core_query/cypher/

exec.rs

//! Cypher executor: `PlanOp` sequence → `ResultSet` over a binding table.

use crate::cypher::ast::{
    ret_val_label, AggArg, AggFunc, Expr, LimitSkip, Operand, OrderItem, OrderTarget, RetItem,
    RetVal, UnwindExpr,
};
use crate::cypher::plan::PlanOp;
use crate::cypher::RelDir;
use crate::filter::eval_cmp;
use crate::result::ResultSet;
use crate::traverse::{expand, Dir, EdgeRef};
use crate::value_ops::{cmp_optional, values_equal};
use crate::view::GraphView;
use core_storage::{Value, ValueKey};
use std::collections::{BTreeMap, BTreeSet, HashMap, HashSet};

/// Test-only counter incremented each time the fused ScanLabel+Filter arm
/// executes.  Lets property tests assert the fast path actually fires for
/// matching query shapes (and does NOT fire for fallback shapes).
#[cfg(test)]
static FUSED_SCAN_FIRES: std::sync::atomic::AtomicUsize = std::sync::atomic::AtomicUsize::new(0);

/// Test-only counter incremented each time a `ScanKey` arm executes.
/// Lets tests assert the point-lookup path actually fires (and does not
/// walk `view.labels`).
#[cfg(test)]
static SCAN_KEY_FIRES: std::sync::atomic::AtomicUsize = std::sync::atomic::AtomicUsize::new(0);

/// Test-only counter incremented each time an `IndexScan` takes the *indexed*
/// path (not the scan fallback). Lets tests assert the equality index is used.
#[cfg(test)]
pub static INDEX_SCAN_FIRES: std::sync::atomic::AtomicUsize =
    std::sync::atomic::AtomicUsize::new(0);

/// Test-only counter incremented each time an `IndexIntersect` takes the
/// indexed path (at least one field resolved via `nodes_with_prop`). Does not
/// advance on the all-unindexed full-scan fallback.
#[cfg(test)]
pub static INDEX_INTERSECT_FIRES: std::sync::atomic::AtomicUsize =
    std::sync::atomic::AtomicUsize::new(0);

/// Query parameters. Missing names anywhere in the plan are an error at
/// execution start (the plan is walked before any rows are produced).
pub struct Params<'a>(pub &'a BTreeMap<String, Value>);

#[derive(Clone, Debug)]
enum Cell {
    Node(u32),
    Rel(EdgeRef),
    /// Virtual path cell produced by `VarExpand` / `ShortestPath`.
    /// The only accessible property is `length` → `Value::Int(hops)`.
    Path(u8),
    /// Scalar value produced by a WITH projection (property alias, aggregate result, etc.).
    Scalar(Value),
}

/// Binding-table row: one slot per interned variable. Cheaper to clone than
/// `BTreeMap<String, Cell>` on every Expand/Scan (the two-hop hot path).
type Row = Vec<Option<Cell>>;

struct VarTable {
    names: Vec<String>,
}

impl VarTable {
    fn intern(&mut self, name: &str) -> usize {
        if let Some(i) = self.names.iter().position(|n| n == name) {
            return i;
        }
        self.names.push(name.to_string());
        self.names.len() - 1
    }

    fn slot(&self, name: &str) -> Option<usize> {
        self.names.iter().position(|n| n == name)
    }
}

struct Projected {
    columns: Vec<String>,
    rows: Vec<Vec<Option<Value>>>,
}

/// Production cap on the executor binding table after each `scan_label` /
/// `expand`. Unjoined multi-MATCH cross-joins OOM without this.
const MAX_INTERMEDIATE_ROWS: usize = 1_000_000;

/// Cap on the number of distinct groups a `GroupAggregate` plan may produce.
const MAX_GROUPS: usize = 1_000_000;

/// Group-key type: one `Option<ValueKey>` per group-key RETURN item.
/// `None` represents a null value; null keys group together (openCypher).
type GroupKey = Vec<Option<ValueKey>>;
/// Per-group entry: first-seen display values for key columns, plus accumulators.
/// Display values are the original `Value`s before normalization so that
/// `Int(42)` groups still display as `Int(42)` even though they are hashed as
/// `FloatBits`.  For mixed Int/Float groups the first-seen value wins.
type GroupEntry = (Vec<Option<Value>>, Vec<AggAcc>);

#[cfg(test)]
thread_local! {
    static TEST_MAX_INTERMEDIATE_ROWS: std::cell::Cell<Option<usize>> =
        const { std::cell::Cell::new(None) };
    /// Accumulator for rows emitted by `exec_expand` during a bounded test run.
    /// `None` means the counter is inactive (no test is watching).
    static TEST_EXPAND_PRODUCED: std::cell::Cell<Option<usize>> =
        const { std::cell::Cell::new(None) };
    /// Override for `MAX_GROUPS` used in tests.
    static TEST_MAX_GROUPS: std::cell::Cell<Option<usize>> =
        const { std::cell::Cell::new(None) };
}

fn max_intermediate_rows() -> usize {
    #[cfg(test)]
    {
        TEST_MAX_INTERMEDIATE_ROWS
            .with(|c| c.get())
            .unwrap_or(MAX_INTERMEDIATE_ROWS)
    }
    #[cfg(not(test))]
    {
        MAX_INTERMEDIATE_ROWS
    }
}

fn max_groups() -> usize {
    #[cfg(test)]
    {
        TEST_MAX_GROUPS.with(|c| c.get()).unwrap_or(MAX_GROUPS)
    }
    #[cfg(not(test))]
    {
        MAX_GROUPS
    }
}

/// Test hook: run `f` with a smaller group cap so the group-count error path
/// can fire without creating a million groups.
#[cfg(test)]
pub(crate) