_fn_computed();
        let person = schema.types.iter_mut().find(|t| t.name == "Person").unwrap();
        person.computed = vec![ComputedDescriptor {
            name: "recent".into(),
            expression: "latest()".into(),
            return_type: None,
            link_target: None,
            link_multi: false,
        }];
        let ast = parse::parse("SELECT Person { t := .recent.title }").unwrap();
        let err = match ir::compile(&ast, &schema) {
            Ok(_) => panic!("expected a compile error"),
            Err(e) => format!("{e}"),
        };
        assert!(err.contains("expects 1 argument(s), got 0"), "{err}");
    }

    #[test]
    fn test_path_traverses_through_a_computed_pointer() {
        // `.published` is itself computed, so it has no column to join on —
        // its own path takes the step's place and its filter rides along.
        let schema = make_schema_with_computed_links();
        let out = compile_and_emit_with("SELECT Person { t := .published.title }", &schema);
        assert!(out.sql.contains("\"public\".\"Person.posts\""), "{}", out.sql);
        assert!(
            out.sql
                .contains("(\"t1\".\"id\" = \"t0\".\"id\") AND (\"t2\".\"title\" <> '')"),
            "{}",
            out.sql
        );
    }

    #[test]
    fn test_sub_select_over_a_computed_pointer_projects_a_property() {
        // The shape the jaldis translation model is built on:
        // `(select .translation filter .attribute = X limit 1).value`.
        let schema = make_schema_with_computed_links();
        let out = compile_and_emit_with(
            "SELECT Person { t := (select .published filter .title = 'a' limit 1).title }",
            &schema,
        );
        assert!(out.sql.contains("LIMIT 1"), "{}", out.sql);
        // Both the computed's own filter and the use-site one apply.
        assert!(out.sql.contains("\"t2\".\"title\" = 'a'"), "{}", out.sql);
        assert!(out.sql.contains("\"t2\".\"title\" <> ''"), "{}", out.sql);
    }

    #[test]
    fn test_chained_computed_pointers_each_contribute_their_filter() {
        let schema = make_schema_with_computed_links();
        let out = compile_and_emit_with("SELECT Person { t := .featured.title }", &schema);
        assert_eq!(out.sql.matches("\"public\".\"Person.posts\"").count(), 1, "{}", out.sql);
        assert!(out.sql.contains("\"t2\".\"title\" <> 'draft'"), "{}", out.sql);
        assert!(out.sql.contains("\"t2\".\"title\" <> ''"), "{}", out.sql);
    }

    #[test]
    fn test_computed_pointer_naming_a_bare_link_is_traversable() {
        let schema = make_schema_with_computed_links();
        let out = compile_and_emit_with("SELECT Person { t := .plain.title }", &schema);
        assert!(out.sql.contains("ARRAY(SELECT \"t2\".\"title\""), "{}", out.sql);
    }

    #[test]
    fn test_a_cycle_of_computed_pointers_is_reported_not_hung_on() {
        let schema = make_schema_with_computed_links();
        let ast = parse::parse("SELECT Person { t := .looper.title }").unwrap();
        let err = match ir::compile(&ast, &schema) {
            Ok(_) => panic!("expected a compile error"),
            Err(e) => format!("{e}"),
        };
        assert!(err.contains("expands into itself"), "{err}");
    }

    #[test]
    fn test_a_cycle_of_computed_pointers_read_as_a_value_is_reported_not_hung_on() {
        // Read as the last step, the computed was compiled as an expression,
        // which walked the same path again with no count of how often.
        let schema = make_schema_with_computed_links();
        for query in ["SELECT Person { looper }", "SELECT Person { * }"] {
            let ast = parse::parse(query).unwrap();
            let err = match ir::compile(&ast, &schema) {
                Ok(_) => panic!("expected a compile error for {query}"),
                Err(e) => format!("{e}"),
            };
            assert!(err.contains("expands into itself"), "{query}: {err}");
        }
    }

    #[test]
    fn test_traversing_through_a_limited_computed_uses_a_lateral() {
        // `capped` picks one row per Person, so the traversal correlates
        // through a LATERAL rather than a plain join — and its `limit 1`
        // means the result is one title, not an array of them.
        let schema = make_schema_with_computed_links();
        let out = compile_and_emit_with("SELECT Person { t := .capped.title }", &schema);
        assert!(out.sql.contains("JOIN LATERAL ("), "{}", out.sql);
        assert!(out.sql.contains("LIMIT 1"), "{}", out.sql);
        assert!(!out.sql.contains("ARRAY(SELECT"), "{}", out.sql);
    }

    #[test]
    fn test_backlink_from_a_type_whose_interface_the_link_targets() {
        use crate::schema::{LinkDescriptor, TypeDescriptor};
        let mut schema = make_interface_schema();
        let individual = schema.types.iter().find(|t| t.name == "Individual").unwrap().clone();
        schema.types.push(TypeDescriptor {
            name: "Note".into(),
            module: "default".into(),
            table: "Note".into(),
            abstract_: false,
            materialized: true,
            description: None,
            parents: vec![],
            interfaces: vec![],
            bases: vec![],
            properties: individual.properties[..1].to_vec(),
            // Targets the *interface*, which `Individual` implements.
            links: vec![LinkDescriptor {
                name: "owner".into(),
                target: "default::Account".into(),
                nullable: true,
                through: None,
                description: None,
                default_pyql: None,
                is_exclusive: false,
                is_readonly: false,
                rewrites: vec![],
                on_delete: vec![],
            }],
            multilinks: vec![],
            computed: vec![],
            constraints: vec![],
            indexes: vec![],
            partition: None,
            vector_indexes: vec![],
            search_indexes: vec![],
            triggers: vec![],
            junction: false,
            signals: vec![],
        });
        // Reported "link 'Note::owner' does not target 'Individual'" — but a
        // link to an interface accepts every implementor, and they share ids.
        let out = compile_and_emit_with("SELECT Individual { notes := .<owner[is Note] }", &schema);
        assert!(out.sql.contains("\"public\".\"Note\""), "{}", out.sql);
        assert!(out.sql.contains("\"owner_id\" = \"t0\".id"), "{}", out.sql);
    }

    #[test]
    fn test_with_block_in_expression_position_hoists_its_bindings() {
        // A `with` has nowhere to put its CTEs mid-expression, so they move
        // to the enclosing statement's own WITH clause.
        let out = compile_and_emit("SELECT Person { t := (with c := (select Company.name limit 1) select c) }");
        assert!(out.sql.starts_with("WITH\n\"c\" AS ("), "{}", out.sql);
        assert!(out.sql.contains("(SELECT \"v\" FROM \"c\")"), "{}", out.sql);
    }

    #[test]
    fn test_hoisted_binding_is_usable_in_the_expression_around_it() {
        let out =
            compile_and_emit("SELECT Person { t := (with c := (select Company.name limit 1) select .name ++ c) }");
        assert!(
            out.sql.contains("(\"t0\".\"name\" || (SELECT \"v\" FROM \"c\"))"),
            "{}",
            out.sql
        );
    }

    #[test]
    fn test_hoisted_binding_joins_the_statements_own_with_clause() {
        let out = compile_and_emit(
            "WITH top := (select Person.name limit 1) \
             SELECT Person { t := (with c := (select Company.name limit 1) select c) }",
        );
        let top = out.sql.find("\"top\" AS (").expect("top-level binding");
        let inner = out.sql.find("\"c\" AS (").expect("hoisted binding");
        assert!(top < inner, "the statement's own bindings come first:\n{}", out.sql);
    }

    #[test]
    fn test_a_binding_named_after_a_keyword_is_usable() {
        // The lexer is case-insensitive, so `order` arrives as the ORDER
        // token and read as "expected an expression, found 'order'" — and the
        // binding keeps the casing it was written with, not the keyword
        // table's.
        let out = compile_and_emit("WITH order := (select Person.name) SELECT Person { name } FILTER .name = order");
        assert!(out.sql.contains("(SELECT \"v\" FROM \"order\")"), "{}", out.sql);
    }

    #[test]
    fn test_computed_declared_on_an_interface_is_visible_from_an_implementor() {
        use crate::schema::ComputedDescriptor;
        let mut schema = make_interface_schema();
        let account = schema.types.iter_mut().find(|t| t.name == "Account").unwrap();
        account.computed = vec![ComputedDescriptor {
            name: "tier".into(),
            expression: "'gold'".into(),
            return_type: None,
            link_target: None,
            link_multi: false,
        }];
        let individual = schema.types.iter_mut().find(|t| t.name == "Individual").unwrap();
        individual.computed = vec![];
        // Reported as "has no link or property 'tier'. Did you mean 'tier'?"
        // — the suggester searched interfaces, the resolver didn't.
        for q in ["SELECT Individual { tier }", "SELECT Individual { t := .tier }"] {
            let out = compile_and_emit_with(q, &schema);
            assert!(out.sql.contains("'gold'"), "{q}: {}", out.sql);
        }
    }

    #[test]
    fn test_computed_may_carry_trailing_modifiers_without_a_select() {
        use crate::schema::ComputedDescriptor;
        let mut schema = make_schema();
        let person = schema.types.iter_mut().find(|t| t.name == "Person").unwrap();
        person.computed = vec![ComputedDescriptor {
            name: "recent".into(),
            expression: ".posts order by .title desc limit 5".into(),
            return_type: None,
            link_target: None,
            link_multi: false,
        }];
        // A schema fragment has no statement around it to hang modifiers
        // off, so this used to be "expected an expression, found 'order'".
        let out = compile_and_emit_with("SELECT Person { recent { title } }", &schema);
        assert!(out.sql.contains("LIMIT 5"), "{}", out.sql);
        assert!(out.sql.contains("ORDER BY \"t1\".\"title\" DESC"), "{}", out.sql);
    }

    #[test]
    fn test_computed_may_lead_with_a_bare_select() {
        use crate::schema::ComputedDescriptor;
        let mut schema = make_schema();
        let person = schema.types.iter_mut().find(|t| t.name == "Person").unwrap();
        person.computed = vec![ComputedDescriptor {
            name: "recent".into(),
            expression: "select .posts order by .title desc limit 5".into(),
            return_type: None,
            link_target: None,
            link_multi: false,
        }];
        let out = compile_and_emit_with("SELECT Person { recent { title } }", &schema);
        assert!(out.sql.contains("LIMIT 5"), "{}", out.sql);
    }

    #[test]
    fn test_sub_select_with_a_shape_may_project_a_property_off_it() {
        let out = compile_and_emit("SELECT Person { t := (select .posts { title } limit 1).title }");
        assert!(out.sql.contains("\"t2\".\"title\""), "{}", out.sql);
        assert!(out.sql.contains("LIMIT 1"), "{}", out.sql);
    }

    #[test]
    fn test_sub_statement_error_names_what_it_got() {
        let err = compile_err("SELECT Person { t := (insert Company { name := 'a' }) }");
        assert!(err.contains("an insert cannot stand in for a value"), "{err}");
    }

    #[test]
    fn test_with_bound_scalar_path_select_emits_a_real_select() {
        // The CTE body used to be just the FROM clause — `"xs" AS (
        // "public"."Person" AS "t0" )` — which is not SQL at all.
        let out = compile_and_emit("WITH xs := (select Person.name) SELECT Person { name }");
        assert!(
            out.sql
                .contains("SELECT ROW(v) AS result, v FROM (SELECT \"t0\".\"name\" AS v"),
            "{}",
            out.sql
        );
    }

    #[test]
    fn test_with_bound_scalar_path_select_is_typed_as_its_result() {
        // Binding `Person.name` binds text, not `default::Person`, so a
        // reference to it reads the CTE's value column rather than an `id`.
        let out = compile_and_emit("WITH xs := (select Person.name) SELECT Person { name } FILTER .name IN xs");
        assert!(out.sql.contains("ANY((SELECT \"v\" FROM \"xs\"))"), "{}", out.sql);
    }

    #[test]
    fn test_distinct_survives_a_with_binding() {
        let out = compile_and_emit("WITH xs := (select distinct Person.name) SELECT Person { name }");
        assert!(out.sql.contains("SELECT DISTINCT \"t0\".\"name\" AS v"), "{}", out.sql);
    }

    #[test]
    fn test_for_over_a_derived_set_iterates_every_row() {
        // A non-literal iterator used to be wrapped in a one-row VALUES, so
        // the body ran once against a scalar subquery instead of per row.
        let out = compile_and_emit("FOR x IN (select Person.name) UNION (SELECT Person { name } FILTER .name = x)");
        assert!(!out.sql.contains("VALUES"), "{}", out.sql);
        assert!(out.sql.contains(") AS \"_for_x\""), "{}", out.sql);
        assert!(out.sql.contains("CROSS JOIN LATERAL"), "{}", out.sql);
        assert!(out.sql.contains("\"t1\".\"name\" = \"_for_x\".\"v\""), "{}", out.sql);
    }

    #[test]
    fn test_for_insert_over_a_derived_set_uses_it_as_the_cte() {
        let out = compile_and_emit("FOR x IN (select Person.name) UNION (INSERT Company { name := x })");
        assert!(out.sql.starts_with("WITH \"_for_x\" AS ("), "{}", out.sql);
        assert!(
            out.sql.contains("SELECT \"_for_x\".\"v\" FROM \"_for_x\""),
            "{}",
            out.sql
        );
    }

    #[test]
    fn test_for_over_objects_binds_their_id() {
        let out = compile_and_emit("FOR p IN (select Person) UNION (INSERT Company { name := <str>p })");
        assert!(out.sql.contains("SELECT \"id\" AS v FROM ("), "{}", out.sql);
    }

    #[test]
    fn test_for_over_a_non_select_statement_is_rejected() {
        let ast = parse::parse("FOR x IN (INSERT Company { name := 'a' }) UNION (SELECT Person { name })").unwrap();
        let schema = make_schema();
        let err = match ir::compile(&ast, &schema) {
            Ok(_) => panic!("expected a compile error"),
            Err(e) => format!("{e}"),
        };
        assert!(err.contains("only a select can be iterated"), "{err}");
    }

    #[test]
    fn test_multi_link_path_outside_a_comparison_is_an_array() {
        // It stands for a set of values — as a scalar subquery Postgres
        // would reject it the moment a Person had two posts.
        let out = compile_and_emit("SELECT Person { t := .posts.title }");
        assert!(out.sql.contains("ARRAY(SELECT \"t2\".\"title\""), "{}", out.sql);
        assert!(out.sql.contains("(\"t1\".\"id\" = \"t0\".\"id\")"), "{}", out.sql);
    }

    #[test]
    fn test_multi_link_path_inside_a_comparison_is_still_exists() {
        let out = compile_and_emit("SELECT Person { name } filter .posts.title = 'x'");
        assert!(out.sql.contains("WHERE EXISTS("), "{}", out.sql);
        assert!(!out.sql.contains("ARRAY("), "{}", out.sql);
    }

    #[test]
    fn test_path_traversal_deeper_than_two_steps() {
        let schema = make_schema_with_through();
        let out = compile_and_emit_with("SELECT Person { t := .friends.friends.name }", &schema);
        // Two junction hops, then the property — one subquery, not an error.
        assert_eq!(out.sql.matches("\"public\".\"PersonFriend\"").count(), 2, "{}", out.sql);
        assert!(out.sql.contains("ARRAY(SELECT \"t4\".\"name\""), "{}", out.sql);
    }

    #[test]
    fn test_coalesce_over_backlinked_values_is_a_set_choice() {
        // The backlink was taken for a comparison and rewritten to an EXISTS,
        // with a COALESCE of a value and an array inside it.
        let out = compile_and_emit("SELECT Post ORDER BY max(.<posts[is Person].age ?? .<posts[is Person].age)");
        assert!(!out.sql.contains("EXISTS"), "{}", out.sql);
        assert!(!out.sql.contains("COALESCE("), "{}", out.sql);
        assert!(out.sql.contains("CASE WHEN (cardinality("), "{}", out.sql);
        assert!(
            out.sql
                .contains("(SELECT coalesce(max(\"_s\".\"v\"), NULL) FROM unnest("),
            "{}",
            out.sql
        );
    }

    #[test]
    fn test_coalesce_over_multilink_values_is_a_set_choice() {
        let out = compile_and_emit("SELECT Person { t := .posts.title ?? .posts.title }");
        assert!(!out.sql.contains("EXISTS"), "{}", out.sql);
        assert!(out.sql.contains("CASE WHEN (cardinality("), "{}", out.sql);
    }

    #[test]
    fn test_coalesce_of_a_set_with_a_single_value_falls_back_to_it() {
        let out = compile_and_emit("SELECT Post { a := .<posts[is Person].age ?? 0 }");
        assert!(out.sql.contains("ELSE array_remove(ARRAY[0], NULL) END"), "{}", out.sql);
    }

    #[test]
    fn test_traversal_after_a_backlink() {
        let out = compile_and_emit("SELECT Post { t := .<posts[is Person].name }");
        assert!(out.sql.contains("ARRAY(SELECT \"t2\".\"name\""), "{}", out.sql);
        assert!(out.sql.contains("\"t3\".\"target\" = \"t1\".\"id\""), "{}", out.sql);
    }

    #[test]
    fn test_computed_pointer_on_a_linked_type_is_reachable() {
        use crate::schema::ComputedDescriptor;
        let mut schema = make_schema();
        let company = schema.types.iter_mut().find(|t| t.name == "Company").unwrap();
        company.computed = vec![ComputedDescriptor {
            name: "shout".into(),
            expression: ".name ++ '!'".into(),
            return_type: Some("text".into()),
            link_target: None,
            link_multi: false,
        }];
        // `.company.shout` used to report "has no link or property 'shout'"
        // while helpfully suggesting 'shout' — path traversal never looked
        // at the target type's computed pointers.
        let out = compile_and_emit_with("SELECT Person { t := .company.shout }", &schema);
        assert!(out.sql.contains("(\"t2\".\"name\" || '!')"), "{}", out.sql);
    }

    #[test]
    fn test_computed_pointer_behind_a_type_intersection() {
        let schema = make_interface_schema();
        let out = compile_and_emit_with("SELECT Account { n := [is Individual].full_name }", &schema);
        assert!(out.sql.contains("upper(\"t1\".\"first_name\")"), "{}", out.sql);
        assert!(out.sql.contains("FROM \"public\".\"Individual\""), "{}", out.sql);
        assert!(out.sql.contains("(\"t1\".\"id\" = \"t0\".\"id\")"), "{}", out.sql);
    }

    #[test]
    fn test_computed_pointer_cannot_be_traversed_through() {
        use crate::schema::ComputedDescriptor;
        let mut schema = make_schema();
        let company = schema.types.iter_mut().find(|t| t.name == "Company").unwrap();
        company.computed = vec![ComputedDescriptor {
            name: "shout".into(),
            expression: ".name ++ '!'".into(),
            return_type: Some("text".into()),
            link_target: None,
            link_multi: false,
        }];
        let ast = parse::parse("SELECT Person { t := .company.shout.nope }").unwrap();
        let err = match ir::compile(&ast, &schema) {
            Ok(_) => panic!("expected a compile error"),
            Err(e) => format!("{e}"),
        };
        assert!(err.contains("is a computed pointer"), "{err}");
    }

    #[test]
    fn test_sub_select_over_a_multi_link_path_keeps_its_modifiers() {
        let out = compile_and_emit("SELECT Person { t := (select .posts.title order by .title desc limit 2) }");
        assert!(out.sql.contains("ARRAY(SELECT"), "{}", out.sql);
        assert!(out.sql.contains("ORDER BY \"t2\".\"title\" DESC"), "{}", out.sql);
        assert!(out.sql.contains("LIMIT 2"), "{}", out.sql);
    }

    #[test]
    fn test_limit_one_over_a_multi_link_path_stays_a_scalar() {
        // `limit 1` is what makes the set single-valued, so this one is a
        // scalar subquery rather than a one-element array.
        let out = compile_and_emit("SELECT Person { t := (select .posts.title limit 1) }");
        assert!(!out.sql.contains("ARRAY("), "{}", out.sql);
        assert!(out.sql.contains("LIMIT 1"), "{}", out.sql);
    }

    #[test]
    fn test_computed_multilink_can_carry_a_nested_shape() {
        let out = compile_and_emit("SELECT Person { p := .posts { title } }");
        let ShapeNode::Object { pointers, .. } = &out.shape.root else {
            panic!()
        };
        let ShapeNode::Array { name, element, .. } = &pointers[2] else {
            panic!("{:?}", pointers[2])
        };
        assert_eq!(name, "p");
        let ShapeNode::Object {
            pointers: elem_pointers,
            ..
        } = element.as_ref()
        else {
            panic!()
        };
        assert!(matches!(&elem_pointers[1], ShapeNode::Scalar { name, .. } if name == "title"));
    }

    #[test]
    fn test_schema_declared_backlink_computed_is_an_object_pointer() {
        use crate::schema::ComputedDescriptor;
        let mut schema = make_schema();
        let post = schema.types.iter_mut().find(|t| t.name == "Post").unwrap();
        post.computed = vec![ComputedDescriptor {
            name: "authors".into(),
            expression: ".<posts[is Person]".into(),
            return_type: None,
            link_target: None,
            link_multi: false,
        }];
        // A backlink computed used to compile as an EXISTS boolean here —
        // the pointer builders were only reachable from an inline `:=`.
        let out = compile_and_emit_with("SELECT Post { authors { name } }", &schema);
        assert!(out.sql.contains("array_agg(ROW("), "{}", out.sql);
        assert!(out.sql.contains("\"t1\".\"name\"::text"), "{}", out.sql);
        let ShapeNode::Object { pointers, .. } = &out.shape.root else {
            panic!()
        };
        assert!(matches!(&pointers[1], ShapeNode::Array { name, .. } if name == "authors"));
    }

    #[test]
    fn test_schema_declared_multilink_computed_defaults_to_ids_without_a_shape() {
        use crate::schema::ComputedDescriptor;
        let mut schema = make_schema();
        schema.types[0].computed = vec![ComputedDescriptor {
            name: "everything".into(),
            expression: ".posts".into(),
            return_type: None,
            link_target: None,
            link_multi: false,
        }];
        let out = compile_and_emit_with("SELECT Person { everything }", &schema);
        let ShapeNode::Object { pointers, .. } = &out.shape.root else {
            panic!()
        };
        assert!(matches!(&pointers[2], ShapeNode::Array { name, .. } if name == "everything"));
    }

    #[test]
    fn test_schema_declared_computed_sub_select_is_a_link_pointer() {
        use crate::schema::ComputedDescriptor;
        let mut schema = make_schema();
        schema.types[0].computed = vec![
            ComputedDescriptor {
                name: "recent".into(),
                expression: "(select .posts order by .title desc limit 1)".into(),
                return_type: None,
                link_target: None,
                link_multi: false,
            },
            ComputedDescriptor {
                name: "recent_title".into(),
                expression: "(select .posts order by .title desc limit 1).title".into(),
                return_type: Some("text".into()),
                link_target: None,
                link_multi: false,
            },
        ];
        let out = compile_and_emit_with("SELECT Person { recent { title }, recent_title }", &schema);
        let ShapeNode::Object { pointers, .. } = &out.shape.root else {
            panic!()
        };
        // A declared computed gets the same treatment as one written inline:
        // object-valued when it selects a link, scalar when it projects one.
        assert!(matches!(&pointers[2], ShapeNode::Object { name, .. } if name == "recent"));
        assert!(matches!(&pointers[3], ShapeNode::Scalar { name, .. } if name == "recent_title"));
        assert!(out.sql.contains("\"jt\".source = \"t0\".id"), "{}", out.sql);
        assert_eq!(out.sql.matches("LIMIT 1").count(), 2, "{}", out.sql);
        crate::validate::validate_schema_types(&schema).expect("schema should validate");
    }

    /// `select (insert T { … }).id` — the mutation is hoisted into the
    /// statement's own WITH and the field read off it, which is what the
    /// `with x := (insert …) select x.id` spelling already produced. Without
    /// the CTE the projection would read the whole table instead.
    #[test]
    fn test_a_mutation_read_through_a_path_runs_as_a_hoisted_cte() {
        let out = compile_and_emit("SELECT (INSERT Person { name := 'x' }).name");
        assert!(
            out.sql.contains("\"_nested_dml_0\" AS (") && out.sql.contains("INSERT INTO"),
            "the mutation must run as a CTE:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("FROM \"_nested_dml_0\""),
            "the projection must read the mutation's own rows:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_dml_sub_statement_in_expression_position_still_rejected() {
        let err = compile_err("SELECT Person { x := (insert Company { name := 'c' }).name }");
        assert!(err.contains("an insert cannot stand in for a value"), "{err}");
    }

    #[test]
    fn test_sub_select_with_a_shape_in_expression_position_is_rejected() {
        // Only when the shape *is* the result — projecting a property off
        // it is fine, since the shape then says nothing the projection
        // doesn't (see test_sub_select_with_a_shape_may_project_a_property_off_it).
        let err = compile_err("SELECT Person { name } filter (select .posts { title }) = 1");
        assert!(err.contains("sub-select with a shape"), "{err}");
    }

    fn make_schema_with_through() -> SchemaDescriptor {
        let id_prop = || PropertyDescriptor {
            name: "id".into(),
            pg_type: "uuid".into(),
            nullable: false,
            default_sql: Some("gen_random_uuid()".into()),
            description: None,
            default_pyql: None,
            check_constraints: vec![],
            is_exclusive: true,
            is_pk: true,
            is_readonly: true,
            rewrites: vec![],
            tuple_members: None,
            column_type: None,
        };
        let name_prop = || PropertyDescriptor {
            name: "name".into(),
            pg_type: "text".into(),
            nullable: false,
            default_sql: None,
            description: None,
            check_constraints: vec![],
            default_pyql: None,
            is_exclusive: false,
            is_pk: false,
            is_readonly: false,
            rewrites: vec![],
            tuple_members: None,
            column_type: None,
        };
        SchemaDescriptor {
            types: vec![
                TypeDescriptor {
                    name: "Person".into(),
                    module: "default".into(),
                    table: "Person".into(),
                    abstract_: false,
                    materialized: false,
                    description: None,
                    parents: vec![],
                    interfaces: vec![],
                    bases: vec![],
                    properties: vec![id_prop(), name_prop()],
                    links: vec![],
                    multilinks: vec![MultiLinkDescriptor {
                        name: "friends".into(),
                        target: "default::Person".into(),
                        through: Some("default::PersonFriend".into()),
                        nullable: false,
                        description: None,
                        default_pyql: None,
                        on_delete: vec![],
                        is_exclusive: false,
                    }],
                    computed: vec![],
                    constraints: vec![],
                    indexes: vec![],
                    partition: None,
                    vector_indexes: vec![],
                    search_indexes: vec![],
                    triggers: vec![],
                    junction: false,
                    signals: vec![],
                },
                TypeDescriptor {
                    name: "PersonFriend".into(),
                    module: "default".into(),
                    table: "PersonFriend".into(),
                    abstract_: false,
                    materialized: false,
                    description: None,
                    parents: vec![],
                    interfaces: vec![],
                    bases: vec![],
                    properties: vec![id_prop()],
                    links: vec![
                        LinkDescriptor {
                            name: "person".into(),
                            target: "default::Person".into(),
                            nullable: false,
                            through: None,
                            description: None,
                            default_pyql: None,
                            is_exclusive: false,
                            is_readonly: false,
                            rewrites: vec![],
                            on_delete: vec![],
                        },
                        LinkDescriptor {
                            name: "friend".into(),
                            target: "default::Person".into(),
                            nullable: false,
                            through: None,
                            description: None,
                            default_pyql: None,
                            is_exclusive: false,
                            is_readonly: false,
                            rewrites: vec![],
                            on_delete: vec![],
                        },
                    ],
                    multilinks: vec![],
                    computed: vec![],
                    constraints: vec![],
                    indexes: vec![],
                    partition: None,
                    vector_indexes: vec![],
                    search_indexes: vec![],
                    triggers: vec![],
                    junction: false,
                    signals: vec![],
                },
            ],
            scalars: vec![],
            enums: vec![],
            named_tuples: vec![],
            globals: vec![],
            functions: vec![],
            aliases: vec![],
            channels: vec![],
            ..Default::default()
        }
    }

    #[test]
    fn test_select_through_multi_link() {
        let schema = make_schema_with_through();
        let ast = crate::parse::parse("SELECT Person { name, friends { name } }").unwrap();
        let ir = crate::ir::compile(&ast, &schema).unwrap();
        let out = emit(&ir);
        // Junction table is the PersonFriend table, not the standard dotted name
        assert!(out.sql.contains("\"public\".\"PersonFriend\""));
        // Source FK column (person → Person) and target FK column (friend → Person)
        assert!(out.sql.contains("\"friend\""));
        assert!(out.sql.contains("\"person\""));
        // Still emits array_agg pattern
        assert!(out.sql.contains("array_agg(ROW("));
    }

    /// Person/Org — a junction-backed single link (`Person.spouse: Link[Org,
    /// through(Marriage)]`, exclusive). Deliberately targets a different
    /// type (Org) rather than a self-link, for the same reason
    /// `test_select_through_multi_link` avoids Person-to-Person — see the
    /// comment on `make_schema_with_through` above.
    #[test]
    fn test_limited_multi_link_keeps_a_nested_multi_link_in_scope() {
        // The nested pointer's own correlated subquery moves inside the
        // derived table the LIMIT sits on, where its alias is still bound.
        let schema = make_schema_with_through();
        let ast = crate::parse::parse("SELECT Person { friends: { name, friends { name } } limit 1 }").unwrap();
        let ir = crate::ir::compile(&ast, &schema).unwrap();
        let out = emit(&ir);
        assert_eq!(out.sql.matches("array_agg(ROW(").count(), 2, "{}", out.sql);
        let derived = out.sql.find("AS \"__lim\"").unwrap();
        let nested = out.sql.rfind("array_agg(ROW(").unwrap();
        assert!(
            nested < derived,
            "nested aggregate must sit inside the derived table:\n{}",
            out.sql
        );
    }

    fn make_schema_with_junction_backed_link() -> SchemaDescriptor {
        let id_prop = || PropertyDescriptor {
            name: "id".into(),
            pg_type: "uuid".into(),
            nullable: false,
            default_sql: Some("gen_random_uuid()".into()),
            description: None,
            default_pyql: None,
            check_constraints: vec![],
            is_exclusive: true,
            is_pk: true,
            is_readonly: true,
            rewrites: vec![],
            tuple_members: None,
            column_type: None,
        };
        let name_prop = || PropertyDescriptor {
            name: "name".into(),
            pg_type: "text".into(),
            nullable: false,
            default_sql: None,
            description: None,
            check_constraints: vec![],
            default_pyql: None,
            is_exclusive: false,
            is_pk: false,
            is_readonly: false,
            rewrites: vec![],
            tuple_members: None,
            column_type: None,
        };
        SchemaDescriptor {
            types: vec![
                TypeDescriptor {
                    name: "Person".into(),
                    module: "default".into(),
                    table: "Person".into(),
                    abstract_: false,
                    materialized: false,
                    description: None,
                    parents: vec![],
                    interfaces: vec![],
                    bases: vec![],
                    properties: vec![id_prop(), name_prop()],
                    links: vec![LinkDescriptor {
                        name: "spouse".into(),
                        target: "default::Org".into(),
                        nullable: true,
                        through: Some("default::Marriage".into()),
                        description: None,
                        default_pyql: None,
                        is_exclusive: true,
                        is_readonly: false,
                        rewrites: vec![],
                        on_delete: vec![],
                    }],
                    multilinks: vec![],
                    computed: vec![],
                    constraints: vec![],
                    indexes: vec![],
                    partition: None,
                    vector_indexes: vec![],
                    search_indexes: vec![],
                    triggers: vec![],
                    junction: false,
                    signals: vec![],
                },
                TypeDescriptor {
                    name: "Org".into(),
                    module: "default".into(),
                    table: "Org".into(),
                    abstract_: false,
                    materialized: false,
                    description: None,
                    parents: vec![],
                    interfaces: vec![],
                    bases: vec![],
                    properties: vec![id_prop(), name_prop()],
                    links: vec![],
                    multilinks: vec![],
                    computed: vec![],
                    constraints: vec![],
                    indexes: vec![],
                    partition: None,
                    vector_indexes: vec![],
                    search_indexes: vec![],
                    triggers: vec![],
                    junction: false,
                    signals: vec![],
                },
                TypeDescriptor {
                    name: "Marriage".into(),
                    module: "default".into(),
                    table: "Person.spouse".into(),
                    abstract_: false,
                    materialized: false,
                    description: None,
                    parents: vec![],
                    interfaces: vec![],
                    bases: vec![],
                    properties: vec![
                        id_prop(),
                        PropertyDescriptor {
                            name: "since".into(),
                            pg_type: "int8".into(),
                            nullable: true,
                            default_sql: None,
                            description: None,
                            check_constraints: vec![],
                            default_pyql: None,
                            is_exclusive: false,
                            is_pk: false,
                            is_readonly: false,
                            rewrites: vec![],
                            tuple_members: None,
                            column_type: None,
                        },
                    ],
                    links: vec![
                        LinkDescriptor {
                            name: "source".into(),
                            target: "default::Person".into(),
                            nullable: false,
                            through: None,
                            description: None,
                            default_pyql: None,
                            is_exclusive: false,
                            is_readonly: false,
                            rewrites: vec![],
                            on_delete: vec![],
                        },
                        LinkDescriptor {
                            name: "target".into(),
                            target: "default::Org".into(),
                            nullable: false,
                            through: None,
                            description: None,
                            default_pyql: None,
                            is_exclusive: false,
                            is_readonly: false,
                            rewrites: vec![],
                            on_delete: vec![],
                        },
                    ],
                    multilinks: vec![],
                    computed: vec![],
                    constraints: vec![],
                    indexes: vec![],
                    partition: None,
                    vector_indexes: vec![],
                    search_indexes: vec![],
                    triggers: vec![],
                    junction: true,
                    signals: vec![],
                },
            ],
            scalars: vec![],
            enums: vec![],
            named_tuples: vec![],
            globals: vec![],
            functions: vec![],
            aliases: vec![],
            channels: vec![],
            ..Default::default()
        }
    }

    #[test]
    fn test_select_shape_over_junction_backed_single_link() {
        let schema = make_schema_with_junction_backed_link();
        let ast = crate::parse::parse("SELECT Person { name, spouse { name } }").unwrap();
        let ir = crate::ir::compile(&ast, &schema).unwrap();
        let out = emit(&ir);
        // Junction table is the through type's own table, joined like a multi-link's.
        assert!(out.sql.contains("\"public\".\"Person.spouse\""), "got:\n{}", out.sql);
        assert!(out.sql.contains("\"jt\".source"), "got:\n{}", out.sql);
        assert!(out.sql.contains("\"jt\".target"), "got:\n{}", out.sql);
        // Cardinality-one: a scalar correlated subquery, never array_agg.
        assert!(!out.sql.contains("array_agg"), "got:\n{}", out.sql);
        assert!(!out.sql.contains("spouse_id"), "got:\n{}", out.sql);
    }

    #[test]
    fn test_select_shape_over_junction_backed_single_link_with_link_property() {
        // A junction-backed single link can read its own link properties
        // (`@since`) in a nested shape, same as a multi-link's `@weight` —
        // this is the one read site with no reusable FK-based form
        // (`IrSingleLinkPointer` needed its own `link_properties`, mirroring
        // `IrMultiLinkPointer`'s).
        let schema = make_schema_with_junction_backed_link();
        let ast = crate::parse::parse("SELECT Person { name, spouse { name, @since } }").unwrap();
        let ir = crate::ir::compile(&ast, &schema).unwrap();
        let out = emit(&ir);
        assert!(out.sql.contains("\"jt\".\"since\""), "got:\n{}", out.sql);
        assert!(!out.sql.contains("array_agg"), "got:\n{}", out.sql);
    }

    #[test]
    fn test_select_path_over_junction_backed_single_link() {
        let schema = make_schema_with_junction_backed_link();
        let ast = crate::parse::parse("SELECT Person.spouse { name }").unwrap();
        let ir = crate::ir::compile(&ast, &schema).unwrap();
        let out = emit(&ir);
        assert!(out.sql.contains("\"public\".\"Person.spouse\""), "got:\n{}", out.sql);
        assert!(out.sql.contains("\"source\""), "got:\n{}", out.sql);
        assert!(out.sql.contains("\"target\""), "got:\n{}", out.sql);
        assert!(!out.sql.contains("spouse_id"), "got:\n{}", out.sql);
    }

    #[test]
    fn test_insert_with_junction_backed_single_link() {
        let schema = make_schema_with_junction_backed_link();
        let out = compile_and_emit_with(
            "INSERT Person { name := $name, \
                spouse := (SELECT Org FILTER .id = $oid) { @since := <int64>$since } }",
            &schema,
        );
        assert!(out.sql.contains("\"public\".\"Person.spouse\""), "got:\n{}", out.sql);
        assert!(out.sql.contains("\"since\""), "got:\n{}", out.sql);
        assert!(!out.sql.contains("spouse_id"), "got:\n{}", out.sql);
        // A fresh row has no prior junction entry, but the emitted ON
        // CONFLICT target must still name only `source` — no `(source,
        // target)` composite unique constraint exists on this table (D3).
        // The conflict action is always an upsert (not DO NOTHING) since a
        // same-statement reassignment's paired DELETE shares this
        // statement's snapshot and can still "conflict" with the old row.
        assert!(out.sql.contains("ON CONFLICT (\"source\") DO UPDATE SET \"target\" = EXCLUDED.\"target\", \"since\" = EXCLUDED.\"since\""), "got:\n{}", out.sql);
    }

    #[test]
    fn test_update_replace_junction_backed_single_link() {
        let schema = make_schema_with_junction_backed_link();
        let out = compile_and_emit_with(
            "UPDATE Person FILTER .id = $id SET { \
                spouse := (SELECT Org FILTER .id = $oid) { @since := <int64>$since } }",
            &schema,
        );
        // Replace = clear the existing junction row, then insert the new one.
        assert!(
            out.sql.contains("DELETE FROM \"public\".\"Person.spouse\""),
            "got:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("INSERT INTO \"public\".\"Person.spouse\""),
            "got:\n{}",
            out.sql
        );
        assert!(out.sql.contains("ON CONFLICT (\"source\") DO UPDATE SET \"target\" = EXCLUDED.\"target\", \"since\" = EXCLUDED.\"since\""), "got:\n{}", out.sql);
        // The paired DELETE must not race the new INSERT for a target that's
        // being kept, but for a single link (PK is `source` alone) ANY
        // existing row conflicts regardless of target, so the upsert above
        // is what actually makes reassignment work, not the exclusion here.
        assert!(out.sql.contains("NOT IN"), "got:\n{}", out.sql);
    }

    /// `unless conflict on .k else (update T filter … set { … })` — the ON
    /// CONFLICT target says which row conflicts, the filter says whether to
    /// touch it. Dropping the filter turned automator's advisory lock into an
    /// unconditional steal of a live holder's row.
    #[test]
    fn test_unless_conflict_else_update_keeps_its_filter() {
        let out = compile_and_emit(
            "INSERT Person { name := 'a' } UNLESS CONFLICT ON .name \
             ELSE (UPDATE Person FILTER .age < 30 SET { name := 'b' })",
        );
        assert!(
            out.sql.contains("DO UPDATE SET") && out.sql.contains("WHERE"),
            "the ELSE UPDATE's filter must survive as a DO UPDATE predicate:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("\"Person\".\"age\""),
            "the predicate must read the existing row, not `excluded`:\n{}",
            out.sql
        );
    }

    /// No filter, no predicate — an unconditional ELSE UPDATE still updates.
    #[test]
    fn test_unless_conflict_else_update_without_a_filter_has_no_predicate() {
        let out = compile_and_emit(
            "INSERT Person { name := 'a' } UNLESS CONFLICT ON .name ELSE (UPDATE Person SET { name := 'b' })",
        );
        let after_set = out.sql.split("DO UPDATE SET").nth(1).unwrap_or_default();
        assert!(
            !after_set.contains("WHERE"),
            "got a predicate we never asked for:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_junction_backed_single_link_rejected_in_unless_conflict_else() {
        // `compile_conflict_else` calls `compile_assignments_for_update`
        // directly on the raw ELSE-clause shape, bypassing the
        // shape-classification loop `compile_insert`/`compile_update` use to
        // intercept a junction-backed link before it reaches the plain
        // FK-column assignment path — must fail loudly here instead of
        // emitting a `spouse_id` column that doesn't exist.
        let schema = make_schema_with_junction_backed_link();
        let ast = parse::parse(
            "INSERT Person { name := $name } \
             UNLESS CONFLICT ON .name ELSE (UPDATE Person SET { \
                spouse := (SELECT Org FILTER .id = $oid) })",
        )
        .unwrap();
        assert!(ir::compile(&ast, &schema).is_err());
    }

    #[test]
    fn test_update_clear_junction_backed_single_link() {
        let schema = make_schema_with_junction_backed_link();
        let out = compile_and_emit_with("UPDATE Person FILTER .id = $id SET { spouse := {} }", &schema);
        assert!(
            out.sql.contains("DELETE FROM \"public\".\"Person.spouse\""),
            "got:\n{}",
            out.sql
        );
        assert!(
            !out.sql.contains("INSERT INTO \"public\".\"Person.spouse\""),
            "clearing must not also insert:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_update_clear_junction_backed_single_link_with_cast_empty_set() {
        // Regression: the frontend (generateStatements.ts's `setNull`
        // handling) always emits the cast form `<Type>{}`, never the bare
        // `{}` — the junction-backed clear path originally only recognized
        // bare `Expr::Set([])`, so this form fell through to
        // `compile_multilink_values` and errored with "multilink value must
        // be a CTE reference, parenthesised subquery, or type path" instead
        // of clearing (confirmed live via the Data Explorer).
        let schema = make_schema_with_junction_backed_link();
        let out = compile_and_emit_with("UPDATE Person FILTER .id = $id SET { spouse := <Org>{} }", &schema);
        assert!(
            out.sql.contains("DELETE FROM \"public\".\"Person.spouse\""),
            "got:\n{}",
            out.sql
        );
        assert!(
            !out.sql.contains("INSERT INTO \"public\".\"Person.spouse\""),
            "clearing must not also insert:\n{}",
            out.sql
        );
    }

    /// Product/Tag/ProductTag — the actual real-world shape link properties
    /// were built for (mirrors pylon-demo). Deliberately *not* built off
    /// `make_schema_with_through` (Person self-linking to Person): a
    /// same-type through-link hits a pre-existing, unrelated column-
    /// resolution bug in `multilink_junction_info` (both source_col and
    /// target_col resolve to the same link name when source type == target
    /// type) — tracked separately, not fixed here.
    fn make_schema_with_through_and_prop() -> SchemaDescriptor {
        let id_prop = || PropertyDescriptor {
            name: "id".into(),
            pg_type: "uuid".into(),
            nullable: false,
            default_sql: Some("gen_random_uuid()".into()),
            description: None,
            default_pyql: None,
            check_constraints: vec![],
            is_exclusive: true,
            is_pk: true,
            is_readonly: true,
            rewrites: vec![],
            tuple_members: None,
            column_type: None,
        };
        let name_prop = || PropertyDescriptor {
            name: "name".into(),
            pg_type: "text".into(),
            nullable: false,
            default_sql: None,
            description: None,
            check_constraints: vec![],
            default_pyql: None,
            is_exclusive: false,
            is_pk: false,
            is_readonly: false,
            rewrites: vec![],
            tuple_members: None,
            column_type: None,
        };
        SchemaDescriptor {
            types: vec![
                TypeDescriptor {
                    name: "Product".into(),
                    module: "default".into(),
                    table: "Product".into(),
                    abstract_: false,
                    materialized: false,
                    description: None,
                    parents: vec![],
                    interfaces: vec![],
                    bases: vec![],
                    properties: vec![id_prop(), name_prop()],
                    links: vec![],
                    multilinks: vec![MultiLinkDescriptor {
                        name: "tags".into(),
                        target: "default::Tag".into(),
                        through: Some("default::ProductTag".into()),
                        nullable: false,
                        description: None,
                        default_pyql: None,
                        on_delete: vec![],
                        is_exclusive: false,
                    }],
                    computed: vec![],
                    constraints: vec![],
                    indexes: vec![],
                    partition: None,
                    vector_indexes: vec![],
                    search_indexes: vec![],
                    triggers: vec![],
                    junction: false,
                    signals: vec![],
                },
                TypeDescriptor {
                    name: "Tag".into(),
                    module: "default".into(),
                    table: "Tag".into(),
                    abstract_: false,
                    materialized: false,
                    description: None,
                    parents: vec![],
                    interfaces: vec![],
                    bases: vec![],
                    properties: vec![id_prop(), name_prop()],
                    links: vec![],
                    multilinks: vec![],
                    computed: vec![],
                    constraints: vec![],
                    indexes: vec![],
                    partition: None,
                    vector_indexes: vec![],
                    search_indexes: vec![],
                    triggers: vec![],
                    junction: false,
                    signals: vec![],
                },
                TypeDescriptor {
                    name: "ProductTag".into(),
                    module: "default".into(),
                    table: "Product.tags".into(),
                    abstract_: false,
                    materialized: false,
                    description: None,
                    parents: vec![],
                    interfaces: vec![],
                    bases: vec![],
                    properties: vec![
                        id_prop(),
                        PropertyDescriptor {
                            name: "weight".into(),
                            pg_type: "float8".into(),
                            nullable: false,
                            default_sql: None,
                            default_pyql: None,
                            description: None,
                            check_constraints: vec![],
                            is_exclusive: false,
                            is_pk: false,
                            is_readonly: false,
                            rewrites: vec![],
                            tuple_members: None,
                            column_type: None,
                        },
                    ],
                    // No declared Link pointers — matches pylon-demo's actual
                    // ProductTag, which relies on multilink_junction_info's
                    // "source"/"target" defaults.
                    links: vec![],
                    multilinks: vec![],
                    computed: vec![],
                    constraints: vec![],
                    indexes: vec![],
                    partition: None,
                    vector_indexes: vec![],
                    search_indexes: vec![],
                    triggers: vec![],
                    junction: true,
                    signals: vec![],
                },
            ],
            scalars: vec![],
            enums: vec![],
            named_tuples: vec![],
            globals: vec![],
            functions: vec![],
            aliases: vec![],
            channels: vec![],
            ..Default::default()
        }
    }

    #[test]
    fn test_link_property_in_a_links_own_filter() {
        // `filter @weight > x` reads the junction row of the link being
        // filtered — `@prop` used to parse only as a postfix on a named
        // path, so a bare one was "expected an expression, found '@'".
        let schema = make_schema_with_through_and_prop();
        let out = compile_and_emit_with("SELECT Product { tags: { } filter @weight > 0.5 }", &schema);
        assert!(out.sql.contains("\"jt\".\"weight\" > (0.5::float8)"), "{}", out.sql);
    }

    #[test]
    fn test_exists_on_a_link_property() {
        // `exists @weight` — the junction column the link's own modifiers are
        // compiled against. `exists` only knew how to take a pointer name.
        let schema = make_schema_with_through_and_prop();
        let out = compile_and_emit_with("SELECT Product { tags: { } filter exists @weight }", &schema);
        assert!(out.sql.contains("\"jt\".\"weight\" IS NOT NULL"), "{}", out.sql);
    }

    #[test]
    fn test_link_property_in_a_computed_link_filter() {
        let schema = make_schema_with_through_and_prop();
        let out = compile_and_emit_with(
            "SELECT Product { best := (select .tags filter @weight > 0.5 limit 1) { @weight } }",
            &schema,
        );
        assert!(out.sql.contains("\"jt\".\"weight\" > (0.5::float8)"), "{}", out.sql);
        assert!(out.sql.contains("LIMIT 1"), "{}", out.sql);
    }

    #[test]
    fn test_link_property_in_a_projected_sub_select_reads_the_junction() {
        // The flat-join form names the junction with an ordinary alias
        // rather than "jt", so the scope has to carry the alias too.
        let schema = make_schema_with_through_and_prop();
        let out = compile_and_emit_with(
            "SELECT Product { t := (select .tags filter @weight > 0.5 limit 1).id }",
            &schema,
        );
        assert!(out.sql.contains("\"t3\".\"weight\" > (0.5::float8)"), "{}", out.sql);
    }

    #[test]
    fn test_unknown_link_property_names_the_through_type() {
        let schema = make_schema_with_through_and_prop();
        let ast = parse::parse("SELECT Product { tags: { } filter @nope > 0.5 }").unwrap();
        let err = match ir::compile(&ast, &schema) {
            Ok(_) => panic!("expected a compile error"),
            Err(e) => format!("{e}"),
        };
        assert!(err.contains("default::ProductTag") && err.contains("nope"), "{err}");
    }

    #[test]
    fn test_link_property_outside_a_link_says_where_it_belongs() {
        let schema = make_schema_with_through_and_prop();
        let ast = parse::parse("SELECT Product { t := @weight }").unwrap();
        let err = match ir::compile(&ast, &schema) {
            Ok(_) => panic!("expected a compile error"),
            Err(e) => format!("{e}"),
        };
        assert!(err.contains("is a link property"), "{err}");
    }

    #[test]
    fn test_multilink_append_with_link_property() {
        let schema = make_schema_with_through_and_prop();
        let out = compile_and_emit_with(
            "UPDATE Product FILTER .id = $id SET { tags += (SELECT Tag FILTER .id = $tid) { @weight := <float64>$w } }",
            &schema,
        );
        // Extra junction column present in both the INSERT column list and the SELECT list.
        assert!(out.sql.contains("\"weight\""), "missing weight column:\n{}", out.sql);
        // Re-linking an existing pair with a new weight must update in place,
        // not silently keep the old value (a bare `DO NOTHING` would).
        assert!(
            out.sql
                .contains("ON CONFLICT (\"source\", \"target\") DO UPDATE SET \"weight\" = EXCLUDED.\"weight\""),
            "missing upsert conflict clause:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_multilink_append_union_with_different_link_property_values() {
        // The realistic Data Explorer scenario: multiple checked targets in
        // one `+=`, each with its own distinct property value — expressed as
        // a `union` of individually-shaped target selects.
        let schema = make_schema_with_through_and_prop();
        let out = compile_and_emit_with(
            "UPDATE Product FILTER .id = $id SET { \
                tags += (SELECT Tag FILTER .id = $aid) { @weight := <float64>$w1 } \
                    union (SELECT Tag FILTER .id = $bid) { @weight := <float64>$w2 } \
            }",
            &schema,
        );
        assert!(
            out.sql.contains("UNION ALL"),
            "expected a UNION ALL between the two shaped targets:\n{}",
            out.sql
        );
        // Both branches must select the same "weight" column (with their own
        // value) so the union has a consistent column list.
        assert_eq!(
            out.sql.matches("AS \"weight\"").count(),
            2,
            "each union branch must project its own weight:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_multilink_append_without_link_property_keeps_do_nothing() {
        // No `@prop := ...` anywhere — must keep the original DO NOTHING
        // behavior (no spurious upsert/extra columns for the common case).
        let schema = make_schema_with_through_and_prop();
        let out = compile_and_emit_with(
            "UPDATE Product FILTER .id = $id SET { tags += (SELECT Tag FILTER .id = $tid) }",
            &schema,
        );
        assert!(
            out.sql.contains("ON CONFLICT DO NOTHING"),
            "expected plain DO NOTHING when no link properties are set:\n{}",
            out.sql
        );
        assert!(
            !out.sql.contains("\"weight\""),
            "unexpected weight column with no link property assignment:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_multilink_link_property_rejected_on_standard_junction() {
        // Person.posts is a Standard (implicit) junction — no user-declared
        // properties, so `@prop := ...` must be rejected at compile time.
        let ast = crate::parse::parse(
            "UPDATE Person FILTER .id = $id SET { posts += (SELECT Post FILTER .title = $t) { @weight := <float64>$w } }",
        ).unwrap();
        match crate::ir::compile(&ast, &make_schema()) {
            Ok(_) => panic!("expected a compile error for link property on a Standard junction"),
            Err(e) => assert!(
                e.to_string().contains("Through"),
                "expected a Through[...]-related error, got: {e}"
            ),
        }
    }

    #[test]
    fn test_multilink_link_property_rejected_on_remove() {
        let schema = make_schema_with_through_and_prop();
        let ast = crate::parse::parse(
            "UPDATE Product FILTER .id = $id SET { tags -= (SELECT Tag FILTER .id = $tid) { @weight := <float64>$w } }",
        )
        .unwrap();
        match crate::ir::compile(&ast, &schema) {
            Ok(_) => panic!("expected a compile error for link property on a remove (-=)"),
            Err(e) => assert!(
                e.to_string().contains("removing"),
                "expected a remove-related error, got: {e}"
            ),
        }
    }

    #[test]
    fn test_insert_with_multilink_assignment() {
        // Regression test: `insert Product { tags := ... }` used to fail with
        // "object type 'default::Product' has no link or property 'tags'" —
        // compile_assignments_inner never checked resolve_multilink, and
        // IrInsert had no multi-link handling at all.
        let schema = make_schema_with_through_and_prop();
        let out = compile_and_emit_with(
            "INSERT Product { name := $name, tags := (SELECT Tag FILTER .id = $tid) { @weight := <float64>$w } }",
            &schema,
        );
        // The row insert must happen before the junction insert references its id.
        assert!(
            out.sql.contains("\"_w__ids\" AS (\nINSERT INTO"),
            "missing row-insert CTE:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("\"_w__ml_add_0\" AS ("),
            "missing junction-append CTE:\n{}",
            out.sql
        );
        assert!(out.sql.contains("\"weight\""), "missing weight column:\n{}", out.sql);
        assert!(
            out.sql.contains("\"_w\" AS (\n    SELECT * FROM \"_w__ids\"\n)"),
            "missing _w passthrough:\n{}",
            out.sql
        );
        assert_eq!(
            out.sql.matches("WITH\n").count(),
            1,
            "must be a single flat top-level WITH block:\n{}",
            out.sql
        );
    }

    /// A user `with` binding *consumed by* a statement that builds CTEs of
    /// its own. The binding is emitted by the generic `ir.ctes` path while
    /// the multi-link/junction machinery has already opened its own `WITH`;
    /// prefixing a second one produced `WITH ... WITH ...`, which compiles
    /// cleanly and then fails at execution with `syntax error at or near
    /// "WITH"`. Covers both link kinds, since each reaches the junction
    /// emitter by a different route.
    #[test]
    fn test_user_with_binding_merges_into_a_statements_own_ctes() {
        let schema = make_schema_with_through_and_prop();
        for query in [
            "with t := (select Tag filter .id = $tid) insert Product { name := $name, tags := t }",
            "with t := (select Tag filter .id = $tid) \
             update Product filter .name = $name set { tags += t }",
        ] {
            let out = compile_and_emit_with(query, &schema);
            let flat: String = out.sql.split_whitespace().collect::<Vec<_>>().join(" ");
            assert!(
                !flat.contains(") WITH "),
                "a second WITH clause is invalid SQL:\n{}",
                out.sql
            );
            assert_eq!(
                out.sql.matches("WITH").count(),
                1,
                "exactly one WITH keyword expected:\n{}",
                out.sql
            );
        }
    }

    #[test]
    fn test_with_bound_insert_with_multilink_assignment() {
        // Same as above, but bound via WITH (the Data Explorer's actual
        // shape) — exercises emit_user_cte_parts's insert_has_any_multilink
        // branch instead of emit_insert_stmt's bare-statement path.
        let schema = make_schema_with_through_and_prop();
        let out = compile_and_emit_with(
            "with insert0 := (insert Product { name := $name, tags := (select Tag filter .id = $tid) }) select insert0",
            &schema,
        );
        assert!(
            out.sql.contains("\"insert0__ids\" AS (\nINSERT INTO"),
            "missing row-insert CTE:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("\"insert0__ml_add_0\" AS ("),
            "missing junction-append CTE:\n{}",
            out.sql
        );
        assert!(
            out.sql
                .contains("\"insert0\" AS (\n    SELECT * FROM \"insert0__ids\"\n)"),
            "missing insert0 passthrough:\n{}",
            out.sql
        );
        assert_eq!(
            out.sql.matches("WITH\n").count(),
            1,
            "must be a single flat top-level WITH block:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_with_block_cte_over_computed_global_merges_into_single_with_clause() {
        // Regression: `with user := (select global current_user) select user;`
        // produced two separate top-level `WITH` keywords ("syntax error at
        // or near WITH" from Postgres) whenever the computed global's own
        // expression referenced a session global (registering a global CTE
        // in addition to the user-defined "user" CTE). The merge check in
        // emit() only recognized the "WITH " (space) prefix convention, not
        // emit_cte_prefix's "WITH\n" (newline) convention used here, so it
        // prepended a second WITH block instead of merging into the first.
        let mut schema = make_schema();
        schema.globals.push(GlobalDescriptor {
            name: "current_user_id".into(),
            module: "default".into(),
            scalar_type: "std::uuid".into(),
            required: false,
            default_expr: None,
            computed_expr: None,
        });
        schema.globals.push(GlobalDescriptor {
            name: "current_user".into(),
            module: "default".into(),
            scalar_type: "Person".into(),
            required: false,
            default_expr: None,
            computed_expr: Some("select default::Person filter .id = global current_user_id".into()),
        });
        let out = compile_and_emit_with("with\n  user := (select global current_user)\nselect user;", &schema);
        assert_eq!(
            out.sql.matches("WITH").count(),
            1,
            "must be a single WITH clause, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_several_mutations_in_a_free_set_each_become_a_cte() {
        let schema = make_schema();
        let out = compile_and_emit_with(
            "select { (update Person filter .name = 'a' set { age := 1 }), \
                      (update Company filter .name = 'b' set { name := 'c' }) }",
            &schema,
        );
        assert_eq!(
            out.sql.matches("UPDATE").count(),
            2,
            "both mutations have to run, got:\n{}",
            out.sql
        );
        assert!(out.sql.contains("WITH"), "each becomes a CTE, got:\n{}", out.sql);
    }

    #[test]
    fn test_updating_a_binding_touches_only_its_own_rows() {
        // The binding names the rows to update, so resolving it to its type
        // without narrowing would rewrite every row in the table.
        let schema = make_schema();
        let out = compile_and_emit_with(
            "with mine := (select Person filter .name = 'a') select { (update mine set { age := 1 }) }",
            &schema,
        );
        assert!(
            out.sql.contains("UPDATE") && out.sql.contains("= ANY(ARRAY(SELECT"),
            "expected the update narrowed to the binding's rows, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_updating_a_traversal_touches_only_the_rows_it_lands_on() {
        let mut schema = make_schema();
        // The traversal's target needs an id to narrow against; every real
        // schema declares one, this one leaves it off Company.
        let company = schema
            .types
            .iter_mut()
            .find(|t| t.name == "Company")
            .expect("test schema has a Company type");
        company.properties.push(crate::schema::PropertyDescriptor {
            name: "id".into(),
            pg_type: "uuid".into(),
            nullable: false,
            default_sql: Some("uuidv7()".into()),
            default_pyql: None,
            description: None,
            check_constraints: vec![],
            is_exclusive: true,
            is_pk: true,
            is_readonly: true,
            rewrites: vec![],
            tuple_members: None,
            column_type: None,
        });
        let out = compile_and_emit_with(
            "with mine := (select Person filter .name = 'a') select { (update mine.company set { name := 'x' }) }",
            &schema,
        );
        assert!(
            out.sql.contains("= ANY(ARRAY(SELECT"),
            "expected the update narrowed to the traversal's rows, got:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("\"Company\""),
            "expected the table the traversal ends on, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_aggregate_over_a_multi_valued_path_keeps_the_set_flat() {
        // Regression: `array_agg(a.posts.title)` compiled the path as an
        // expression, where a multi-valued path stands for the array of its
        // elements — so the aggregate wrapped that array and the result came
        // back one level deep, `[[t1, t2]]` where Gel gives `[t1, t2]`.
        let schema = make_schema();
        let out = compile_and_emit_with(
            "with a := (select Person limit 1) select { titles := array_agg(a.posts.title) }",
            &schema,
        );
        assert!(
            !out.sql.contains("array_agg(ARRAY("),
            "the aggregate must take the set, not the array standing for it, got:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("array_agg(") && out.sql.contains("\"title\""),
            "expected the aggregate over the traversal's own column, got:\n{}",
            out.sql
        );
        // The free object keeps its composite emission, so values stay typed
        // rather than passing through jsonb as strings.
        assert!(
            out.sql.contains("ROW(") && !out.sql.contains("jsonb_build_object"),
            "expected a composite row, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_with_bound_root_read_by_name_inside_a_nested_select() {
        // Regression: `compile_path_select` resolved a `with`-bound root
        // correctly, but the multi-valued check straight after it re-resolved
        // the same root as a plain type, so reading the binding by name from a
        // nested sub-select failed with "unknown type 'owner'".
        let schema = make_schema();
        let out = compile_and_emit_with(
            "with
  owner := (select Person limit 1)
select owner { posts := (select owner.posts.title) };",
            &schema,
        );
        assert!(
            out.sql.contains("WITH") && out.sql.contains("\"title\""),
            "expected the nested select to traverse the binding, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_path_traversal_into_with_bound_cte_of_object_type() {
        // Regression: `with user := (select global current_user) select
        // user.name;` failed with "unknown type 'user'" — compile_path_select
        // (used for multi-step absolute paths like `user.name`) only ever
        // tried resolve_type(root_name), never checking whether the root
        // name is a WITH-block CTE bound to an object type.
        let mut schema = make_schema();
        schema.globals.push(GlobalDescriptor {
            name: "current_user_id".into(),
            module: "default".into(),
            scalar_type: "std::uuid".into(),
            required: false,
            default_expr: None,
            computed_expr: None,
        });
        schema.globals.push(GlobalDescriptor {
            name: "current_user".into(),
            module: "default".into(),
            scalar_type: "Person".into(),
            required: false,
            default_expr: None,
            computed_expr: Some("select default::Person filter .id = global current_user_id".into()),
        });
        let out = compile_and_emit_with(
            "with\n  user := (select global current_user)\nselect user.name;",
            &schema,
        );
        assert!(
            out.sql.contains("FROM \"user\""),
            "expected path traversal from the CTE, got:\n{}",
            out.sql
        );
        assert_eq!(
            out.sql.matches("WITH").count(),
            1,
            "must be a single WITH clause, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_with_bound_schema_object_property_access_in_filter_expression() {
        // Regression: `with person := (select detached Person filter ...)
        // select Person filter .id not in person.id;` failed with "absolute
        // paths are not valid in expression context" — compile_path's
        // schema-bound-CTE-in-expression-position case only ever handled a
        // WITH-bound *free object* (resolve_cte_path), never a WITH-bound
        // *schema object* CTE, even though its own SELECT * CTE source
        // (emit_dml_as_cte_source) always exposes every raw column.
        let schema = make_schema();
        let out = compile_and_emit_with(
            "with\n  person := (select detached Person filter .id = <uuid>$id)\n\
             select Person filter .id not in person.id;",
            &schema,
        );
        assert!(
            out.sql.contains("<> ALL((SELECT") && out.sql.contains(".\"id\"") && out.sql.contains("FROM \"person\""),
            "got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_with_bound_schema_object_scalar_property_access() {
        // Not just `.id` — any scalar property of the bound type is
        // reachable, since the CTE source is `SELECT *` regardless of the
        // binding's own declared shape.
        let schema = make_schema();
        let out = compile_and_emit_with(
            "with\n  person := (select detached Person filter .id = <uuid>$id)\n\
             select Person filter .name = person.name;",
            &schema,
        );
        assert!(
            out.sql.contains("(SELECT") && out.sql.contains(".\"name\"") && out.sql.contains("FROM \"person\""),
            "got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_with_bound_schema_object_link_traversal_reaches_nested_property() {
        // Generalization: a with-bound schema object's CTE source is
        // `SELECT *`, so it's traversable exactly like a real type name —
        // not just its own scalar properties, but a *link's* properties too
        // (`person.company.name`), via the same path-traversal machinery
        // `TypeName.a.b.c` already uses (compile_path_select), wrapped as a
        // correlated subquery (IrExpr::PathSubquery).
        let schema = make_schema();
        let out = compile_and_emit_with(
            "with\n  person := (select detached Person filter .id = <uuid>$id)\n\
             select Company filter .name = person.company.name;",
            &schema,
        );
        assert!(
            out.sql.contains("(SELECT") && out.sql.contains("\"name\"") && out.sql.contains("\"Company\""),
            "got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_with_bound_schema_object_bare_link_reduces_to_id() {
        // A bare link at the end of the chain (`person.company`, no further
        // traversal) reduces to the linked object's id, matching the
        // existing bare-CTE-reference convention (IrExpr::CteRef).
        let schema = make_schema();
        let out = compile_and_emit_with(
            "with\n  person := (select detached Person filter .id = <uuid>$id)\n\
             select Person filter .id = person.company;",
            &schema,
        );
        assert!(
            out.sql.contains("(SELECT") && out.sql.contains("\"id\""),
            "got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_with_bound_schema_object_unknown_property_suggests_a_close_match() {
        let schema = make_schema();
        let ast = parse::parse(
            "with\n  person := (select detached Person filter .id = <uuid>$id)\n\
             select Person filter .name = person.nam;",
        )
        .unwrap();
        match ir::compile(&ast, &schema) {
            Err(err) => assert!(format!("{err}").contains("Did you mean 'name'"), "got: {err}"),
            Ok(_) => panic!("expected a compile error"),
        }
    }

    #[test]
    fn test_with_bound_free_object_passthrough_preserves_all_fields() {
        // Regression: `with test := { test2 := 1.0, test3 := 'str' } select
        // test;` decoded as just `1.0` (the CTE's first field) — a bare
        // reference to a free-object CTE (`IrFreeExpr::CtePassthrough`) had
        // its shape hardcoded to `ShapeNode::Scalar` regardless of what the
        // CTE actually held.
        let out = compile_and_emit("with\n  test := { test2 := 1.0, test3 := 'str' }\nselect test;");
        let ShapeNode::Object { pointers, .. } = &out.shape.root else {
            panic!("expected Object shape, got {:?}", out.shape.root)
        };
        assert_eq!(pointers.len(), 2);
        assert!(matches!(&pointers[0], ShapeNode::Scalar { name, .. } if name == "test2"));
        assert!(matches!(&pointers[1], ShapeNode::Scalar { name, .. } if name == "test3"));
    }

    #[test]
    fn test_with_bound_free_object_field_access() {
        // Regression: `with test := {...} select test.test2;` failed with
        // "unknown type 'test'" — any absolute path with 2+ steps
        // unconditionally routed to compile_path_select, which only knows
        // how to resolve real schema types, never a free-value CTE.
        let out = compile_and_emit("with\n  test := { test2 := 1.0, test3 := 'str' }\nselect test.test2;");
        assert!(out.sql.contains("\"test2\" FROM \"test\""), "got:\n{}", out.sql);
    }

    #[test]
    fn test_with_bound_free_object_nested_field_access_chain() {
        // A field-access chain through a *nested* free object literal
        // (`test.test3.foo`) must resolve the first hop via the CTE's own
        // materialized column, then extract `foo` as jsonb from that value
        // — the fix must not be hardcoded to exactly 2 path steps.
        let out =
            compile_and_emit("with\n  test := { test2 := 1.0, test3 := { foo := 'bar' } }\nselect test.test3.foo;");
        assert!(out.sql.contains("\"test3\" FROM \"test\""), "got:\n{}", out.sql);
        assert!(
            out.sql.contains("->'foo'"),
            "expected jsonb field extraction, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_with_bound_free_object_nested_field_access_wrong_field_errors() {
        // A typo'd field name anywhere in the chain must still be a compile
        // error, not silently emit SQL that returns NULL at runtime.
        let schema = make_schema();
        let ast = parse::parse("with\n  test := { test2 := 1.0, test3 := { foo := 'bar' } }\nselect test.test3.nope;")
            .unwrap();
        assert!(ir::compile(&ast, &schema).is_err());
    }

    #[test]
    fn test_nested_free_object_literal_in_computed_shape_element() {
        // Regression: a free object literal (`{ foo := 'bar' }`) nested
        // inside a computed shape element — not the top-level SELECT
        // result — hit "shapes and set literals are not valid in
        // expression context"; free object literals were only ever handled
        // at the statement level.
        let out = compile_and_emit("select default::Person { id, test := { foo := 'bar' } };");
        let ShapeNode::Object { pointers, .. } = &out.shape.root else {
            panic!("expected Object shape")
        };
        let test_node = pointers
            .iter()
            .find(|p| matches!(p, ShapeNode::NamedTuple { name, .. } if name == "test"))
            .unwrap_or_else(|| panic!("expected a NamedTuple shape node for 'test', got {:?}", pointers));
        assert!(matches!(
            test_node,
            ShapeNode::NamedTuple {
                is_free_object: true,
                ..
            }
        ));
    }

    #[test]
    fn test_bare_free_cte_reference_in_computed_shape_collapses_to_empty() {
        // A free-object CTE referenced bare (no shape) as a computed shape
        // element's value has nothing to project — a free object needs an
        // explicit shape to know what to expose (unlike a tuple, which is
        // a plain value with no such requirement).
        let out = compile_and_emit(
            "with\n  test := { test2 := 1.0, test3 := 'str' }\n\
             select default::Person { id, test := test };",
        );
        assert!(
            out.sql.contains("jsonb_build_object()"),
            "expected an empty free object, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_shaped_free_cte_reference_projects_fields() {
        // `test := test { test2 }` must project just the named field out
        // of the underlying free object — neither collapsing to empty nor
        // returning every field.
        let out = compile_and_emit(
            "with\n  test := { test2 := 1.0, test3 := 'str' }\n\
             select default::Person { id, test := test { test2 } };",
        );
        assert!(out.sql.contains("jsonb_build_object('test2'"), "got:\n{}", out.sql);
        assert!(
            !out.sql.contains("'test3'"),
            "test3 should not be projected, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_insert_multilink_remove_rejected() {
        let schema = make_schema_with_through_and_prop();
        let ast =
            crate::parse::parse("INSERT Product { name := $name, tags -= (SELECT Tag FILTER .id = $tid) }").unwrap();
        match crate::ir::compile(&ast, &schema) {
            Ok(_) => panic!("expected a compile error for `-=` on a multi-link at insert time"),
            Err(e) => assert!(
                e.to_string().contains("nothing to remove"),
                "expected a 'nothing to remove yet' error, got: {e}"
            ),
        }
    }

    #[test]
    fn test_multilink_junction_info_disambiguates_self_referencing_through_type() {
        // Regression test for a self-referencing through-link (Person.friends
        // via PersonFriend, which declares two Person-typed links: "person"
        // and "friend"). Naively matching purely by target type resolves
        // *both* source_col and target_col to the same first-matching link
        // ("person"), silently collapsing the junction to a single column
        // and losing the other side entirely.
        let schema = make_schema_with_through();
        let out = compile_and_emit_with(
            "UPDATE Person FILTER .id = $id SET { friends += (SELECT Person FILTER .id = $fid) }",
            &schema,
        );
        assert!(
            out.sql.contains("(\"person\", \"friend\")"),
            "expected two distinct FK columns:\n{}",
            out.sql
        );
        assert!(
            !out.sql.contains("(\"person\", \"person\")"),
            "source/target collapsed to the same column:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_shape_descriptor_scalars() {
        let out = compile_and_emit("SELECT Person { name, age }");
        let ShapeNode::Object { pointers, .. } = &out.shape.root else {
            panic!()
        };
        assert_eq!(pointers.len(), 4); // __type__, the implicit id, name, age
        assert!(matches!(&pointers[0], ShapeNode::Scalar { name, position: 0 } if name == "__type__"));
        assert!(matches!(&pointers[1], ShapeNode::Scalar { name, position: 1 } if name == "id"));
        assert!(matches!(&pointers[2], ShapeNode::Scalar { name, position: 2 } if name == "name"));
        assert!(matches!(&pointers[3], ShapeNode::Scalar { name, position: 3 } if name == "age"));
    }

    #[test]
    fn test_shape_descriptor_multi_link() {
        let out = compile_and_emit("SELECT Person { name, posts { title } }");
        let ShapeNode::Object { pointers, .. } = &out.shape.root else {
            panic!()
        };
        // pointers: [__type__, id, name, posts]
        assert_eq!(pointers.len(), 4);
        let ShapeNode::Array {
            name,
            position,
            element,
        } = &pointers[3]
        else {
            panic!()
        };
        assert_eq!(name, "posts");
        assert_eq!(*position, 3);
        let ShapeNode::Object {
            pointers: elem_pointers,
            ..
        } = element.as_ref()
        else {
            panic!()
        };
        // element pointers: [__type__, title] — `Post` has no pk property in
        // this fixture, so nothing is injected into its shape.
        assert_eq!(elem_pointers.len(), 2);
    }

    #[test]
    fn test_select_order_by_limit() {
        let out = compile_and_emit("SELECT Person { name } ORDER BY .name ASC LIMIT 10");
        assert!(out.sql.contains("ORDER BY"));
        assert!(out.sql.contains("LIMIT 10"));
    }

    #[test]
    fn test_insert_returning() {
        let out = compile_and_emit("INSERT Person { name := 'Alice', age := 30 }");
        assert!(out.sql.contains("INSERT INTO \"public\".\"Person\""));
        assert!(out.sql.contains("RETURNING"));
        assert!(out.sql.contains("'default::Person'::text"));
        assert!(out.sql.contains(") AS result"));
        // Bare INSERT returns pk only
        let ShapeNode::Object {
            cardinality, pointers, ..
        } = &out.shape.root
        else {
            panic!()
        };
        assert_eq!(*cardinality, Cardinality::Required);
        // Only __type__ and id — not name or age
        assert!(
            pointers
                .iter()
                .any(|f| matches!(f, ShapeNode::Scalar { name, .. } if name == "id"))
        );
        assert!(
            !pointers
                .iter()
                .any(|f| matches!(f, ShapeNode::Scalar { name, .. } if name == "name"))
        );
    }

    #[test]
    fn test_update_returning() {
        let out = compile_and_emit("UPDATE Person FILTER .name = $name SET { age := 31 }");
        assert!(out.sql.contains("UPDATE \"public\".\"Person\""));
        assert!(out.sql.contains("SET"));
        // Bare UPDATE returns pk only
        assert!(out.sql.contains("RETURNING"));
        assert!(out.sql.contains("'default::Person'::text"));
        assert!(!out.sql.contains("\"name\"::text"), "bare UPDATE must not return name");
    }

    #[test]
    fn test_update_set_tuple_param_cast_uses_direct_jsonb_cast_not_to_jsonb() {
        // Regression: to_jsonb($N) on a bare, still-untyped parameter fails at
        // execution time with "could not determine polymorphic type because
        // input has type unknown" — Postgres can't dispatch a polymorphic
        // function against an unknown-typed placeholder. A direct
        // ($N)::jsonb cast resolves the parameter's type from the cast
        // itself instead, matching how e.g. `$1::uuid` already works.
        let out = compile_and_emit("UPDATE Person FILTER .id = $id SET { age := <tuple<x: float64>>$val }");
        assert!(
            out.sql.contains(")::jsonb"),
            "expected a direct ::jsonb cast, got:\n{}",
            out.sql
        );
        assert!(
            !out.sql.contains("to_jsonb($"),
            "must not pass a bare param straight into to_jsonb(): got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_empty_set_cast_to_object_type_clears_optional_link() {
        // Regression: unsetting an optional single-link generates
        // `<TargetType>{}` (a cast of the empty set, not the bare `{}` the
        // assignment-position special case in compile_assignments_inner
        // already handled) — that TypeCast wraps the empty set, so it fell
        // through to the generic Set/Shape rejection instead ("shapes and
        // set literals are not valid in expression context"). Must compile
        // to a plain NULL for the link's FK column: `<AnyType>{}` is always
        // NULL regardless of context.
        let out = compile_and_emit("UPDATE Person FILTER .id = $id SET { company := <default::Company>{} }");
        assert!(out.sql.contains("\"company_id\" = NULL"), "got:\n{}", out.sql);
    }

    #[test]
    fn test_delete_returning() {
        let out = compile_and_emit("DELETE Person FILTER .id = $id");
        assert!(out.sql.contains("DELETE FROM \"public\".\"Person\""));
        // Bare DELETE returns pk only
        assert!(out.sql.contains("RETURNING"));
        assert!(out.sql.contains("'default::Person'::text"));
        assert!(!out.sql.contains("\"name\"::text"), "bare DELETE must not return name");
    }

    #[test]
    fn test_select_over_insert() {
        let out = compile_and_emit("SELECT (INSERT Person { name := $name, age := $age }) { id, name }");
        // Must use a CTE
        assert!(out.sql.contains("WITH\n\"_dml\" AS ("));
        assert!(out.sql.contains("INSERT INTO"));
        assert!(out.sql.contains("RETURNING *"));
        // Outer SELECT shapes the result
        assert!(out.sql.contains("'default::Person'::text"));
        assert!(out.sql.contains("\"name\"::text"));
    }

    #[test]
    fn test_select_over_update() {
        let out = compile_and_emit("SELECT (UPDATE Person FILTER .id = $id SET { name := $name }) { id, name }");
        assert!(out.sql.contains("WITH\n\"_dml\" AS ("));
        assert!(out.sql.contains("UPDATE"));
        assert!(out.sql.contains("RETURNING \"t1\".*"), "{}", out.sql);
        assert!(out.sql.contains("\"name\"::text"));
    }

    #[test]
    fn test_insert_user_specified_id_denied_by_default() {
        let schema = make_schema();
        let ast = parse::parse("INSERT Person { id := <uuid>$id, name := $name, age := $age }").unwrap();
        match ir::compile_with_config(&ast, &schema, &ir::SessionConfig::default()) {
            Err(err) => assert!(err.to_string().contains("cannot assign to property 'id'"), "got: {err}"),
            Ok(_) => panic!("expected id assignment to be denied by default"),
        }
    }

    #[test]
    fn test_insert_user_specified_id_allowed_when_configured() {
        let schema = make_schema();
        let ast = parse::parse("INSERT Person { id := <uuid>$id, name := $name, age := $age }").unwrap();
        let config = ir::SessionConfig {
            allow_user_specified_id: true,
        };
        let ir_out = ir::compile_with_config(&ast, &schema, &config)
            .expect("expected id assignment to be allowed with allow_user_specified_id");
        let out = emit(&ir_out);
        assert!(out.sql.contains("INSERT INTO"));
    }

    #[test]
    fn test_update_user_specified_id_denied_even_when_configured() {
        let schema = make_schema();
        let ast = parse::parse("UPDATE Person FILTER .name = $name SET { id := <uuid>$id }").unwrap();
        let config = ir::SessionConfig {
            allow_user_specified_id: true,
        };
        match ir::compile_with_config(&ast, &schema, &config) {
            Err(err) => assert!(err.to_string().contains("cannot assign to property 'id'"), "got: {err}"),
            Ok(_) => panic!("expected UPDATE to always deny reassigning id"),
        }
    }

    #[test]
    fn test_select_over_update_multilink_only() {
        // Regression test: an UPDATE bound to a single external name (here,
        // the implicit "_dml" wrapper for `SELECT (UPDATE ...)`) whose SET
        // clause is *only* a multi-link mutation, with no scalar/single-link
        // assignment. This used to either emit an empty `SET` clause
        // ("UPDATE ... SET  WHERE ..." — invalid SQL) or, in an earlier fix
        // attempt, nest the junction INSERT inside "_dml"'s own CTE body —
        // which Postgres rejects outright ("WITH clause containing a
        // data-modifying statement must be at the top level"). The junction
        // CTEs must instead be hoisted out as top-level siblings of "_dml".
        let out = compile_and_emit(
            "SELECT (UPDATE Person FILTER .id = $id SET { posts += (SELECT Post FILTER .title = $title) }) { id, name }",
        );
        assert!(
            out.sql.contains("\"_dml__ml_add_0\""),
            "missing junction-append CTE:\n{}",
            out.sql
        );
        assert!(out.sql.contains("INSERT INTO"), "missing junction INSERT:\n{}", out.sql);
        // No scalar changes -> the row-source CTE must be a SELECT, not an UPDATE
        // with an empty SET clause.
        assert!(
            out.sql.contains("\"_dml__ids\" AS (\nSELECT"),
            "expected SELECT-based _ids CTE:\n{}",
            out.sql
        );
        assert!(
            !out.sql.contains("SET\n\nWHERE") && !out.sql.contains("SET \nWHERE"),
            "empty SET clause regression:\n{}",
            out.sql
        );
        // The junction CTE must be a *sibling* at the top-level WITH, not
        // nested inside another CTE's body — only one "WITH" keyword total.
        assert_eq!(
            out.sql.matches("WITH\n").count(),
            1,
            "junction CTE must not be nested in a second WITH:\n{}",
            out.sql
        );
        // "_dml" itself must still resolve (as a passthrough) for the outer SELECT.
        assert!(
            out.sql.contains("\"_dml\" AS (\n    SELECT * FROM \"_dml__ids\"\n)"),
            "missing _dml passthrough:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_select_over_update_scalar_and_multilink() {
        // Mixed case: a scalar assignment alongside a multi-link mutation, both
        // bound to the same external name — the scalar SET must survive *and*
        // the junction mutation must still be emitted as its own top-level CTE.
        let out = compile_and_emit(
            "SELECT (UPDATE Person FILTER .id = $id SET { name := $name, posts += (SELECT Post FILTER .title = $title) }) { id, name }",
        );
        assert!(
            out.sql.contains("\"_dml__ml_add_0\""),
            "missing junction-append CTE:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("\"_dml__ids\" AS (\nUPDATE"),
            "expected UPDATE-based _ids CTE:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("\"name\" = "),
            "missing scalar SET assignment:\n{}",
            out.sql
        );
        assert_eq!(
            out.sql.matches("WITH\n").count(),
            1,
            "junction CTE must not be nested in a second WITH:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_with_bound_insert_and_multilink_update_forward_ref() {
        // The actual shape pylon-ui's Data Explorer generates for its
        // flagship "insert + link in one batch" scenario: a same-batch
        // forward-reference (Person.posts += a not-yet-existing Post, bound
        // to insert0) inside a multi-statement WITH. This is the exact query
        // that failed live against a real Postgres before this fix ("WITH
        // clause containing a data-modifying statement must be at the top
        // level"), so it's the most important regression to pin down.
        let out = compile_and_emit(
            "with insert0 := (insert Post { title := $title }), update0 := (update Person filter .id = $id set { posts += (select insert0) }) select { insert0, update0 }",
        );
        assert!(
            out.sql.contains("\"insert0\" AS (\n    INSERT INTO"),
            "missing insert0 CTE:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("\"update0__ml_add_0\""),
            "missing junction-append CTE for update0:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("\"update0__ids\" AS (\nSELECT"),
            "expected SELECT-based update0 ids CTE (no scalar changes):\n{}",
            out.sql
        );
        assert!(
            out.sql
                .contains("\"update0\" AS (\n    SELECT * FROM \"update0__ids\"\n)"),
            "missing update0 passthrough:\n{}",
            out.sql
        );
        // Exactly one WITH keyword — every CTE (insert0, update0__ids,
        // update0__ml_add_0, update0) must be a top-level sibling.
        assert_eq!(
            out.sql.matches("WITH\n").count(),
            1,
            "must be a single flat top-level WITH block:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_select_over_delete() {
        let out = compile_and_emit("SELECT (DELETE Person FILTER .id = $id) { id, name }");
        assert!(out.sql.contains("WITH\n\"_dml\" AS ("));
        assert!(out.sql.contains("DELETE FROM"));
        assert!(out.sql.contains("RETURNING *"));
        assert!(out.sql.contains("\"name\"::text"));
    }

    fn make_schema_with_rewrite() -> SchemaDescriptor {
        use crate::schema::RewriteEntry;
        let mut schema = make_schema();
        // Add a `slug` property to Person with an INSERT+UPDATE rewrite: lower(.name)
        let person = schema.types.iter_mut().find(|t| t.name == "Person").unwrap();
        person.properties.push(PropertyDescriptor {
            name: "slug".into(),
            pg_type: "text".into(),
            nullable: true,
            default_sql: None,
            default_pyql: None,
            description: None,
            check_constraints: vec![],
            is_exclusive: false,
            is_pk: false,
            is_readonly: false,
            rewrites: vec![
                RewriteEntry {
                    on: 1,
                    handler: "str_lower(.name)".into(),
                }, // INSERT
                RewriteEntry {
                    on: 2,
                    handler: "str_lower(.name)".into(),
                }, // UPDATE
            ],
            tuple_members: None,
            column_type: None,
        });
        schema
    }

    #[test]
    fn a_link_assigned_from_a_select_keeps_its_clauses() {
        let out = compile_and_emit(
            "INSERT Person { name := 'a', age := 1, company := (SELECT Company ORDER BY .name LIMIT 1) }",
        );
        assert!(
            out.sql.contains("ORDER BY") && out.sql.contains("LIMIT 1"),
            "{}",
            out.sql
        );
    }

    #[test]
    fn a_rewrite_leaves_the_statement_alone() {
        // It runs in the table's own `BEFORE` trigger, on the row as written.
        let schema = make_schema_with_rewrite();
        for query in [
            "INSERT Person { name := $name, age := 30 }",
            "UPDATE Person FILTER .id = $id SET { name := $name }",
        ] {
            let out = compile_and_emit_with(query, &schema);
            assert!(!out.sql.contains("\"slug\""), "{}", out.sql);
        }
    }

    #[test]
    fn a_rewrite_reads_the_row_being_written() {
        let schema = make_schema_with_rewrite();
        let assignments = crate::ir::compile_rewrite_assignments("default::Person", 1, &schema).unwrap();
        let [assignment] = assignments.as_slice() else {
            panic!("one rewrite on insert")
        };
        assert_eq!(assignment.column, "slug");
        assert!(assignment.sql.contains("NEW.\"name\""), "{}", assignment.sql);
    }

    #[test]
    fn rewrites_become_before_triggers() {
        let ddl = crate::export::export_schema(&make_schema_with_rewrite()).unwrap();
        for event in ["BEFORE INSERT", "BEFORE UPDATE"] {
            assert!(
                ddl.contains(&format!("{event} ON \"public\".\"Person\"")),
                "{event}:\n{ddl}"
            );
        }
        assert!(ddl.contains("NEW.\"slug\" := _pylon_rewrites.\"v0\";"), "{ddl}");
    }

    #[test]
    fn a_rewrite_walking_a_link_starts_from_the_new_row() {
        // The row is not in its table yet when a `BEFORE INSERT` trigger runs.
        use crate::schema::RewriteEntry;
        let mut schema = make_schema_with_rewrite();
        let person = schema.types.iter_mut().find(|t| t.name == "Person").unwrap();
        let slug = person.properties.iter_mut().find(|p| p.name == "slug").unwrap();
        slug.rewrites = vec![RewriteEntry {
            on: 1,
            handler: "<str>(.company is Company) ++ .company.name".into(),
        }];
        let assignments = crate::ir::compile_rewrite_assignments("default::Person", 1, &schema).unwrap();
        let sql = &assignments[0].sql;
        assert!(!sql.contains("\"public\".\"Person\""), "{sql}");
        assert!(
            sql.contains("(SELECT (NEW).*)") && sql.contains("NEW.\"company_id\""),
            "{sql}"
        );
    }

    #[test]
    fn test_unless_conflict_do_nothing() {
        let out = compile_and_emit("INSERT Person { name := $name } UNLESS CONFLICT");
        assert!(out.sql.contains("ON CONFLICT DO NOTHING"));
    }

    #[test]
    fn test_unless_conflict_on_do_nothing() {
        let out = compile_and_emit("INSERT Person { name := $name } UNLESS CONFLICT ON .name");
        assert!(out.sql.contains("ON CONFLICT (\"name\") DO NOTHING"));
    }

    #[test]
    fn test_unless_conflict_do_update() {
        let out = compile_and_emit(
            "INSERT Person { name := $name, age := $age } \
             UNLESS CONFLICT ON .name \
             ELSE (UPDATE Person SET { age := $age })",
        );
        assert!(out.sql.contains("ON CONFLICT (\"name\") DO UPDATE SET"));
        assert!(out.sql.contains("\"age\" = $2"));
        // The param $age is shared — same index as in the INSERT VALUES
        assert!(!out.sql.contains("DO NOTHING"));
    }

    #[test]
    fn test_unless_conflict_else_update_self_reference_is_qualified() {
        // Regression: a self-referencing ELSE-update assignment (`age :=
        // .age + 1`) used to compile its RHS with an empty (unqualified)
        // alias, emitting a bare `"age"` — ambiguous in Postgres's `ON
        // CONFLICT DO UPDATE SET` context between the existing row and the
        // `excluded` pseudo-row (confirmed live: "column reference ... is
        // ambiguous"). It must be qualified with the target table's own
        // bare name instead.
        let out = compile_and_emit(
            "INSERT Person { name := $name, age := 0 } \
             UNLESS CONFLICT ON .name \
             ELSE (UPDATE Person SET { age := .age + 1 })",
        );
        assert!(
            out.sql.contains("\"Person\".\"age\""),
            "self-reference must be qualified with the table's own name, got:\n{}",
            out.sql
        );
        assert!(
            !out.sql.contains("SET \"age\" = (\"age\""),
            "must not emit an unqualified (ambiguous) self-reference, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_unless_conflict_do_update_no_on() {
        let out = compile_and_emit(
            "INSERT Person { name := $name } \
             UNLESS CONFLICT \
             ELSE (UPDATE Person SET { age := 0 })",
        );
        assert!(out.sql.contains("ON CONFLICT DO UPDATE SET"));
        assert!(out.sql.contains("\"age\" = 0"));
    }

    #[test]
    fn test_select_over_select() {
        let out = compile_and_emit("SELECT (SELECT Person FILTER .age > 18) { name }");
        // Must use a CTE
        assert!(out.sql.contains("WITH\n\"_dml\" AS ("));
        // CTE exposes raw columns via SELECT *
        assert!(out.sql.contains("SELECT *"));
        assert!(out.sql.contains("FROM \"public\".\"Person\""));
        // CTE carries the inner filter
        assert!(out.sql.contains("WHERE"));
        // Outer SELECT projects its own shape
        assert!(out.sql.contains("'default::Person'::text"));
        assert!(out.sql.contains("\"name\"::text"));
    }

    #[test]
    fn test_select_over_select_with_outer_filter() {
        let out = compile_and_emit("SELECT (SELECT Person FILTER .age > 18) { name } FILTER .name = $name");
        assert!(out.sql.contains("WITH\n\"_dml\" AS ("));
        assert!(out.sql.contains("SELECT *"));
        // Both filters present: one inside CTE, one in outer SELECT
        assert_eq!(out.sql.matches("WHERE").count(), 2);
        assert!(out.sql.contains("$1"));
    }

    #[test]
    fn test_insert_link_subquery() {
        let out = compile_and_emit("INSERT Person { name := $name, company := (SELECT Company FILTER .name = $co) }");
        // The company FK column should be assigned via a scalar subquery
        assert!(out.sql.contains("\"company_id\""));
        assert!(out.sql.contains("SELECT"));
        // The subquery must select the pk (id) of Company
        assert!(out.sql.contains("\"id\""));
        assert!(out.sql.contains("FROM \"public\".\"Company\""));
        // Filter param must appear
        assert!(out.sql.contains("$2")); // $1 = name, $2 = co
    }

    #[test]
    fn test_update_link_subquery() {
        let out =
            compile_and_emit("UPDATE Person FILTER .id = $id SET { company := (SELECT Company FILTER .name = $co) }");
        assert!(out.sql.contains("\"company_id\""));
        assert!(out.sql.contains("SELECT"));
        assert!(out.sql.contains("FROM \"public\".\"Company\""));
    }

    #[test]
    fn test_insert_link_value_from_nested_insert_hoists_a_with_cte() {
        // `company := (select (insert Company {...}) { id })` — Postgres has
        // no way to run a nested INSERT inside another statement's VALUES
        // list without hoisting it into its own WITH CTE first (confirmed
        // live against real Postgres — see live_execution_insert_nested.rs).
        // Regression: an earlier attempt at this compiled cleanly but
        // silently emitted a subquery that dropped the nested INSERT
        // entirely and selected an arbitrary unrelated row instead.
        let out = compile_and_emit(
            "INSERT Person { name := 'Alice', company := (select (insert Company { name := 'Acme' }) { id }) }",
        );
        assert!(
            out.sql.starts_with("WITH"),
            "expected a WITH-hoisted CTE, got:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("INSERT INTO \"public\".\"Company\""),
            "expected the nested insert to be its own CTE, got:\n{}",
            out.sql,
        );
        // The outer insert must reference the nested CTE's own id, not a
        // freestanding subquery against the Company table.
        assert!(
            out.sql.contains("\"company_id\") SELECT") && out.sql.contains(".\"id\" FROM"),
            "expected the outer insert to switch from VALUES to SELECT ... FROM <cte>, got:\n{}",
            out.sql,
        );
        assert!(
            !out.sql.contains("FROM \"public\".\"Company\" AS"),
            "must not select from the real Company table, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_update_link_value_from_nested_insert_hoists_a_with_cte_and_adds_from_clause() {
        let out = compile_and_emit(
            "UPDATE Person FILTER .id = $id SET { company := (select (insert Company { name := 'Acme' }) { id }) }",
        );
        assert!(
            out.sql.starts_with("WITH"),
            "expected a WITH-hoisted CTE, got:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("INSERT INTO \"public\".\"Company\""),
            "expected the nested insert to be its own CTE, got:\n{}",
            out.sql,
        );
        // Plain `UPDATE ... SET ...` has no FROM clause of its own — one
        // must be added so the SET clause can reference the CTE's column.
        assert!(
            out.sql.contains("\nFROM \""),
            "expected a FROM clause referencing the nested CTE, got:\n{}",
            out.sql
        );
        assert!(out.sql.contains("SET \"company_id\" = "), "got:\n{}", out.sql);
    }

    #[test]
    fn test_update_link_value_from_nested_insert_combines_with_multilink_mutation() {
        // A nested-insert link value and a `+=` multi-link mutation in the
        // same UPDATE route through a completely different emitter branch
        // (the junction-CTE "_ids" path, not the plain-UPDATE path the
        // previous test covers) — both must still see the hoisted CTE.
        let out = compile_and_emit(
            "UPDATE Person FILTER .id = $id SET { \
                 company := (select (insert Company { name := 'Acme' }) { id }), \
                 posts += (SELECT Post FILTER .title = $t) \
             }",
        );
        assert!(out.sql.starts_with("WITH"), "got:\n{}", out.sql);
        assert!(
            out.sql.contains("INSERT INTO \"public\".\"Company\""),
            "expected the nested insert to be its own CTE, got:\n{}",
            out.sql,
        );
        assert!(
            out.sql.contains("\"_ids\" AS (\nUPDATE") && out.sql.contains("\nFROM \""),
            "expected the _ids UPDATE to gain a FROM clause referencing the nested CTE, got:\n{}",
            out.sql,
        );
        assert!(
            out.sql.contains("\"_ml_add_0\""),
            "expected the junction-append CTE to still be present, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_update_link_value_from_nested_insert_combines_with_interface_target() {
        // A nested-insert link value on an UPDATE targeting an interface
        // (abstract + materialized) type routes through the poly fan-out
        // branch (emit_poly_update_stmt) — a third, independent emitter path
        // from the previous two tests. The hoisted CTE must be materialized
        // once and referenced from every per-implementor UPDATE branch.
        fn id_prop() -> PropertyDescriptor {
            PropertyDescriptor {
                name: "id".into(),
                pg_type: "uuid".into(),
                nullable: false,
                default_sql: Some("uuidv7()".into()),
                default_pyql: None,
                description: None,
                check_constraints: vec![],
                is_exclusive: true,
                is_pk: true,
                is_readonly: true,
                rewrites: vec![],
                tuple_members: None,
                column_type: None,
            }
        }
        fn text_prop(name: &str) -> PropertyDescriptor {
            PropertyDescriptor {
                name: name.into(),
                pg_type: "text".into(),
                nullable: false,
                default_sql: None,
                default_pyql: None,
                description: None,
                check_constraints: vec![],
                is_exclusive: false,
                is_pk: false,
                is_readonly: false,
                rewrites: vec![],
                tuple_members: None,
                column_type: None,
            }
        }
        fn company_link() -> LinkDescriptor {
            LinkDescriptor {
                name: "company".into(),
                target: "default::Company".into(),
                nullable: true,
                through: None,
                description: None,
                default_pyql: None,
                is_exclusive: false,
                is_readonly: false,
                rewrites: vec![],
                on_delete: vec![],
            }
        }
        let schema = SchemaDescriptor {
            types: vec![
                TypeDescriptor {
                    name: "Company".into(),
                    module: "default".into(),
                    table: "Company".into(),
                    abstract_: false,
                    materialized: true,
                    description: None,
                    parents: vec![],
                    interfaces: vec![],
                    bases: vec![],
                    properties: vec![id_prop(), text_prop("name")],
                    links: vec![],
                    multilinks: vec![],
                    computed: vec![],
                    constraints: vec![],
                    indexes: vec![],
                    partition: None,
                    vector_indexes: vec![],
                    search_indexes: vec![],
                    triggers: vec![],
                    junction: false,
                    signals: vec![],
                },
                TypeDescriptor {
                    name: "Account".into(),
                    module: "default".into(),
                    table: "Account".into(),
                    abstract_: true,
                    materialized: true,
                    description: None,
                    parents: vec![],
                    interfaces: vec![],
                    bases: vec![],
                    properties: vec![id_prop(), text_prop("email")],
                    links: vec![company_link()],
                    multilinks: vec![],
                    computed: vec![],
                    constraints: vec![],
                    indexes: vec![],
                    partition: None,
                    vector_indexes: vec![],
                    search_indexes: vec![],
                    triggers: vec![],
                    junction: false,
                    signals: vec![],
                },
                TypeDescriptor {
                    name: "Individual".into(),
                    module: "default".into(),
                    table: "Individual".into(),
                    abstract_: false,
                    materialized: true,
                    description: None,
                    parents: vec![],
                    interfaces: vec!["default::Account".into()],
                    bases: vec![],
                    properties: vec![id_prop(), text_prop("email"), text_prop("first_name")],
                    links: vec![company_link()],
                    multilinks: vec![],
                    computed: vec![],
                    constraints: vec![],
                    indexes: vec![],
                    partition: None,
                    vector_indexes: vec![],
                    search_indexes: vec![],
                    triggers: vec![],
                    junction: false,
                    signals: vec![],
                },
            ],
            scalars: vec![],
            enums: vec![],
            named_tuples: vec![],
            globals: vec![],
            functions: vec![],
            aliases: vec![],
            channels: vec![],
            ..Default::default()
        };
        let out = compile_and_emit_with(
            "UPDATE Account FILTER .email = $email \
             SET { company := (select (insert Company { name := 'Acme' }) { id }) }",
            &schema,
        );
        assert!(out.sql.starts_with("WITH"), "got:\n{}", out.sql);
        assert!(
            out.sql.contains("INSERT INTO \"public\".\"Company\""),
            "expected the nested insert to be its own CTE, got:\n{}",
            out.sql,
        );
        assert!(
            out.sql.contains("UPDATE \"public\".\"Individual\"") && out.sql.contains("\nFROM \"_nested_dml_0\""),
            "expected the per-implementor UPDATE to gain a FROM clause referencing the nested CTE, got:\n{}",
            out.sql,
        );
    }

    #[test]
    fn test_a_backlink_narrowed_to_a_supertype() {
        // `.<company[is Account]` — the intersection narrows what comes back,
        // while the link itself is declared further down. Pylon looked for it
        // on the narrowed type alone and reported it had no such link.
        let mut schema = interface_link_schema();
        // Leave `company` declared only on the implementor, so the narrowed
        // type genuinely does not have it.
        schema
            .types
            .iter_mut()
            .find(|t| t.name == "Account")
            .expect("the helper declares Account")
            .links
            .clear();
        let out = compile_and_emit_with("SELECT Company { owners := .<company[is Account] { email } }", &schema);
        assert!(
            out.sql.contains("\"public\".\"Individual\""),
            "the backlink must read the type that declares it:\n{}",
            out.sql
        );
    }

    #[test]
    fn a_path_through_a_backlink_narrowed_to_a_supertype() {
        let mut schema = interface_link_schema();
        schema
            .types
            .iter_mut()
            .find(|t| t.name == "Account")
            .expect("the helper declares Account")
            .links
            .clear();
        let out = compile_and_emit_with(
            "SELECT Company { name } FILTER EXISTS (SELECT .<company[is Account])",
            &schema,
        );
        assert!(out.sql.contains("\"public\".\"Individual\""), "{}", out.sql);
    }

    /// Company, the Account interface, its Individual implementor and a
    /// Token linking to the interface -- enough to exercise reading and
    /// backlinking across one.
    fn interface_link_schema() -> SchemaDescriptor {
        // A link whose target is an interface is read by fanning the
        // implementors out inline. That union used to project only the
        // interface's *properties*, so a shape that then followed one of the
        // interface's own links failed with "column t1.company_id does not
        // exist".
        fn id_prop() -> PropertyDescriptor {
            PropertyDescriptor {
                name: "id".into(),
                pg_type: "uuid".into(),
                nullable: false,
                default_sql: None,
                default_pyql: None,
                description: None,
                check_constraints: vec![],
                is_exclusive: true,
                is_pk: true,
                is_readonly: true,
                rewrites: vec![],
                tuple_members: None,
                column_type: None,
            }
        }
        fn text_prop(name: &str) -> PropertyDescriptor {
            PropertyDescriptor {
                name: name.into(),
                pg_type: "text".into(),
                nullable: false,
                default_sql: None,
                default_pyql: None,
                description: None,
                check_constraints: vec![],
                is_exclusive: false,
                is_pk: false,
                is_readonly: false,
                rewrites: vec![],
                tuple_members: None,
                column_type: None,
            }
        }
        fn link(name: &str, target: &str) -> LinkDescriptor {
            LinkDescriptor {
                name: name.into(),
                target: target.into(),
                nullable: true,
                through: None,
                description: None,
                default_pyql: None,
                is_exclusive: false,
                is_readonly: false,
                rewrites: vec![],
                on_delete: vec![],
            }
        }
        fn ty(
            name: &str,
            abstract_: bool,
            interfaces: Vec<String>,
            properties: Vec<PropertyDescriptor>,
            links: Vec<LinkDescriptor>,
        ) -> TypeDescriptor {
            TypeDescriptor {
                name: name.into(),
                module: "default".into(),
                table: name.into(),
                abstract_,
                materialized: true,
                description: None,
                parents: vec![],
                interfaces,
                bases: vec![],
                properties,
                links,
                multilinks: vec![],
                computed: vec![],
                constraints: vec![],
                indexes: vec![],
                partition: None,
                vector_indexes: vec![],
                search_indexes: vec![],
                triggers: vec![],
                junction: false,
                signals: vec![],
            }
        }
        SchemaDescriptor {
            types: vec![
                ty("Company", false, vec![], vec![id_prop(), text_prop("name")], vec![]),
                ty(
                    "Account",
                    true,
                    vec![],
                    vec![id_prop(), text_prop("email")],
                    vec![link("company", "default::Company")],
                ),
                ty(
                    "Individual",
                    false,
                    vec!["default::Account".into()],
                    vec![id_prop(), text_prop("email")],
                    vec![link("company", "default::Company")],
                ),
                ty(
                    "Token",
                    false,
                    vec![],
                    vec![id_prop(), text_prop("value")],
                    vec![link("account", "default::Account")],
                ),
            ],
            scalars: vec![],
            enums: vec![],
            named_tuples: vec![],
            globals: vec![],
            functions: vec![],
            aliases: vec![],
            channels: vec![],
            ..Default::default()
        }
    }

    /// `Token`, but reached through a second single link and narrowed to the
    /// interface's implementor -- conduit's trust-tier walk,
    /// `installation.connector.provider[is Individual].staff`.
    fn two_hop_interface_schema() -> SchemaDescriptor {
        let mut schema = interface_link_schema();
        let token = schema
            .types
            .iter()
            .find(|t| t.name == "Token")
            .expect("the helper declares Token")
            .clone();
        let mut connector = token.clone();
        connector.name = "Connector".into();
        connector.table = "Connector".into();
        connector.links[0].name = "provider".into();
        let mut installation = token;
        installation.name = "Installation".into();
        installation.table = "Installation".into();
        installation.links[0].name = "connector".into();
        installation.links[0].target = "default::Connector".into();
        schema.types.push(connector);
        schema.types.push(installation);
        schema
    }

    #[test]
    fn a_narrowed_two_hop_walk_off_a_pinned_row_is_one_value() {
        // Every step is a forward single link and the filter pins the id, so
        // the `[is Individual]` narrowing does not make the walk set-valued.
        // Gathered as an array the `??` default became `array_remove(ARRAY[…])`
        // and the result decoded as a one-element set where Gel gives the
        // value itself.
        let out = compile_and_emit_with(
            "SELECT { s := ((SELECT Installation FILTER .id = <uuid>$0).connector.provider[is Individual].email) ?? 'y' }",
            &two_hop_interface_schema(),
        );
        assert!(
            !out.sql.contains("ARRAY(SELECT"),
            "expected a value, not a set:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("COALESCE("),
            "the default must stay a scalar one:\n{}",
            out.sql
        );
    }

    #[test]
    fn a_narrowed_two_hop_walk_off_an_unpinned_row_stays_a_set() {
        let out = compile_and_emit_with(
            "SELECT { s := (SELECT Installation FILTER .value = 'x').connector.provider[is Individual].email }",
            &two_hop_interface_schema(),
        );
        assert!(
            out.sql.contains("ARRAY(SELECT"),
            "a walk off many rows is a set:\n{}",
            out.sql
        );
    }

    /// `exists .account[is Individual]` — a walk narrowed to an implementor
    /// with no pointer read after it. The two-step form needs a trailing
    /// column to read, so it rejected the intersection outright ("type
    /// intersections are not valid in expression context") even though the
    /// same walk with a pointer on the end (`[is Individual].email`) has
    /// always compiled.
    #[test]
    fn a_walk_ending_in_a_type_intersection_is_valid_in_a_filter() {
        let out = compile_and_emit_with(
            "SELECT Token { id } FILTER EXISTS .account[is Individual]",
            &interface_link_schema(),
        );
        assert!(
            out.sql.contains("\"public\".\"Individual\""),
            "the narrowing must reach the implementor's table:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("IS NOT NULL"),
            "exists over the narrowed walk is a nullness test:\n{}",
            out.sql
        );
    }

    #[test]
    fn a_walk_ending_in_a_type_intersection_is_valid_off_a_binding() {
        let out = compile_and_emit_with(
            "WITH t := (SELECT Token LIMIT 1) SELECT { e := EXISTS t.account[is Individual] }",
            &interface_link_schema(),
        );
        assert!(
            out.sql.contains("\"public\".\"Individual\""),
            "the narrowing must reach the implementor's table:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_reading_a_link_through_an_interface_target() {
        let schema = interface_link_schema();
        let out = compile_and_emit_with("SELECT Token { account: { email, company: { name } } }", &schema);
        assert!(
            out.sql
                .contains("\"email\", \"company_id\" FROM \"public\".\"Individual\""),
            "the fanned-out interface must carry its own link columns:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_computed_pointer_in_shape_emits_expression() {
        let mut schema = make_schema();
        schema.types[0].computed.push(crate::schema::ComputedDescriptor {
            name: "upper_name".into(),
            expression: "str_upper(.name)".into(),
            return_type: Some("text".into()),
            link_target: None,
            link_multi: false,
        });
        let out = compile_and_emit_with("SELECT Person { upper_name }", &schema);
        assert!(
            out.sql.to_lowercase().contains("upper"),
            "expected upper() in SQL, got:\n{}",
            out.sql
        );
    }

    /// Interface `Account` (non-materialized) with one implementor,
    /// `Individual`, carrying a stored property and a computed pointer.
    fn make_interface_schema() -> SchemaDescriptor {
        fn id_prop() -> PropertyDescriptor {
            PropertyDescriptor {
                name: "id".into(),
                pg_type: "uuid".into(),
                nullable: false,
                default_sql: Some("uuidv7()".into()),
                default_pyql: None,
                description: None,
                check_constraints: vec![],
                is_exclusive: true,
                is_pk: true,
                is_readonly: true,
                rewrites: vec![],
                tuple_members: None,
                column_type: None,
            }
        }
        SchemaDescriptor {
            types: vec![
                TypeDescriptor {
                    name: "Account".into(),
                    module: "default".into(),
                    table: "Account".into(),
                    abstract_: true,
                    materialized: false,
                    description: None,
                    parents: vec![],
                    interfaces: vec![],
                    bases: vec![],
                    properties: vec![
                        id_prop(),
                        PropertyDescriptor {
                            name: "email".into(),
                            pg_type: "text".into(),
                            nullable: false,
                            default_sql: None,
                            default_pyql: None,
                            description: None,
                            check_constraints: vec![],
                            is_exclusive: false,
                            is_pk: false,
                            is_readonly: false,
                            rewrites: vec![],
                            tuple_members: None,
                            column_type: None,
                        },
                    ],
                    links: vec![],
                    multilinks: vec![],
                    computed: vec![],
                    constraints: vec![],
                    indexes: vec![],
                    partition: None,
                    vector_indexes: vec![],
                    search_indexes: vec![],
                    triggers: vec![],
                    junction: false,
                    signals: vec![],
                },
                TypeDescriptor {
                    name: "Individual".into(),
                    module: "default".into(),
                    table: "Individual".into(),
                    abstract_: false,
                    materialized: true,
                    description: None,
                    parents: vec![],
                    interfaces: vec!["default::Account".into()],
                    bases: vec![],
                    properties: vec![
                        id_prop(),
                        PropertyDescriptor {
                            name: "first_name".into(),
                            pg_type: "text".into(),
                            nullable: false,
                            default_sql: None,
                            default_pyql: None,
                            description: None,
                            check_constraints: vec![],
                            is_exclusive: false,
                            is_pk: false,
                            is_readonly: false,
                            rewrites: vec![],
                            tuple_members: None,
                            column_type: None,
                        },
                    ],
                    links: vec![],
                    multilinks: vec![],
                    computed: vec![crate::schema::ComputedDescriptor {
                        name: "full_name".into(),
                        expression: "str_upper(.first_name)".into(),
                        return_type: Some("text".into()),
                        link_target: None,
                        link_multi: false,
                    }],
                    constraints: vec![],
                    indexes: vec![],
                    partition: None,
                    vector_indexes: vec![],
                    search_indexes: vec![],
                    triggers: vec![],
                    junction: false,
                    signals: vec![],
                },
            ],
            scalars: vec![],
            enums: vec![],
            named_tuples: vec![],
            globals: vec![],
            functions: vec![],
            aliases: vec![],
            channels: vec![],
            ..Default::default()
        }
    }

    #[test]
    fn test_type_intersection_splat_includes_concrete_computed_pointers() {
        // `[is Concrete].*` must expand to the concrete type's own computed
        // pointers too (e.g. `full_name`), not just its stored properties —
        // querying `Individual` directly with `*` already included them, but
        // the interface-splat path never even looked at `td.computed`.
        let schema = make_interface_schema();
        let out = compile_and_emit_with("SELECT Account { *, [is Individual].* }", &schema);
        assert!(
            out.sql.to_lowercase().contains("upper"),
            "expected the concrete type's computed pointer (str_upper(...)) in the shape, got:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("\"first_name\""),
            "expected the concrete type's stored property too, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_enum_array_property_casts_to_text_array_and_shapes_as_a_list() {
        // Regression: an `array<enum>` property's pg_type is
        // `"module"."Type"[]`, which starts with a quote exactly as a scalar
        // enum's does. Read as a scalar it was cast to a single `::text`,
        // so `{MagicLink}` -- Postgres's text rendering of the whole array --
        // came back as one enum value, and the polymorphic path, which casts
        // nothing, failed outright with "no decoder for PostgreSQL type OID".
        let mut schema = make_schema();
        schema.enums.push(crate::schema::EnumDescriptor {
            name: "Gender".into(),
            module: "default".into(),
            members: vec!["Male".into(), "Female".into()],
        });
        schema.types[0].properties.push(crate::schema::PropertyDescriptor {
            name: "genders".into(),
            pg_type: r#""default"."Gender"[]"#.into(),
            nullable: true,
            default_sql: None,
            default_pyql: None,
            description: None,
            check_constraints: vec![],
            is_exclusive: false,
            is_pk: false,
            is_readonly: false,
            rewrites: vec![],
            tuple_members: None,
            column_type: None,
        });

        let out = compile_and_emit_with("SELECT Person { genders }", &schema);
        assert!(
            out.sql.contains(r#""genders"::text[]"#),
            "expected a text[] cast, got:\n{}",
            out.sql
        );

        let crate::query::ShapeNode::Object { pointers, .. } = &out.shape.root else {
            panic!("expected Object shape, got {:?}", out.shape.root)
        };
        let genders = pointers
            .iter()
            .find(|node| matches!(node, crate::query::ShapeNode::Array { name, .. } if name == "genders"))
            .unwrap_or_else(|| panic!("expected an Array-shaped pointer, got {pointers:?}"));
        let crate::query::ShapeNode::Array { element, .. } = genders else {
            unreachable!()
        };
        assert!(
            matches!(
                element.as_ref(),
                crate::query::ShapeNode::Enum { enum_type, position: 0, .. }
                    if enum_type == "default::Gender"
            ),
            "expected the elements to be enum-shaped, got {element:?}",
        );
    }

    #[test]
    fn test_scalar_enum_property_still_casts_to_a_single_text() {
        let mut schema = make_schema();
        schema.enums.push(crate::schema::EnumDescriptor {
            name: "Gender".into(),
            module: "default".into(),
            members: vec!["Male".into(), "Female".into()],
        });
        schema.types[0].properties.push(crate::schema::PropertyDescriptor {
            name: "gender".into(),
            pg_type: r#""default"."Gender""#.into(),
            nullable: true,
            default_sql: None,
            default_pyql: None,
            description: None,
            check_constraints: vec![],
            is_exclusive: false,
            is_pk: false,
            is_readonly: false,
            rewrites: vec![],
            tuple_members: None,
            column_type: None,
        });

        let out = compile_and_emit_with("SELECT Person { gender }", &schema);
        assert!(
            out.sql.contains(r#""gender"::text"#) && !out.sql.contains(r#""gender"::text[]"#),
            "expected a plain text cast, got:\n{}",
            out.sql
        );
        let crate::query::ShapeNode::Object { pointers, .. } = &out.shape.root else {
            panic!("expected Object shape")
        };
        assert!(
            pointers.iter().any(|node| matches!(
                node,
                crate::query::ShapeNode::Enum { name, enum_type, .. }
                    if name == "gender" && enum_type == "default::Gender"
            )),
            "expected an Enum-shaped pointer, got {pointers:?}",
        );
    }

    #[test]
    fn test_type_intersection_splat_keeps_enum_pointers_enum_shaped() {
        // Regression: `[is Concrete].*` wraps each property in a correlated
        // subquery, so the outer shape saw an `IrExpr::Subquery` rather than a
        // column and fell back to a plain Scalar — handing back `'System'`
        // where a direct select gave `InterfaceAppearance.System`.
        let mut schema = make_interface_schema();
        schema.enums.push(crate::schema::EnumDescriptor {
            name: "Gender".into(),
            module: "default".into(),
            members: vec!["Male".into(), "Female".into()],
        });
        // An enum that lives only on the concrete type, so it can only be
        // reached through the intersection.
        let individual = schema
            .types
            .iter_mut()
            .find(|t| t.name == "Individual")
            .expect("the interface schema has an Individual type");
        individual.properties.push(crate::schema::PropertyDescriptor {
            name: "gender".into(),
            pg_type: r#""default"."Gender""#.into(),
            nullable: true,
            default_sql: None,
            default_pyql: None,
            description: None,
            check_constraints: vec![],
            is_exclusive: false,
            is_pk: false,
            is_readonly: false,
            rewrites: vec![],
            tuple_members: None,
            column_type: None,
        });

        let out = compile_and_emit_with("SELECT Account { [is Individual].* }", &schema);
        let crate::query::ShapeNode::Object { pointers, .. } = &out.shape.root else {
            panic!("expected Object shape, got {:?}", out.shape.root)
        };
        assert!(
            pointers.iter().any(|node| matches!(
                node,
                crate::query::ShapeNode::Enum { name, enum_type, .. }
                    if name == "gender" && enum_type == "default::Gender"
            )),
            "expected the enum property to stay enum-shaped through the intersection, got {pointers:?}",
        );
    }

    #[test]
    fn test_shallow_splat_leaves_object_valued_computeds_to_the_deep_form() {
        // Regression: `*` expanded every computed, link-valued ones included,
        // which Gel's `*` does not -- it is properties only, `**` adds links.
        // On jaldis's schema that pulled a computed multi-link
        // (`members := .memberships.member`) into every splat query.
        let mut schema = make_schema();
        schema.types[0].computed.push(crate::schema::ComputedDescriptor {
            name: "authors".into(),
            expression: ".posts".into(),
            return_type: None,
            link_target: None,
            link_multi: false,
        });
        schema.types[0].computed.push(crate::schema::ComputedDescriptor {
            name: "age_next".into(),
            expression: ".age + 1".into(),
            return_type: Some("int8".into()),
            link_target: None,
            link_multi: false,
        });

        let shallow = compile_and_emit_with("SELECT Person { * }", &schema);
        let crate::query::ShapeNode::Object { pointers, .. } = &shallow.shape.root else {
            panic!("expected Object shape")
        };
        let names: Vec<&str> = pointers.iter().map(shape_pointer_name).collect();
        assert!(
            names.contains(&"age_next"),
            "a computed property belongs in `*`: {names:?}"
        );
        assert!(!names.contains(&"authors"), "a computed link does not: {names:?}");

        let deep = compile_and_emit_with("SELECT Person { ** }", &schema);
        let crate::query::ShapeNode::Object { pointers, .. } = &deep.shape.root else {
            panic!("expected Object shape")
        };
        let names: Vec<&str> = pointers.iter().map(shape_pointer_name).collect();
        assert!(
            names.contains(&"authors"),
            "`**` includes links, computed ones included: {names:?}"
        );
    }

    fn shape_pointer_name(node: &crate::query::ShapeNode) -> &str {
        use crate::query::ShapeNode;
        match node {
            ShapeNode::Scalar { name, .. }
            | ShapeNode::Enum { name, .. }
            | ShapeNode::NamedTuple { name, .. }
            | ShapeNode::Object { name, .. }
            | ShapeNode::Array { name, .. } => name,
            _ => "",
        }
    }

    #[test]
    fn test_any_over_a_multilink_comparison_does_not_warn() {
        // Regression: the warning fired whether or not the comparison was
        // wrapped, so `any()` -- the very thing it asks for -- did not silence
        // it. The argument is compiled before the call is resolved, so the
        // suppression has to be in place while the argument is compiled.
        let schema = make_schema();
        let ast = parse::parse("SELECT Person FILTER any(.posts.title = 'x')").unwrap();
        let ir = ir::compile(&ast, &schema).unwrap();
        assert!(
            !ir.warnings.iter().any(|w| w.contains("FILTER clause")),
            "any() states the intent, so nothing should warn: {:?}",
            ir.warnings
        );
    }

    #[test]
    fn test_bare_multilink_comparison_still_warns() {
        let schema = make_schema();
        let ast = parse::parse("SELECT Person FILTER .posts.title = 'x'").unwrap();
        let ir = ir::compile(&ast, &schema).unwrap();
        assert!(
            ir.warnings.iter().any(|w| w.contains("FILTER clause")),
            "an unwrapped set-valued comparison still warns: {:?}",
            ir.warnings
        );
    }

    #[test]
    fn test_link_to_an_interface_expands_over_its_implementors() {
        // Regression: an interface's view carries only the interface's own
        // columns, so reading a link through it tagged every row as the
        // interface and hydrated `Profile` where the row was an
        // `IndividualProfile`. The root of a query already expands inline;
        // a nested link did not.
        let mut schema = make_interface_schema();
        let account = schema
            .types
            .iter_mut()
            .find(|t| t.name == "Account")
            .expect("the interface schema has an Account type");
        account.materialized = true;
        // A second implementor, so the fan-out is a real union.
        let mut organization = schema
            .types
            .iter()
            .find(|t| t.name == "Individual")
            .expect("the interface schema has an Individual type")
            .clone();
        organization.name = "Organization".into();
        organization.table = "Organization".into();
        organization.computed.clear();
        schema.types.push(organization);

        let owner = TypeDescriptor {
            name: "Note".into(),
            module: "default".into(),
            table: "Note".into(),
            abstract_: false,
            materialized: false,
            description: None,
            parents: vec![],
            interfaces: vec![],
            bases: vec![],
            properties: vec![],
            links: vec![LinkDescriptor {
                name: "owner".into(),
                target: "default::Account".into(),
                nullable: true,
                description: None,
                default_pyql: None,
                is_exclusive: false,
                is_readonly: false,
                rewrites: vec![],
                on_delete: vec![],
                through: None,
            }],
            multilinks: vec![],
            computed: vec![],
            constraints: vec![],
            indexes: vec![],
            partition: None,
            vector_indexes: vec![],
            search_indexes: vec![],
            triggers: vec![],
            junction: false,
            signals: vec![],
        };
        schema.types.push(owner);

        let out = compile_and_emit_with("SELECT Note { owner: { id } }", &schema);
        assert!(
            out.sql.contains("UNION ALL"),
            "the link's target should expand over its implementors, got:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("'default::Individual'::text AS \"__type__\""),
            "each branch should carry its own discriminator, got:\n{}",
            out.sql
        );
        assert!(
            !out.sql.contains("FROM \"default\".\"Account\" AS"),
            "the interface's own view should no longer be read directly, got:\n{}",
            out.sql
        );
    }

    /// `make_interface_schema` with a second implementor of `Account`, so a
    /// read of it is a real union of two concrete types.
    fn make_two_implementor_schema() -> SchemaDescriptor {
        let mut schema = make_interface_schema();
        let mut organization = schema
            .types
            .iter()
            .find(|t| t.name == "Individual")
            .expect("the interface schema has an Individual type")
            .clone();
        organization.name = "Organization".into();
        organization.table = "Organization".into();
        organization.computed.clear();
        schema.types.push(organization);
        schema
    }

    /// Every row read of an interface names its own concrete type, so none may
    /// be labelled with the interface's.
    fn assert_rows_carry_their_concrete_type(query: &str) {
        let out = compile_and_emit_with(query, &make_two_implementor_schema());
        assert!(
            !out.sql.contains("'default::Account'::text"),
            "rows are labelled with the interface instead of their own type:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("'default::Individual'::text AS \"__type__\""),
            "the rows should come from the implementors, each tagging its own:\n{}",
            out.sql
        );
    }

    #[test]
    fn an_aggregated_select_of_an_interface_carries_the_concrete_type() {
        assert_rows_carry_their_concrete_type("SELECT (xs := array_agg((SELECT Account { id })))");
    }

    #[test]
    fn an_aggregated_binding_of_an_interface_carries_the_concrete_type() {
        assert_rows_carry_their_concrete_type("WITH a := (SELECT Account) SELECT (xs := array_agg((SELECT a { id })))");
    }

    #[test]
    fn a_single_object_read_off_a_binding_carries_the_concrete_type() {
        assert_rows_carry_their_concrete_type("WITH a := (SELECT Account) SELECT (x := (SELECT a { id } LIMIT 1))");
    }

    #[test]
    fn grouped_elements_of_an_interface_carry_the_concrete_type() {
        assert_rows_carry_their_concrete_type("group Account { id } by .email");
    }

    /// `make_two_implementor_schema` with a `Note` that links to an `Account`,
    /// and an `Account` that links back to a `Note`.
    fn make_note_schema() -> SchemaDescriptor {
        fn link_to(name: &str, target: &str) -> LinkDescriptor {
            LinkDescriptor {
                name: name.into(),
                target: target.into(),
                nullable: true,
                description: None,
                default_pyql: None,
                is_exclusive: false,
                is_readonly: false,
                rewrites: vec![],
                on_delete: vec![],
                through: None,
            }
        }
        let mut schema = make_two_implementor_schema();
        for account in schema
            .types
            .iter_mut()
            .filter(|t| ["Account", "Individual", "Organization"].contains(&t.name.as_str()))
        {
            account.links.push(link_to("pinned", "default::Note"));
        }
        schema.types.push(TypeDescriptor {
            name: "Note".into(),
            module: "default".into(),
            table: "Note".into(),
            abstract_: false,
            materialized: false,
            description: None,
            parents: vec![],
            interfaces: vec![],
            bases: vec![],
            properties: vec![],
            links: vec![link_to("owner", "default::Account")],
            multilinks: vec![],
            computed: vec![],
            constraints: vec![],
            indexes: vec![],
            partition: None,
            vector_indexes: vec![],
            search_indexes: vec![],
            triggers: vec![],
            junction: false,
            signals: vec![],
        });
        schema
    }

    #[test]
    fn a_backlink_to_an_interface_carries_the_concrete_type() {
        let out = compile_and_emit_with(
            "SELECT Note { pinners := .<pinned[is Account] { id } }",
            &make_note_schema(),
        );
        assert!(
            !out.sql.contains("'default::Account'::text"),
            "the backlink's rows are labelled with the interface instead of their own type:\n{}",
            out.sql
        );
    }

    #[test]
    fn a_link_reached_through_a_deep_splat_carries_the_concrete_type() {
        let out = compile_and_emit_with("SELECT Note { ** }", &make_note_schema());
        assert!(
            !out.sql.contains("'default::Account'::text"),
            "the splatted link's rows are labelled with the interface instead of their own type:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_path_traversal_onto_an_interface_carries_the_concrete_type() {
        // The hole the chained-pointer fix landed on: a *path join* onto an
        // interface still read the interface's view, so the rows a traversal
        // returned were tagged with the interface and hydrated its class. Links
        // fan out; this is the traversal route they share with computed chains.
        let mut schema = make_interface_schema();
        let account = schema
            .types
            .iter_mut()
            .find(|t| t.name == "Account")
            .expect("the interface schema has an Account type");
        account.materialized = true;
        let mut organization = schema
            .types
            .iter()
            .find(|t| t.name == "Individual")
            .expect("the interface schema has an Individual type")
            .clone();
        organization.name = "Organization".into();
        organization.table = "Organization".into();
        organization.computed.clear();
        schema.types.push(organization);
        schema.types.push(TypeDescriptor {
            name: "Note".into(),
            module: "default".into(),
            table: "Note".into(),
            abstract_: false,
            materialized: false,
            description: None,
            parents: vec![],
            interfaces: vec![],
            bases: vec![],
            properties: vec![],
            links: vec![LinkDescriptor {
                name: "owner".into(),
                target: "default::Account".into(),
                nullable: true,
                description: None,
                default_pyql: None,
                is_exclusive: false,
                is_readonly: false,
                rewrites: vec![],
                on_delete: vec![],
                through: None,
            }],
            multilinks: vec![],
            computed: vec![],
            constraints: vec![],
            indexes: vec![],
            partition: None,
            vector_indexes: vec![],
            search_indexes: vec![],
            triggers: vec![],
            junction: false,
            signals: vec![],
        });

        let out = compile_and_emit_with("SELECT Note.owner { id }", &schema);
        assert!(
            out.sql.contains("UNION ALL") && out.sql.contains("'default::Individual'::text AS \"__type__\""),
            "the traversal's target should fan out over implementors, got:\n{}",
            out.sql
        );
        assert!(
            !out.sql.contains("'default::Account'::text,"),
            "the row's type should be read off the row, not fixed to the interface, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_computed_multilink_chain_comes_back_as_rows() {
        // Regression: `members := .memberships.member` is a chain, and
        // try_compile_pointer_expr only recognised a single-step `.multilink`,
        // so it fell through to expression position and became a scalar
        // subquery — which Postgres rejects the moment a second row matches.
        let mut schema = make_schema();
        schema.types[0].computed.push(crate::schema::ComputedDescriptor {
            name: "coauthors".into(),
            expression: ".posts.author".into(),
            return_type: None,
            link_target: None,
            link_multi: false,
        });
        let post = schema
            .types
            .iter_mut()
            .find(|t| t.name == "Post")
            .expect("test schema has a Post type");
        post.links.push(LinkDescriptor {
            name: "author".into(),
            target: "default::Person".into(),
            nullable: true,
            description: None,
            default_pyql: None,
            is_exclusive: false,
            is_readonly: false,
            rewrites: vec![],
            on_delete: vec![],
            through: None,
        });

        let out = compile_and_emit_with("SELECT Person { coauthors: { name } }", &schema);
        assert!(
            out.sql.contains("ARRAY(SELECT"),
            "the chain should aggregate into an array, got:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("\"name\""),
            "the requested sub-shape should survive, got:\n{}",
            out.sql
        );

        let crate::query::ShapeNode::Object { pointers, .. } = &out.shape.root else {
            panic!("expected Object shape, got {:?}", out.shape.root)
        };
        let coauthors = pointers
            .iter()
            .find(|node| matches!(node, crate::query::ShapeNode::Array { name, .. } if name == "coauthors"))
            .unwrap_or_else(|| panic!("expected an Array-shaped pointer, got {pointers:?}"));
        let crate::query::ShapeNode::Array { element, .. } = coauthors else {
            unreachable!()
        };
        assert!(
            matches!(
                element.as_ref(),
                crate::query::ShapeNode::Object { type_name: Some(t), .. } if t == "default::Person"
            ),
            "the elements should hydrate as objects, got {element:?}",
        );
    }

    #[test]
    fn test_multi_sort_with_then_emits_two_order_keys() {
        let out = compile_and_emit("SELECT Person { name } ORDER BY .name THEN .age DESC");
        assert!(out.sql.contains("ORDER BY"), "expected ORDER BY");
        // Both columns should appear in the ORDER BY clause
        assert!(out.sql.contains("\"name\""));
        assert!(out.sql.contains("\"age\""));
        assert!(out.sql.contains("DESC"));
    }

    #[test]
    fn test_string_index_emits_str_subscript() {
        let out = compile_and_emit("SELECT 'hello'[1]");
        assert!(
            out.sql.contains("_pylon.str_subscript"),
            "expected _pylon.str_subscript() for string index, got:\n{}",
            out.sql
        );
    }

    /// PostgreSQL forbids a set-returning function inside `VALUES`, so a loop
    /// over `array_unpack` has to iterate a select instead.
    #[test]
    fn test_for_over_a_set_returning_call_is_not_a_values_clause() {
        let out = compile_and_emit("FOR n IN array_unpack(<array<int64>>$ns) UNION (SELECT n)");
        assert!(
            !out.sql.contains("VALUES (unnest"),
            "a set-returning iterator must not land in VALUES, got:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("SELECT unnest"),
            "expected the iterator in a select list, got:\n{}",
            out.sql
        );
    }

    /// The loop variable's type is the unpacked array's element type, not
    /// whatever the generically-declared call says.
    #[test]
    fn test_for_over_array_unpack_binds_the_element_type() {
        let out = compile_and_emit("FOR j IN array_unpack(<array<json>>$rows) UNION (SELECT <str>j['k'])");
        assert!(
            out.sql.contains("::jsonb AS v"),
            "the loop variable should carry the element type, got:\n{}",
            out.sql
        );
    }

    /// Indexing json reaches into it with `->`; neither string nor array
    /// subscripting applies.
    #[test]
    fn test_indexing_json_emits_a_jsonb_accessor() {
        let out = compile_and_emit("SELECT (<json>$payload)['label']");
        assert!(
            out.sql.contains("->'label'"),
            "expected a jsonb field access, got:\n{}",
            out.sql
        );
        assert!(
            !out.sql.contains("str_subscript"),
            "json must not be subscripted as a string, got:\n{}",
            out.sql
        );
    }

    /// Casting json to a scalar yields the value, not its JSON spelling —
    /// `<str>` of a json string has no quotes around it.
    #[test]
    fn test_casting_json_to_a_scalar_unwraps_it() {
        let out = compile_and_emit("SELECT <str>to_json(<str>$s)");
        assert!(
            out.sql.contains("#>> '{}'"),
            "expected the json value to be extracted, got:\n{}",
            out.sql
        );
    }

    /// A `with` binding that names the loop variable holds a different value
    /// each iteration. Evaluated once for the statement it reads an iterator
    /// that is not in its FROM; every read of it belongs to one iteration.
    #[test]
    fn test_a_binding_that_reads_the_loop_variable_is_keyed_by_iteration() {
        let out = compile_and_emit(
            "WITH made := (FOR p IN (SELECT Person) UNION ( \
               WITH mine := (SELECT p.posts LIMIT 1) \
               INSERT Company { name := mine.title } \
             )) SELECT count(made)",
        );
        assert!(
            out.sql.contains("\"mine\" AS (\nSELECT \"_for_p\".\"v\" AS \"_outer\""),
            "the binding must be evaluated per iteration, got:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("\"mine\".\"_outer\" = \"_for_p\".\"v\""),
            "a read of it must pin to the iteration in scope, got:\n{}",
            out.sql
        );
    }

    /// A `for … union (insert …)` assigned to a multi link *inside another
    /// loop* writes rows per iteration. They used to be cross-joined onto
    /// every row the outer loop wrote; they belong to one iteration of it.
    #[test]
    fn test_a_loop_nested_in_a_loop_pairs_its_rows_by_iteration() {
        let out = compile_and_emit(
            "FOR n IN {'a', 'b'} UNION (INSERT Person { name := n, posts := (FOR t IN {n} UNION (INSERT Post { title := t })) })",
        );
        assert!(
            out.sql.contains("\"_outer\""),
            "the nested rows must carry an iteration key, got:\n{}",
            out.sql
        );
        assert!(
            !out.sql.contains("CROSS JOIN \"_nested_dml_0\""),
            "they must not be cross-joined onto every row, got:\n{}",
            out.sql
        );
    }

    /// A multi link assigned in a `for`-bodied insert used to be dropped: the
    /// rows went in, the junction rows never did, and nothing reported it.
    #[test]
    fn test_for_bodied_insert_writes_its_multi_link() {
        let out = compile_and_emit("FOR n IN {'a', 'b'} UNION (INSERT Person { name := n, posts := (SELECT Post) })");
        assert!(
            out.sql.contains("INSERT INTO \"public\".\"Person.posts\""),
            "the junction rows must be written too, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_string_slice_emits_substr() {
        let out = compile_and_emit("SELECT 'hello'[1:3]");
        assert!(
            out.sql.contains("substr"),
            "expected substr() for string slice, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_array_index_emits_subscript() {
        let out = compile_and_emit("SELECT [1, 2, 3][1]");
        assert!(
            out.sql.contains("_pylon.array_subscript"),
            "expected _pylon.array_subscript() for array index, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_index_into_array_returning_function_emits_array_subscript() {
        let out = compile_and_emit("SELECT str_split('a::b', '::')[-1]");
        assert!(
            out.sql.contains("_pylon.array_subscript(string_to_array"),
            "expected _pylon.array_subscript() over str_split, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_array_slice_emits_subscript() {
        let out = compile_and_emit("SELECT [1, 2, 3][0:2]");
        assert!(
            !out.sql.contains("substr"),
            "should not use substr for array, got:\n{}",
            out.sql
        );
        assert!(out.sql.contains(")["), "expected array slice syntax, got:\n{}", out.sql);
    }

    #[test]
    fn test_open_ended_string_slice_emits_substr_no_length() {
        let out = compile_and_emit("SELECT 'hello'[2:]");
        // substr(expr, start) without length argument
        assert!(out.sql.contains("substr"), "expected substr(), got:\n{}", out.sql);
        // Should NOT have a third argument (length)
        let substr_idx = out.sql.find("substr").unwrap();
        let after = &out.sql[substr_idx..];
        let commas = after.chars().take_while(|&c| c != ')').filter(|&c| c == ',').count();
        assert_eq!(commas, 1, "open-ended slice should use 2-arg substr, got:\n{}", out.sql);
    }

    #[test]
    fn test_group_by_single_key() {
        let out = compile_and_emit("group Person { name } by .age");
        // GROUP BY clause present
        assert!(out.sql.contains("GROUP BY"), "expected GROUP BY, got:\n{}", out.sql);
        // Key column referenced
        assert!(out.sql.contains("\"age\""), "expected age column, got:\n{}", out.sql);
        // array_agg for elements
        assert!(
            out.sql.contains("array_agg(ROW("),
            "expected array_agg, got:\n{}",
            out.sql
        );
        // grouping names array
        assert!(
            out.sql.contains("ARRAY['age']"),
            "expected grouping array, got:\n{}",
            out.sql
        );
        // shape node is Group
        assert!(matches!(out.shape.root, crate::query::ShapeNode::Group { .. }));
        if let crate::query::ShapeNode::Group {
            key_nodes,
            grouping_position,
            elements_position,
            ..
        } = &out.shape.root
        {
            assert_eq!(key_nodes.len(), 1);
            assert!(matches!(&key_nodes[0], crate::query::ShapeNode::Scalar { name, position: 1 } if name == "age"));
            assert_eq!(*grouping_position, 2);
            assert_eq!(*elements_position, 3);
        }
    }

    #[test]
    fn test_group_using_alias() {
        let out = compile_and_emit("group Person using decade := .age // 10 by decade");
        assert!(out.sql.contains("GROUP BY"), "expected GROUP BY, got:\n{}", out.sql);
        assert!(
            out.sql.contains("ARRAY['decade']"),
            "expected grouping array, got:\n{}",
            out.sql
        );
        if let crate::query::ShapeNode::Group { key_nodes, .. } = &out.shape.root {
            assert_eq!(key_nodes.len(), 1);
            assert!(matches!(&key_nodes[0], crate::query::ShapeNode::Scalar { name, .. } if name == "decade"));
        }
    }

    #[test]
    fn test_group_orders_elements_within_each_group() {
        let out = compile_and_emit("group Person { name } by .age order by .name desc");
        assert!(
            out.sql.contains(")::record ORDER BY \"t0\".\"name\" DESC NULLS LAST)"),
            "{}",
            out.sql
        );
        assert!(!out.sql.contains("row_number()"), "{}", out.sql);
    }

    #[test]
    fn test_group_limit_trims_each_group_not_the_result() {
        // The newest row per key: a trailing LIMIT would drop whole groups,
        // so the rows are ranked within each key instead.
        let out = compile_and_emit("group Person { name } by .age order by .name desc limit 1");
        assert!(
            out.sql
                .contains("row_number() OVER (PARTITION BY \"t0\".\"age\" ORDER BY \"t0\".\"name\" DESC NULLS LAST)"),
            "{}",
            out.sql
        );
        assert!(
            out.sql.contains("WHERE \"__rk\" > 0 AND \"__rk\" <= 0 + 1"),
            "{}",
            out.sql
        );
        assert!(!out.sql.contains("\nLIMIT"), "{}", out.sql);
    }

    #[test]
    fn test_group_by_two_keys_with_a_per_group_limit() {
        let out = compile_and_emit(
            "group Person { name } by .age, .name filter .age > 18 order by .name desc offset 1 limit 2",
        );
        assert!(
            out.sql.contains("PARTITION BY \"t0\".\"age\", \"t0\".\"name\""),
            "{}",
            out.sql
        );
        // The filter picks which rows are grouped, so it sits inside the
        // ranking subquery — ranking must not see rows the filter excluded.
        let where_pos = out.sql.find("WHERE (\"t0\".\"age\" > 18)").expect("filter");
        let rank_pos = out.sql.find("WHERE \"__rk\"").expect("rank filter");
        assert!(where_pos < rank_pos, "{}", out.sql);
        assert!(
            out.sql.contains("GROUP BY \"t0\".\"age\", \"t0\".\"name\""),
            "{}",
            out.sql
        );
    }

    #[test]
    fn test_group_filter_without_a_limit_is_a_plain_where() {
        let out = compile_and_emit("group Person { name } by .age filter .age > 18");
        assert!(out.sql.contains("WHERE (\"t0\".\"age\" > 18)"), "{}", out.sql);
        assert!(!out.sql.contains("row_number()"), "{}", out.sql);
    }

    #[test]
    fn test_abs_path_concat_same_type() {
        let out = compile_and_emit("SELECT Person.name ++ ' ' ++ Person.name");
        assert!(out.sql.contains("\"name\""), "expected name column, got:\n{}", out.sql);
        assert!(out.sql.contains("||"), "expected concat operator, got:\n{}", out.sql);
        assert!(out.sql.contains("FROM"), "expected FROM clause, got:\n{}", out.sql);
    }

    #[test]
    fn test_abs_path_single_property() {
        let out = compile_and_emit("SELECT Person.name");
        assert!(out.sql.contains("\"name\""), "expected name column, got:\n{}", out.sql);
        assert!(out.sql.contains("FROM"), "expected FROM clause, got:\n{}", out.sql);
    }

    #[test]
    fn test_pgvector_cast_emits_vector_type() {
        let out = compile_and_emit("SELECT <pgvector::vector>[1.0, 2.0, 3.0]");
        assert!(
            out.sql.contains("::vector"),
            "expected ::vector cast, got:\n{}",
            out.sql
        );
        assert!(out.sql.contains("ARRAY["), "expected ARRAY literal, got:\n{}", out.sql);
    }

    #[test]
    fn test_pgvector_euclidean_distance_emits_l2_operator() {
        let out = compile_and_emit(
            "SELECT pgvector::euclidean_distance(<pgvector::vector>[1.0, 2.0], <pgvector::vector>[3.0, 4.0])",
        );
        assert!(out.sql.contains("<->"), "expected <-> operator, got:\n{}", out.sql);
    }

    #[test]
    fn test_pgvector_cosine_distance_emits_cosine_operator() {
        let out = compile_and_emit(
            "SELECT pgvector::cosine_distance(<pgvector::vector>[1.0, 2.0], <pgvector::vector>[3.0, 4.0])",
        );
        assert!(out.sql.contains("<=>"), "expected <=> operator, got:\n{}", out.sql);
    }

    #[test]
    fn test_pgvector_neg_inner_product_emits_ip_operator() {
        let out = compile_and_emit(
            "SELECT pgvector::neg_inner_product(<pgvector::vector>[1.0, 2.0], <pgvector::vector>[3.0, 4.0])",
        );
        assert!(out.sql.contains("<#>"), "expected <#> operator, got:\n{}", out.sql);
    }

    #[test]
    fn test_pgvector_inner_product_negates_ip_operator() {
        let out = compile_and_emit(
            "SELECT pgvector::inner_product(<pgvector::vector>[1.0, 2.0], <pgvector::vector>[3.0, 4.0])",
        );
        assert!(out.sql.contains("<#>"), "expected <#> operator, got:\n{}", out.sql);
        assert!(out.sql.contains("0.0"), "expected negation of <#>, got:\n{}", out.sql);
    }

    #[test]
    fn test_crypto_digest_str_and_bytes_overloads_both_use_pgcrypto_digest() {
        let out = compile_and_emit("SELECT crypto::digest('hello', 'sha256')");
        assert!(
            out.sql.contains("digest("),
            "expected pgcrypto's digest(), got:\n{}",
            out.sql
        );

        let out = compile_and_emit("SELECT crypto::digest(std::from_hex('68656c6c6f'), 'sha256')");
        assert!(
            out.sql.contains("digest("),
            "expected pgcrypto's digest(), got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_crypto_hmac_str_and_bytes_overloads_both_use_pgcrypto_hmac() {
        let out = compile_and_emit("SELECT crypto::hmac('hello', 'key', 'sha256')");
        assert!(
            out.sql.contains("hmac("),
            "expected pgcrypto's hmac(), got:\n{}",
            out.sql
        );

        let out =
            compile_and_emit("SELECT crypto::hmac(std::from_hex('68656c6c6f'), std::from_hex('6b6579'), 'sha256')");
        assert!(
            out.sql.contains("hmac("),
            "expected pgcrypto's hmac(), got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_crypto_gen_salt_zero_arg_defaults_to_blowfish() {
        let out = compile_and_emit("SELECT crypto::gen_salt()");
        assert!(
            out.sql.contains("gen_salt('bf')"),
            "expected default 'bf' salt type, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_crypto_gen_salt_one_arg_passes_type_through() {
        let out = compile_and_emit("SELECT crypto::gen_salt('xdes')");
        assert!(
            out.sql.contains("gen_salt("),
            "expected gen_salt() call, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_crypto_gen_salt_iter_count_casts_to_int4() {
        let out = compile_and_emit("SELECT crypto::gen_salt('xdes', 5)");
        assert!(
            out.sql.contains("::int4"),
            "expected int8 -> int4 narrowing cast, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_crypto_crypt_uses_pgcrypto_crypt() {
        let out = compile_and_emit("SELECT crypto::crypt('hunter2', crypto::gen_salt())");
        assert!(
            out.sql.contains("crypt("),
            "expected pgcrypto's crypt(), got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_postgis_cast_emits_geometry_type() {
        let out = compile_and_emit("SELECT <postgis::geometry>'POINT(1 2)'");
        assert!(
            out.sql.contains("::geometry"),
            "expected ::geometry cast, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_postgis_x_uses_st_x_builtin() {
        let out = compile_and_emit("SELECT postgis::x(<postgis::geometry>'POINT(1 2)')");
        assert!(out.sql.contains("st_x("), "expected st_x() call, got:\n{}", out.sql);
    }

    #[test]
    fn test_postgis_area_geometry_and_geography_overloads() {
        let out = compile_and_emit("SELECT postgis::area(<postgis::geometry>'POINT(1 2)')");
        assert!(
            out.sql.contains("st_area("),
            "expected st_area() call, got:\n{}",
            out.sql
        );

        let out = compile_and_emit("SELECT postgis::area(<postgis::geography>'POINT(1 2)', true)");
        assert!(
            out.sql.contains("st_area("),
            "expected st_area() call, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_postgis_setsrid_casts_int64_arg_to_int4() {
        let out = compile_and_emit("SELECT postgis::setsrid(<postgis::geometry>'POINT(1 2)', 4326)");
        assert!(
            out.sql.contains("st_setsrid("),
            "expected st_setsrid() call, got:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("::int4"),
            "expected int8 -> int4 narrowing cast, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_postgis_quantizecoordinates_default_arity_variants_compile() {
        // The reference binding documents this with 3 trailing optional
        // params; Pylon has no notion of default args, so each arity is
        // its own registered overload — confirm both the 2-arg and 4-arg
        // forms resolve.
        let out = compile_and_emit("SELECT postgis::quantizecoordinates(<postgis::geometry>'POINT(1 2)', 5)");
        assert!(out.sql.contains("st_quantizecoordinates("), "got:\n{}", out.sql);

        let out = compile_and_emit("SELECT postgis::quantizecoordinates(<postgis::geometry>'POINT(1 2)', 5, 5, 5)");
        assert!(out.sql.contains("st_quantizecoordinates("), "got:\n{}", out.sql);
    }

    #[test]
    fn test_postgis_op_contains_emits_infix_operator_not_function_call() {
        // Regression: ImplStrategy::SqlOperator was never actually consulted
        // by resolve_fn_call — it fell through to the generic "schema.name(args)"
        // FunctionCall path, which would have emitted a nonexistent
        // `"postgis".op_contains(...)` call instead of the `~` operator.
        let out = compile_and_emit(
            "SELECT postgis::op_contains(<postgis::geometry>'POINT(1 2)', <postgis::geometry>'POINT(3 4)')",
        );
        assert!(out.sql.contains(" ~ "), "expected infix ~ operator, got:\n{}", out.sql);
        assert!(
            !out.sql.contains("op_contains("),
            "must not call a literal op_contains function, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_postgis_op_overlaps_geometry_and_geography_overloads() {
        let out = compile_and_emit(
            "SELECT postgis::op_overlaps(<postgis::geometry>'POINT(1 2)', <postgis::geometry>'POINT(3 4)')",
        );
        assert!(
            out.sql.contains(" && "),
            "expected infix && operator, got:\n{}",
            out.sql
        );

        let out = compile_and_emit(
            "SELECT postgis::op_overlaps(<postgis::geography>'POINT(1 2)', <postgis::geography>'POINT(3 4)')",
        );
        assert!(
            out.sql.contains(" && "),
            "expected infix && operator, got:\n{}",
            out.sql
        );
    }

    // ── User-defined function tests ───────────────────────────────────────────

    fn make_schema_with_fns() -> SchemaDescriptor {
        let mut s = make_schema();
        s.functions = vec![
            FunctionDescriptor {
                name: "mysum".into(),
                module: "default".into(),
                params: vec![
                    FunctionParamDescriptor {
                        name: "a".into(),
                        pg_type: "int8".into(),
                    },
                    FunctionParamDescriptor {
                        name: "b".into(),
                        pg_type: "int8".into(),
                    },
                ],
                return_pg_type: "int8".into(),
                return_is_object: false,
                return_is_set: false,
                return_is_polymorphic: false,
                volatility: "immutable".into(),
                body: "a + b".into(),
            },
            FunctionDescriptor {
                name: "adults".into(),
                module: "default".into(),
                params: vec![],
                return_pg_type: "default::Person".into(),
                return_is_object: true,
                return_is_set: true,
                return_is_polymorphic: false,
                volatility: "stable".into(),
                body: "select Person filter .age > 18".into(),
            },
            // A second, 3-arg overload of `mysum`, declared *before* the
            // 2-arg one is looked up by resolution below — regression for a
            // bug where call resolution grabbed the first name-matching
            // overload regardless of arity, instead of searching all
            // same-named overloads for one whose param count actually
            // matches the call site.
            FunctionDescriptor {
                name: "mysum".into(),
                module: "default".into(),
                params: vec![
                    FunctionParamDescriptor {
                        name: "a".into(),
                        pg_type: "int8".into(),
                    },
                    FunctionParamDescriptor {
                        name: "b".into(),
                        pg_type: "int8".into(),
                    },
                    FunctionParamDescriptor {
                        name: "c".into(),
                        pg_type: "int8".into(),
                    },
                ],
                return_pg_type: "int8".into(),
                return_is_object: false,
                return_is_set: false,
                return_is_polymorphic: false,
                volatility: "immutable".into(),
                body: "a + b + c".into(),
            },
        ];
        s
    }

    #[test]
    fn test_user_fn_scalar_call() {
        let schema = make_schema_with_fns();
        let out = compile_and_emit_with("SELECT mysum(1, 2)", &schema);
        assert!(out.sql.contains("\"public\".\"mysum\""), "got:\n{}", out.sql);
    }

    #[test]
    fn test_user_fn_overload_resolved_by_argument_count() {
        // Regression: with a 2-arg and a 3-arg `mysum` overload both
        // declared, a 3-arg call must resolve to the 3-arg overload rather
        // than erroring against whichever overload happens to be first in
        // `schema.functions`.
        let schema = make_schema_with_fns();
        let out = compile_and_emit_with("SELECT mysum(1, 2, 3)", &schema);
        assert!(
            out.sql
                .contains("\"public\".\"mysum\"((1)::int8, (2)::int8, (3)::int8)"),
            "got:\n{}",
            out.sql,
        );

        // The 2-arg call must still resolve to the 2-arg overload.
        let out = compile_and_emit_with("SELECT mysum(1, 2)", &schema);
        assert!(
            out.sql.contains("\"public\".\"mysum\"((1)::int8, (2)::int8)"),
            "got:\n{}",
            out.sql,
        );
    }

    #[test]
    fn test_user_fn_object_select_no_shape() {
        let schema = make_schema_with_fns();
        let out = compile_and_emit_with("SELECT adults()", &schema);
        assert!(out.sql.contains("\"public\".\"adults\"()"), "got:\n{}", out.sql);
        assert!(out.sql.contains("FROM"), "got:\n{}", out.sql);
    }

    #[test]
    fn test_user_fn_object_select_with_shape() {
        let schema = make_schema_with_fns();
        let out = compile_and_emit_with("SELECT adults() { name }", &schema);
        assert!(out.sql.contains("\"public\".\"adults\"()"), "got:\n{}", out.sql);
        assert!(out.sql.contains("\"name\""), "got:\n{}", out.sql);
    }

    #[test]
    fn test_object_fn_projected_to_a_property_inside_an_expression() {
        let schema = make_schema_with_fns();
        let out = compile_and_emit_with("SELECT Person { n := adults().name }", &schema);
        assert!(out.sql.contains("FROM \"public\".\"adults\"() AS"), "got:\n{}", out.sql);
        assert!(out.sql.contains("\".\"name\"\n"), "got:\n{}", out.sql);
    }

    #[test]
    fn test_object_fn_sub_select_carries_its_modifiers() {
        let schema = make_schema_with_fns();
        let out = compile_and_emit_with(
            "SELECT Person { n := (select adults() filter .age > 21 limit 1).name }",
            &schema,
        );
        assert!(out.sql.contains("FROM \"public\".\"adults\"() AS"), "got:\n{}", out.sql);
        assert!(out.sql.contains("\"age\" > 21"), "got:\n{}", out.sql);
        assert!(out.sql.contains("LIMIT 1"), "got:\n{}", out.sql);
    }

    #[test]
    fn test_object_fn_without_a_property_still_explains_the_restriction() {
        let schema = make_schema_with_fns();
        let ast = parse::parse("SELECT Person { n := adults() + 1 }").unwrap();
        let err = match ir::compile(&ast, &schema) {
            Ok(_) => panic!("expected a compile error"),
            Err(e) => format!("{e}"),
        };
        assert!(err.contains("returns objects"), "{err}");
    }

    #[test]
    fn test_object_fn_projected_to_an_unknown_property_is_rejected() {
        let schema = make_schema_with_fns();
        let ast = parse::parse("SELECT Person { n := adults().nope }").unwrap();
        let err = match ir::compile(&ast, &schema) {
            Ok(_) => panic!("expected a compile error"),
            Err(e) => format!("{e}"),
        };
        assert!(err.contains("nope"), "{err}");
    }

    #[test]
    fn test_user_fn_in_cte_exposes_raw_columns() {
        // Regression: FunctionSelect as a CTE source must emit SELECT * FROM fn()
        // so the outer query can reference raw columns like t1.age.
        let schema = make_schema_with_fns();
        let out = compile_and_emit_with("WITH persons := adults() SELECT persons FILTER .age > 25", &schema);
        assert!(
            out.sql.contains("SELECT * FROM \"public\".\"adults\"()"),
            "CTE source must be SELECT * FROM fn(), got:\n{}",
            out.sql,
        );
        assert!(
            out.sql.contains("\"age\""),
            "outer filter must reference raw column, got:\n{}",
            out.sql
        );
    }

    // ── vector::search tests ──────────────────────────────────────────────────

    fn make_schema_with_vector() -> SchemaDescriptor {
        use crate::schema::VectorIndexDescriptor;
        let mut s = make_schema();
        if let Some(td) = s.types.iter_mut().find(|t| t.name == "Person") {
            td.vector_indexes.push(VectorIndexDescriptor {
                index_name: None,
                pointers: vec!["name".into()],
                model: "test-embed".into(),
                metric: "cosine".into(),
                dimensions: 4,
            });
        }
        s
    }

    #[test]
    fn test_vector_search_bare_type_name() {
        let schema = make_schema_with_vector();
        let out = compile_and_emit_with(
            "WITH search := vector::search(Person, <pgvector::vector>[1.0, 2.0, 3.0, 4.0]) \
             SELECT search { object { name }, distance }",
            &schema,
        );
        assert!(
            out.sql.contains("\"Person\""),
            "expected Person table, got:\n{}",
            out.sql
        );
        assert!(out.sql.contains("<=>"), "expected cosine operator, got:\n{}", out.sql);
    }

    #[test]
    fn test_vector_search_qualified_type_name() {
        let schema = make_schema_with_vector();
        let out = compile_and_emit_with(
            "WITH search := vector::search(default::Person, <pgvector::vector>[1.0, 2.0, 3.0, 4.0]) \
             SELECT search { object { name }, distance }",
            &schema,
        );
        assert!(
            out.sql.contains("\"Person\""),
            "expected Person table, got:\n{}",
            out.sql
        );
        assert!(out.sql.contains("<=>"), "expected cosine operator, got:\n{}", out.sql);
    }

    #[test]
    fn test_vector_search_subquery_filter_included_in_where() {
        let schema = make_schema_with_vector();
        let out = compile_and_emit_with(
            "WITH search := vector::search((select Person filter .name = 'Alice'), <pgvector::vector>[1.0, 2.0, 3.0, 4.0]) \
             SELECT search { object { name }, distance }",
            &schema,
        );
        assert!(
            out.sql.contains("\"Person\""),
            "expected Person table, got:\n{}",
            out.sql
        );
        assert!(out.sql.contains("\"name\""), "expected name filter, got:\n{}", out.sql);
        assert!(out.sql.contains("Alice"), "expected filter value, got:\n{}", out.sql);
        assert!(out.sql.contains("<=>"), "expected cosine operator, got:\n{}", out.sql);
    }

    #[test]
    fn test_vector_search_subquery_filter_combined_with_outer_property_filter() {
        let schema = make_schema_with_vector();
        let out = compile_and_emit_with(
            "WITH search := vector::search((select Person filter .age > 18), <pgvector::vector>[1.0, 2.0, 3.0, 4.0]) \
             SELECT search { object { name }, distance }",
            &schema,
        );
        assert!(
            out.sql.contains("\"age\""),
            "expected age pre-filter, got:\n{}",
            out.sql
        );
        assert!(out.sql.contains("18"), "expected filter value 18, got:\n{}", out.sql);
        assert!(out.sql.contains("<=>"), "expected cosine operator, got:\n{}", out.sql);
    }

    #[test]
    fn test_vector_search_text_overload_with_subquery_filter() {
        let schema = make_schema_with_vector();
        let out = compile_and_emit_with(
            "WITH search := vector::search((select Person filter .name = 'Alice'), query := $q) \
             SELECT search { object { name }, distance }",
            &schema,
        );
        assert!(
            out.sql.contains("\"Person\""),
            "expected Person table, got:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("Alice"),
            "expected pre-filter value, got:\n{}",
            out.sql
        );
        assert!(
            out.sql.contains("float8[]"),
            "expected float8[] cast for deferred vec param, got:\n{}",
            out.sql
        );
        assert!(out.sql.contains("<=>"), "expected cosine operator, got:\n{}", out.sql);
    }

    #[test]
    fn test_count_type_ref_compiles_to_agg_over_query() {
        let out = compile_and_emit("SELECT count(Person)");
        assert!(out.sql.contains("count(*)"), "expected count(*), got:\n{}", out.sql);
        assert!(
            out.sql.contains("\"Person\""),
            "expected Person table, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_count_qualified_type_ref_compiles_to_agg_over_query() {
        let out = compile_and_emit("SELECT count(default::Person)");
        assert!(out.sql.contains("count(*)"), "expected count(*), got:\n{}", out.sql);
        assert!(
            out.sql.contains("\"Person\""),
            "expected Person table, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_count_subquery_compiles_to_agg_over_query() {
        let out = compile_and_emit("SELECT count((select Person))");
        assert!(out.sql.contains("count(*)"), "expected count(*), got:\n{}", out.sql);
        assert!(
            out.sql.contains("\"Person\""),
            "expected Person table, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_count_subquery_with_filter() {
        let out = compile_and_emit("SELECT count((select Person filter .name = 'Alice'))");
        assert!(out.sql.contains("count(*)"), "expected count(*), got:\n{}", out.sql);
        assert!(
            out.sql.contains("\"name\""),
            "expected filter on name, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_positional_param_compiles_to_dollar_n() {
        let out = compile_and_emit("SELECT Person FILTER .name = $0");
        assert!(out.sql.contains("$1"), "expected $1 placeholder, got:\n{}", out.sql);
    }

    #[test]
    fn test_multiple_positional_params_compile_in_order() {
        let out = compile_and_emit("SELECT Person FILTER .name = $0 AND .age > $1");
        assert!(out.sql.contains("$1"), "expected $1, got:\n{}", out.sql);
        assert!(out.sql.contains("$2"), "expected $2, got:\n{}", out.sql);
    }

    #[test]
    fn test_repeated_positional_param_reuses_slot() {
        let out = compile_and_emit("SELECT Person FILTER .name = $0 OR .name = $0");
        assert_eq!(
            out.sql.matches("$1").count(),
            2,
            "both uses must reference $1, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_cast_to_nonexistent_type_names_full_type() {
        let schema = make_schema();
        let ast = parse::parse("SELECT Person FILTER .name = <default::Ghost>$name").unwrap();
        match ir::compile(&ast, &schema) {
            Ok(_) => panic!("expected compile error for unknown type"),
            Err(e) => {
                let msg = e.to_string();
                assert!(
                    msg.contains("unknown type 'default::Ghost'"),
                    "expected full type name in error, got: {msg}",
                );
            }
        }
    }

    #[test]
    fn test_cast_to_nonexistent_unqualified_type_names_type() {
        let schema = make_schema();
        let ast = parse::parse("SELECT Person FILTER .name = <Ghost>$name").unwrap();
        match ir::compile(&ast, &schema) {
            Ok(_) => panic!("expected compile error for unknown type"),
            Err(e) => {
                let msg = e.to_string();
                assert!(
                    msg.contains("unknown type 'Ghost'"),
                    "expected type name in error, got: {msg}",
                );
            }
        }
    }

    #[test]
    fn test_top_level_cast_to_registered_scalar_uses_its_own_domain() {
        // A bare top-level `SELECT <module::Scalar>expr` must compile as a
        // plain scalar cast to that scalar's own DOMAIN — not get misrouted
        // to `compile_schema_cast_select`'s object-by-id lookup (which
        // treated any non-stdlib-module named cast target as an object type,
        // since it never checked `resolve_scalar` before falling through).
        use crate::schema::ScalarDescriptor;
        let mut schema = make_schema();
        schema.scalars.push(ScalarDescriptor {
            name: "Email".into(),
            module: "default".into(),
            base: "Str".into(),
            pg_type: "text".into(),
            check_constraints: vec!["value ~ '^[^@]+@[^@]+\\.[^@]+$'".into()],
            is_sequence: false,
        });
        let out = compile_and_emit_with("SELECT <default::Email>'test@test.de'", &schema);
        assert!(
            out.sql.contains("\"public\".\"Email\""),
            "expected a cast to the scalar's own domain, got:\n{}",
            out.sql
        );
        assert!(
            !out.sql.to_lowercase().contains("\"person\""),
            "must not be misrouted to an object-type lookup, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_top_level_cast_to_unknown_type_names_full_type() {
        // The object-cast fallback (`compile_schema_cast_select`) used to
        // discard the module before building its synthetic path, so an
        // unknown top-level cast target reported a bare name even when
        // written with an explicit module.
        let schema = make_schema();
        let ast = parse::parse("SELECT <default::Ghost>$name").unwrap();
        match ir::compile(&ast, &schema) {
            Ok(_) => panic!("expected compile error for unknown type"),
            Err(e) => {
                let msg = e.to_string();
                assert!(
                    msg.contains("unknown type 'default::Ghost'"),
                    "expected full type name in error, got: {msg}",
                );
            }
        }
    }

    #[test]
    fn test_structural_tuple_cast_unnamed_resolves_to_jsonb() {
        let out = compile_and_emit("SELECT <tuple<str, bool>>$p");
        assert!(out.sql.contains("($1)::jsonb"), "got:\n{}", out.sql);
    }

    #[test]
    fn test_jsonb_to_uuid_cast_extracts_via_text() {
        // PostgreSQL has no native jsonb -> uuid cast; extract raw text then cast.
        let out = compile_and_emit("SELECT <uuid>(<json>$p)");
        assert!(out.sql.contains("#>> '{}'"), "got:\n{}", out.sql);
        assert!(out.sql.contains("::uuid"), "got:\n{}", out.sql);
    }

    #[test]
    fn test_jsonb_to_datetime_cast_extracts_via_text() {
        let out = compile_and_emit("SELECT <datetime>(<json>$p)");
        assert!(out.sql.contains("#>> '{}'"), "got:\n{}", out.sql);
        assert!(out.sql.contains("::timestamptz"), "got:\n{}", out.sql);
    }

    #[test]
    fn test_jsonb_to_duration_cast_extracts_via_text() {
        let out = compile_and_emit("SELECT <duration>(<json>$p)");
        assert!(out.sql.contains("#>> '{}'"), "got:\n{}", out.sql);
        assert!(out.sql.contains("::interval"), "got:\n{}", out.sql);
    }

    #[test]
    fn test_jsonb_to_array_cast_unpacks_each_element() {
        let out = compile_and_emit("SELECT <array<int64>>(<json>$p)");
        assert!(out.sql.contains("jsonb_array_elements("), "got:\n{}", out.sql);
        assert!(out.sql.contains("#>> '{}'"), "got:\n{}", out.sql);
        assert!(out.sql.contains("::int8"), "got:\n{}", out.sql);
    }

    #[test]
    fn test_non_jsonb_cast_is_unaffected_by_jsonb_extraction() {
        // A plain str -> uuid cast (source isn't jsonb) must still use the
        // ordinary `::pg_type` path, not the jsonb-extraction template.
        let out = compile_and_emit("SELECT <uuid>$p");
        assert!(!out.sql.contains("#>>"), "got:\n{}", out.sql);
        assert!(out.sql.contains("::uuid"), "got:\n{}", out.sql);
    }

    #[test]
    fn test_datetime_plus_duration_type_checks() {
        // Regression: types_compatible's bucket-matching (same type, or
        // both-int, or both-float) rejected this outright, even though
        // Postgres supports `timestamptz + interval` natively and this is
        // a legitimate operator overload.
        let out = compile_and_emit("SELECT <datetime>$p + <duration>$d");
        assert!(out.sql.contains(" + "), "got:\n{}", out.sql);
    }

    #[test]
    fn test_duration_plus_datetime_type_checks_either_order() {
        let out = compile_and_emit("SELECT <duration>$d + <datetime>$p");
        assert!(out.sql.contains(" + "), "got:\n{}", out.sql);
    }

    #[test]
    fn test_local_date_plus_duration_type_checks() {
        let out = compile_and_emit("SELECT <cal::local_date>$p + <duration>$d");
        assert!(out.sql.contains(" + "), "got:\n{}", out.sql);
    }

    #[test]
    fn test_datetime_minus_duration_type_checks() {
        let out = compile_and_emit("SELECT <datetime>$p - <duration>$d");
        assert!(out.sql.contains(" - "), "got:\n{}", out.sql);
    }

    #[test]
    fn test_local_time_plus_duration_type_checks() {
        let out = compile_and_emit("SELECT <cal::local_time>$p + <duration>$d");
        assert!(out.sql.contains(" + "), "got:\n{}", out.sql);
    }

    #[test]
    fn test_datetime_plus_duration_only_allowed_for_add_and_sub() {
        // Comparing (or otherwise combining) a datetime and a duration with
        // anything other than +/- still isn't meaningful and must still error.
        let schema = make_schema();
        let ast = parse::parse("SELECT <datetime>$p = <duration>$d").unwrap();
        match ir::compile(&ast, &schema) {
            Err(err) => assert!(format!("{err}").contains("cannot be applied"), "got: {err}"),
            Ok(_) => panic!("expected a compile error"),
        }
    }

    #[test]
    fn test_int_and_float_mix_is_allowed() {
        // Every int width implicitly casts to every float width, so mixed
        // int/float arithmetic type-checks without an explicit cast —
        // Postgres's own operator resolution handles the actual promotion.
        let out = compile_and_emit("SELECT <int16>1 + <float32>2.0");
        assert!(out.sql.contains(" + "), "got:\n{}", out.sql);
    }

    #[test]
    fn test_int_and_decimal_mix_is_allowed() {
        // int64 -> bigint -> decimal is a separate implicit-cast branch
        // from int64 -> float32 -> float64, but every int width still
        // mixes with numeric/decimal directly.
        let out = compile_and_emit("SELECT <int64>1 + <decimal>2.0");
        assert!(out.sql.contains(" + "), "got:\n{}", out.sql);
    }

    #[test]
    fn test_float_and_decimal_mix_still_rejected() {
        // float and numeric/decimal are separate branches past int64 in
        // the implicit-cast graph and don't cast to each other — this
        // still needs an explicit cast.
        let schema = make_schema();
        let ast = parse::parse("SELECT <decimal>$p + <float64>$n").unwrap();
        match ir::compile(&ast, &schema) {
            Err(err) => assert!(format!("{err}").contains("cannot be applied"), "got: {err}"),
            Ok(_) => panic!("expected a compile error"),
        }
    }

    #[test]
    fn test_unknown_pointer_suggests_a_close_match() {
        // `Person` has a `name` property (see make_schema) — `nam` is close
        // enough (missing one trailing character) that a "Did you mean"
        // suggestion should fire.
        let schema = make_schema();
        let ast = parse::parse("SELECT Person { nam }").unwrap();
        match ir::compile(&ast, &schema) {
            Err(err) => assert_eq!(
                format!("{err}"),
                "object type 'default::Person' has no link or property 'nam'. Did you mean 'name'?"
            ),
            Ok(_) => panic!("expected a compile error"),
        }
    }

    #[test]
    fn test_unrelated_unknown_pointer_gets_no_suggestion() {
        // A typo with no plausible match on the type shouldn't force a
        // confusing, unrelated suggestion onto the user.
        let schema = make_schema();
        let ast = parse::parse("SELECT Person { xyzxyzxyz }").unwrap();
        match ir::compile(&ast, &schema) {
            Err(err) => {
                let msg = format!("{err}");
                assert!(msg.contains("has no link or property 'xyzxyzxyz'"), "got: {msg}");
                assert!(!msg.contains("Did you mean"), "got: {msg}");
            }
            Ok(_) => panic!("expected a compile error"),
        }
    }

    #[test]
    fn test_multilink_sourced_backlink_exists_filter_compiles() {
        // Regression for a gap live-execution testing caught: `filter exists
        // .<multilink[is Type]` (a backlink whose source pointer on the
        // target type is a multi-link, e.g. `Person.posts`, not a single
        // FK link) used to fail to compile with "has no link ... pointing
        // to" — `compile_backlink_as_exists` only ever checked
        // `target_td.links`, never `target_td.multilinks`. Fast SQL-text
        // companion to the live-execution test in
        // `tests/live_execution_backlinks.rs`, which additionally confirms
        // the emitted SQL returns the right rows.
        let out = compile_and_emit("SELECT Post filter exists .<posts[is Person]");
        assert!(out.sql.contains("EXISTS"), "got:\n{}", out.sql);
    }

    #[test]
    fn test_backlink_as_computed_shape_pointer_compiles() {
        // Regression for the bigger gap live-execution testing caught: a
        // backlink could only be used in `filter exists .<...>` — using one
        // to select/project actual objects as a computed pointer's value
        // (`authors := .<posts[is Person] { name }`) failed to compile at
        // all ("shapes and set literals are not valid in expression
        // context"). `compile_shape_element`'s computed-pointer dispatch
        // now recognizes a backlink-rooted `Expr::Shape`/`Expr::Path` RHS
        // and routes it through `compile_backlink_pointer`, mirroring
        // `compile_multilink_pointer` for the forward direction. Fast
        // SQL-text companion to the live-execution tests in
        // `tests/live_execution_backlinks.rs`, which additionally confirm
        // the emitted SQL returns the right rows, nested 2 levels deep.
        let out = compile_and_emit("SELECT Post { title, authors := .<posts[is Person] { name } }");
        assert!(out.sql.contains("array_agg(ROW("), "got:\n{}", out.sql);
    }

    fn schema_with_exclusive_posts() -> SchemaDescriptor {
        let mut schema = make_schema();
        for td in schema.types.iter_mut().filter(|td| td.name == "Person") {
            for ml in td.multilinks.iter_mut().filter(|ml| ml.name == "posts") {
                ml.is_exclusive = true;
            }
        }
        schema
    }

    #[test]
    fn test_backlink_through_an_exclusive_multilink_is_single() {
        let out = compile_and_emit_with(
            "SELECT Post { title, author := .<posts[is Person] { name } }",
            &schema_with_exclusive_posts(),
        );
        let ShapeNode::Object { pointers, .. } = &out.shape.root else {
            panic!("expected an object root, got {:?}", out.shape.root)
        };
        assert!(
            matches!(pointers.last(), Some(ShapeNode::Object { name, .. }) if name == "author"),
            "got {pointers:?}"
        );
    }

    #[test]
    fn test_a_path_through_an_exclusive_backlink_is_single() {
        let out = compile_and_emit_with(
            "SELECT Post { title, author_name := .<posts[is Person].name }",
            &schema_with_exclusive_posts(),
        );
        let ShapeNode::Object { pointers, .. } = &out.shape.root else {
            panic!("expected an object root, got {:?}", out.shape.root)
        };
        assert!(
            matches!(pointers.last(), Some(ShapeNode::Scalar { name, .. }) if name == "author_name"),
            "got {pointers:?}"
        );
        assert!(!out.sql.contains("ARRAY("), "got:\n{}", out.sql);
    }

    #[test]
    fn test_structural_tuple_cast_named_resolves_to_jsonb() {
        let out = compile_and_emit("SELECT <tuple<x: float64, y: float64>>$p");
        assert!(out.sql.contains("($1)::jsonb"), "got:\n{}", out.sql);
    }

    #[test]
    fn test_structural_tuple_cast_nested_resolves_to_jsonb() {
        let out = compile_and_emit("SELECT <tuple<point: tuple<x: float64, y: float64>, label: str>>$p");
        assert!(out.sql.contains("($1)::jsonb"), "got:\n{}", out.sql);
    }

    #[test]
    fn test_nominal_named_tuple_cast_resolves_to_jsonb() {
        let mut schema = make_schema();
        schema.named_tuples.push(NamedTupleDescriptor {
            name: "Point".into(),
            module: "default".into(),
            members: vec![],
        });
        let out = compile_and_emit_with("SELECT <default::Point>$p", &schema);
        assert!(out.sql.contains("($1)::jsonb"), "got:\n{}", out.sql);
    }

    #[test]
    fn test_array_literal_cast_resolves_to_native_pg_array_not_jsonb() {
        // The exact query reported as failing — must compile now, and must
        // resolve to a real Postgres array (text[]), never jsonb (arrays
        // decode natively, unlike tuples).
        let out = compile_and_emit("SELECT <array<str>>['foo', 'bar']");
        assert!(
            out.sql.contains("::text[]") || out.sql.contains("ARRAY["),
            "got:\n{}",
            out.sql
        );
        assert!(
            !out.sql.contains("jsonb"),
            "arrays must not use jsonb, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_array_literal_cast_applies_per_element_cast() {
        // Each element gets its own real cast, not a raw untyped ARRAY[...] —
        // '1' and '2' must actually coerce to int8 (mirrors the equivalent
        // tuple test).
        let out = compile_and_emit("SELECT <array<int64>>['1', '2']");
        assert!(out.sql.contains("ARRAY[('1')::int8, ('2')::int8]"), "got:\n{}", out.sql);
    }

    #[test]
    fn test_array_param_cast_uses_direct_suffix_cast() {
        let out = compile_and_emit("SELECT <array<int64>>$p");
        assert!(out.sql.contains("::int8[]"), "got:\n{}", out.sql);
    }

    #[test]
    fn test_array_of_named_tuple_element_casts_to_jsonb_array() {
        // An array's element type can be anything except another array —
        // including a structural tuple, which still resolves that one
        // element to jsonb while the array itself stays a native pg array.
        let out = compile_and_emit("SELECT <array<tuple<x: float64, y: float64>>>$p");
        assert!(out.sql.contains("::jsonb[]"), "got:\n{}", out.sql);
    }

    #[test]
    fn test_contains_on_array_literal_cast_uses_array_overload_not_strpos() {
        // Regression: contains()'s array<Any> overload only matched a bare
        // IrExpr::Array node, never a TypeCast-wrapped one — which is what
        // every `<array<T>>[...]` cast actually compiles to. Overload
        // resolution silently fell back to the *first* registered `contains`
        // overload (str, str), emitting a bogus `strpos(text[], text)` call.
        let out = compile_and_emit("SELECT contains(<array<str>>[1, 2], '2')");
        assert!(out.sql.contains("@> ARRAY["), "got:\n{}", out.sql);
        assert!(
            !out.sql.contains("strpos"),
            "must not fall back to the str/str overload, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_nested_array_type_rejected_at_parse_time() {
        match parse::parse("SELECT <array<array<str>>>$p") {
            Ok(_) => panic!("expected parse error for nested array type"),
            Err(e) => assert!(e.to_string().contains("nested arrays are not supported"), "got: {}", e),
        }
    }

    #[test]
    fn test_array_cast_in_computed_shape_field_schema_bound_context() {
        // Schema-bound counterpart (compile_expr, not compile_free_expr) — a
        // computed shape field casting an array literal.
        let out = compile_and_emit("SELECT Person { name, tags := <array<str>>['a', 'b'] }");
        assert!(out.sql.contains("ARRAY[('a')::text, ('b')::text]"), "got:\n{}", out.sql);
    }

    #[test]
    fn test_nominal_named_tuple_cast_shape_carries_real_members() {
        use crate::schema::{TupleMemberDescriptor, TupleMemberKind};
        let mut schema = make_schema();
        schema.named_tuples.push(NamedTupleDescriptor {
            name: "Point".into(),
            module: "default".into(),
            members: vec![
                TupleMemberDescriptor {
                    name: Some("x".into()),
                    kind: TupleMemberKind::Scalar {
                        pg_type: "float8".into(),
                    },
                },
                TupleMemberDescriptor {
                    name: Some("y".into()),
                    kind: TupleMemberKind::Scalar {
                        pg_type: "float8".into(),
                    },
                },
            ],
        });
        let out = compile_and_emit_with("SELECT <default::Point>$p", &schema);
        match &out.shape.root {
            crate::query::ShapeNode::NamedTuple { type_name, members, .. } => {
                assert_eq!(type_name.as_deref(), Some("default::Point"));
                let members = members.as_ref().expect("expected resolved members");
                assert_eq!(members.len(), 2);
                assert_eq!(members[0].key.as_deref(), Some("x"));
                assert_eq!(members[1].key.as_deref(), Some("y"));
            }
            other => panic!("expected ShapeNode::NamedTuple, got {other:?}"),
        }
    }

    #[test]
    fn test_structural_tuple_property_read_shape_carries_real_members() {
        use crate::schema::{TupleMemberDescriptor, TupleMemberKind};
        let schema = SchemaDescriptor {
            types: vec![TypeDescriptor {
                name: "Person".into(),
                module: "default".into(),
                table: "Person".into(),
                abstract_: false,
                materialized: false,
                description: None,
                parents: vec![],
                interfaces: vec![],
                bases: vec![],
                properties: vec![PropertyDescriptor {
                    name: "address".into(),
                    pg_type: "jsonb".into(),
                    nullable: true,
                    default_sql: None,
                    default_pyql: None,
                    description: None,
                    check_constraints: vec![],
                    is_exclusive: false,
                    is_pk: false,
                    is_readonly: false,
                    rewrites: vec![],
                    tuple_members: Some(vec![
                        TupleMemberDescriptor {
                            name: Some("street".into()),
                            kind: TupleMemberKind::Scalar { pg_type: "text".into() },
                        },
                        TupleMemberDescriptor {
                            name: Some("zip".into()),
                            kind: TupleMemberKind::Scalar { pg_type: "text".into() },
                        },
                    ]),
                    column_type: None,
                }],
                links: vec![],
                multilinks: vec![],
                computed: vec![],
                constraints: vec![],
                indexes: vec![],
                partition: None,
                vector_indexes: vec![],
                search_indexes: vec![],
                triggers: vec![],
                junction: false,
                signals: vec![],
            }],
            scalars: vec![],
            enums: vec![],
            named_tuples: vec![],
            globals: vec![],
            functions: vec![],
            aliases: vec![],
            channels: vec![],
            ..Default::default()
        };
        let out = compile_and_emit_with("SELECT Person { address }", &schema);
        assert!(out.sql.contains("::jsonb"), "got:\n{}", out.sql);
        match &out.shape.root {
            crate::query::ShapeNode::Object { pointers, .. } => {
                let address = pointers
                    .iter()
                    .find(|p| matches!(p, crate::query::ShapeNode::NamedTuple { name, .. } if name == "address"))
                    .expect("expected address pointer in shape");
                match address {
                    crate::query::ShapeNode::NamedTuple { type_name, members, .. } => {
                        assert_eq!(*type_name, None);
                        let members = members.as_ref().expect("expected resolved members");
                        assert_eq!(members.len(), 2);
                        assert_eq!(members[0].key.as_deref(), Some("street"));
                        assert_eq!(members[1].key.as_deref(), Some("zip"));
                    }
                    other => panic!("expected NamedTuple, got {other:?}"),
                }
            }
            other => panic!("expected ShapeNode::Object, got {other:?}"),
        }
    }

    #[test]
    fn test_bare_path_select_structural_tuple_property_shape_carries_real_members() {
        // A bare `select Type.property` path traversal (no `{ }` shape) used
        // to fall back to `ShapeNode::Scalar` for a *structural* tuple
        // property — `emit_path_select`'s IrPathResult::Scalar branch only
        // detected a *nominal* named tuple (via the "__nt__:" pg_type
        // marker), never consulting the property's own `tuple_members`. This
        // decoded every row as an opaque blob instead of a proper tuple
        // literal (reported as `select default::Person.address` rendering
        // wrong in both pylon-ui and the CLI REPL, while a `{ address }`
        // shape query — the test above — already worked).
        use crate::schema::{TupleMemberDescriptor, TupleMemberKind};
        let schema = SchemaDescriptor {
            types: vec![TypeDescriptor {
                name: "Person".into(),
                module: "default".into(),
                table: "Person".into(),
                abstract_: false,
                materialized: false,
                description: None,
                parents: vec![],
                interfaces: vec![],
                bases: vec![],
                properties: vec![PropertyDescriptor {
                    name: "address".into(),
                    pg_type: "jsonb".into(),
                    nullable: true,
                    default_sql: None,
                    default_pyql: None,
                    description: None,
                    check_constraints: vec![],
                    is_exclusive: false,
                    is_pk: false,
                    is_readonly: false,
                    rewrites: vec![],
                    tuple_members: Some(vec![
                        TupleMemberDescriptor {
                            name: Some("street".into()),
                            kind: TupleMemberKind::Scalar { pg_type: "text".into() },
                        },
                        TupleMemberDescriptor {
                            name: Some("zip".into()),
                            kind: TupleMemberKind::Scalar { pg_type: "text".into() },
                        },
                    ]),
                    column_type: None,
                }],
                links: vec![],
                multilinks: vec![],
                computed: vec![],
                constraints: vec![],
                indexes: vec![],
                partition: None,
                vector_indexes: vec![],
                search_indexes: vec![],
                triggers: vec![],
                junction: false,
                signals: vec![],
            }],
            scalars: vec![],
            enums: vec![],
            named_tuples: vec![],
            globals: vec![],
            functions: vec![],
            aliases: vec![],
            channels: vec![],
            ..Default::default()
        };
        let out = compile_and_emit_with("SELECT Person.address", &schema);
        match &out.shape.root {
            crate::query::ShapeNode::NamedTuple { type_name, members, .. } => {
                assert_eq!(*type_name, None);
                let members = members.as_ref().expect("expected resolved members");
                assert_eq!(members.len(), 2);
                assert_eq!(members[0].key.as_deref(), Some("street"));
                assert_eq!(members[1].key.as_deref(), Some("zip"));
            }
            other => panic!("expected ShapeNode::NamedTuple, got {other:?}"),
        }
    }

    #[test]
    fn test_path_traversal_into_structural_tuple_property() {
        use crate::schema::{TupleMemberDescriptor, TupleMemberKind};
        let schema = SchemaDescriptor {
            types: vec![TypeDescriptor {
                name: "Person".into(),
                module: "default".into(),
                table: "Person".into(),
                abstract_: false,
                materialized: false,
                description: None,
                parents: vec![],
                interfaces: vec![],
                bases: vec![],
                properties: vec![PropertyDescriptor {
                    name: "address".into(),
                    pg_type: "jsonb".into(),
                    nullable: true,
                    default_sql: None,
                    default_pyql: None,
                    description: None,
                    check_constraints: vec![],
                    is_exclusive: false,
                    is_pk: false,
                    is_readonly: false,
                    rewrites: vec![],
                    tuple_members: Some(vec![
                        TupleMemberDescriptor {
                            name: Some("street".into()),
                            kind: TupleMemberKind::Scalar { pg_type: "text".into() },
                        },
                        TupleMemberDescriptor {
                            name: Some("zip".into()),
                            kind: TupleMemberKind::Scalar { pg_type: "text".into() },
                        },
                    ]),
                    column_type: None,
                }],
                links: vec![],
                multilinks: vec![],
                computed: vec![],
                constraints: vec![],
                indexes: vec![],
                partition: None,
                vector_indexes: vec![],
                search_indexes: vec![],
                triggers: vec![],
                junction: false,
                signals: vec![],
            }],
            scalars: vec![],
            enums: vec![],
            named_tuples: vec![],
            globals: vec![],
            functions: vec![],
            aliases: vec![],
            channels: vec![],
            ..Default::default()
        };
        // Regression: a structural (unnamed) tuple property previously failed
        // path traversal with "'address' is a scalar property, not a link —
        // cannot traverse further" because compile_path_select only allowed
        // further traversal for the nominal `__nt__:` named-tuple marker,
        // ignoring `tuple_members` on plain jsonb-typed properties.
        let out = compile_and_emit_with("SELECT default::Person.address.street", &schema);
        assert!(out.sql.contains("\"address\"->'street'"), "got:\n{}", out.sql);
    }

    #[test]
    fn test_structural_tuple_cast_shape_carries_real_members() {
        let out = compile_and_emit("SELECT <tuple<street: str, zip: str>>$p");
        match &out.shape.root {
            crate::query::ShapeNode::NamedTuple { type_name, members, .. } => {
                assert_eq!(*type_name, None);
                let members = members.as_ref().expect("expected resolved members");
                assert_eq!(members.len(), 2);
                assert_eq!(members[0].key.as_deref(), Some("street"));
                assert_eq!(members[1].key.as_deref(), Some("zip"));
            }
            other => panic!("expected ShapeNode::NamedTuple, got {other:?}"),
        }
    }

    #[test]
    fn test_tuple_cast_mixed_named_and_unnamed_elements_rejected() {
        match parse::parse("SELECT <tuple<x: float64, bool>>$p") {
            Ok(_) => panic!("expected parse error for mixed named/unnamed tuple elements"),
            Err(e) => assert!(e.to_string().contains("all named or all unnamed"), "got: {}", e),
        }
    }

    #[test]
    fn test_is_with_tuple_type_rejected() {
        let schema = make_schema();
        let ast = parse::parse("SELECT Person FILTER Person is tuple<x: float64, y: float64>").unwrap();
        match ir::compile(&ast, &schema) {
            Ok(_) => panic!("expected error for IS with a tuple type"),
            Err(e) => {
                assert!(
                    e.to_string().contains("cannot use IS with a tuple or array type"),
                    "got: {}",
                    e
                );
            }
        }
    }

    #[test]
    fn test_is_with_array_type_rejected() {
        let schema = make_schema();
        let ast = parse::parse("SELECT Person FILTER Person is array<str>").unwrap();
        match ir::compile(&ast, &schema) {
            Ok(_) => panic!("expected error for IS with an array type"),
            Err(e) => {
                assert!(
                    e.to_string().contains("cannot use IS with a tuple or array type"),
                    "got: {}",
                    e
                );
            }
        }
    }

    #[test]
    fn test_tuple_index_on_non_literal_falls_back_to_jsonb_index() {
        // `.1` only constant-folds against a literal tuple; anything else
        // (a $param, a cast result, …) needs a genuine runtime jsonb index
        // instead of erroring "only supported on tuple literals".
        let out = compile_and_emit("SELECT (<tuple<int64, str>>('1', 3)).1");
        assert!(out.sql.contains("->1"), "got:\n{}", out.sql);
    }

    #[test]
    fn test_tuple_index_out_of_bounds_on_cast_target_errors_at_compile_time() {
        // Regression: `.2` on a 2-element tuple cast silently fell back to a
        // runtime jsonb index (returning null) instead of failing, because
        // the source is a TypeCast, not a literal Tuple/NamedTuple. The cast's
        // target type is statically known here, so this must bounds-check
        // and error instead.
        let ast = parse::parse("SELECT (<tuple<int64, str>>('1', 3)).2").unwrap();
        let schema = make_schema();
        match ir::compile(&ast, &schema) {
            Ok(_) => panic!("expected out-of-bounds tuple index error"),
            Err(e) => {
                assert!(
                    e.to_string()
                        .contains("2 is not a member of tuple<std::int64, std::str>"),
                    "got: {}",
                    e
                );
            }
        }
    }

    #[test]
    fn test_positional_tuple_literal_cast_to_tuple_type_compiles() {
        // Each element must be cast to its own declared type — '1' isn't
        // silently jsonb-wrapped unchanged as a string; it's coerced to
        // int8.
        let out = compile_and_emit("SELECT <tuple<int64, str>>(1, 'x')");
        assert!(
            out.sql.contains("jsonb_build_array((1)::int8, ('x')::text)"),
            "got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_positional_tuple_literal_cast_coerces_mismatched_literal_types() {
        // The exact case that was silently wrong: casting a string literal to
        // int64 and an int literal to str must actually coerce each one, not
        // just pass the raw literal through untouched.
        let out = compile_and_emit("SELECT <tuple<int64, str>>('1', 3)");
        assert!(
            out.sql.contains("jsonb_build_array(('1')::int8, (3)::text)"),
            "got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_nested_tuple_literal_cast_applies_casts_recursively() {
        let out = compile_and_emit(
            "SELECT <tuple<point: tuple<x: float64, y: float64>, label: str>>(point := ('1', 2), label := 5)",
        );
        assert!(
            out.sql.contains(
                "jsonb_build_object('point', jsonb_build_object('x', ('1')::float8, 'y', (2)::float8), 'label', (5)::text)"
            ),
            "got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_positional_tuple_literal_nested_inside_named_tuple_compiles() {
        let out = compile_and_emit("SELECT (point := (1, 2), label := 'origin')");
        assert!(out.sql.contains("jsonb_build_array(1, 2)"), "got:\n{}", out.sql);
        assert!(out.sql.contains("jsonb_build_object("), "got:\n{}", out.sql);
    }

    #[test]
    fn test_positional_tuple_literal_in_schema_bound_shape_field_compiles() {
        let out = compile_and_emit("SELECT Person { name, pair := (1, 2) }");
        assert!(out.sql.contains("jsonb_build_array(1, 2)"), "got:\n{}", out.sql);
    }

    // ── search index outbox enqueue tests ─────────────────────────────────────
    //
    // Regression coverage for a real bug: `collect_search_enqueue` used to
    // filter for `SearchBackend::OpenSearch` only, so a Meilisearch-backed
    // SearchIndex never got an outbox row on insert/update/delete — the
    // Meilisearch worker (Rust or the old Python one) never received real
    // traffic from normal DML, only from a manually-inserted outbox row.

    fn make_schema_with_search_index(backend: crate::schema::SearchBackend) -> SchemaDescriptor {
        use crate::schema::{SearchIndexDescriptor, SearchPointerDescriptor, SearchWeight};
        let mut s = make_schema();
        if let Some(td) = s.types.iter_mut().find(|t| t.name == "Person") {
            td.search_indexes.push(SearchIndexDescriptor {
                index_name: None,
                backend,
                pointers: vec![SearchPointerDescriptor {
                    name: "name".into(),
                    weight: SearchWeight::A,
                }],
            });
        }
        s
    }

    #[test]
    fn test_insert_enqueues_a_meilisearch_outbox_row() {
        let schema = make_schema_with_search_index(crate::schema::SearchBackend::Meilisearch);
        let out = compile_and_emit_with("INSERT Person { name := 'Alice', age := 30 }", &schema);
        assert!(
            out.sql.contains("'Meilisearch'::_pylon.\"IndexKind\""),
            "expected a Meilisearch outbox enqueue CTE, got:\n{}",
            out.sql,
        );
        assert!(
            out.sql.contains("INSERT INTO _pylon.\"IndexOutbox\""),
            "got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_insert_enqueues_an_opensearch_outbox_row() {
        let schema = make_schema_with_search_index(crate::schema::SearchBackend::OpenSearch);
        let out = compile_and_emit_with("INSERT Person { name := 'Alice', age := 30 }", &schema);
        assert!(
            out.sql.contains("'OpenSearch'::_pylon.\"IndexKind\""),
            "expected an OpenSearch outbox enqueue CTE, got:\n{}",
            out.sql,
        );
    }

    #[test]
    fn test_insert_does_not_enqueue_an_outbox_row_for_a_postgres_backed_search_index() {
        // Postgres-backed search indexes are maintained synchronously by a
        // trigger-updated tsvector column — no async worker involved.
        let schema = make_schema_with_search_index(crate::schema::SearchBackend::Postgres);
        let out = compile_and_emit_with("INSERT Person { name := 'Alice', age := 30 }", &schema);
        assert!(
            !out.sql.contains("_pylon.\"IndexOutbox\""),
            "did not expect an outbox enqueue, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_update_enqueues_a_meilisearch_outbox_row() {
        let schema = make_schema_with_search_index(crate::schema::SearchBackend::Meilisearch);
        let out = compile_and_emit_with("UPDATE Person FILTER .name = 'Alice' SET { age := 31 }", &schema);
        assert!(
            out.sql.contains("'Meilisearch'::_pylon.\"IndexKind\""),
            "expected a Meilisearch outbox enqueue CTE, got:\n{}",
            out.sql,
        );
    }

    #[test]
    fn test_delete_enqueues_a_meilisearch_outbox_delete_job() {
        let schema = make_schema_with_search_index(crate::schema::SearchBackend::Meilisearch);
        let out = compile_and_emit_with("DELETE Person FILTER .name = 'Alice'", &schema);
        assert!(
            out.sql.contains("'Meilisearch'::_pylon.\"IndexKind\""),
            "expected a Meilisearch outbox enqueue CTE, got:\n{}",
            out.sql,
        );
        assert!(
            out.sql.contains("'delete'"),
            "expected the delete operation literal, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_range_intrinsic_resolves_int_literals_to_int8range() {
        // Regression: std::range's ImplStrategy::TranspilerIntrinsic had no
        // actual substitution anywhere — it fell through resolve_fn_call's
        // catch-all, which emitted a literal (nonexistent) `"std"."range"(...)`
        // call instead of a real PostgreSQL range constructor.
        let out = compile_and_emit("SELECT std::overlaps(std::range(1, 3), std::range(2, 5))");
        assert!(out.sql.contains("int8range(1, 3)"), "got:\n{}", out.sql);
        assert!(out.sql.contains("int8range(2, 5)"), "got:\n{}", out.sql);
        assert!(
            out.sql.contains(" && "),
            "expected infix && for overlaps, got:\n{}",
            out.sql
        );
        assert!(
            !out.sql.contains("\"std\""),
            "must not emit a literal std schema call, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_range_intrinsic_resolves_datetime_to_tstzrange() {
        let out =
            compile_and_emit("SELECT std::range(<datetime>'2024-01-01T00:00:00Z', <datetime>'2024-06-01T00:00:00Z')");
        assert!(out.sql.contains("tstzrange("), "got:\n{}", out.sql);
    }

    #[test]
    fn test_range_intrinsic_four_arg_form_computes_bounds_string() {
        let out = compile_and_emit("SELECT std::range(1, 3, true, false)");
        assert!(out.sql.contains("int8range(1, 3,"), "got:\n{}", out.sql);
        assert!(
            out.sql.contains("CASE WHEN"),
            "expected a dynamic bounds-string CASE, got:\n{}",
            out.sql
        );
    }

    #[test]
    fn test_multirange_intrinsic_resolves_from_range_element() {
        let out = compile_and_emit("SELECT std::multirange([std::range(1, 3), std::range(5, 7)])");
        assert!(out.sql.contains("int8multirange(VARIADIC "), "got:\n{}", out.sql);
        assert!(
            !out.sql.contains("\"std\""),
            "must not emit a literal std schema call, got:\n{}",
            out.sql
        );
    }
}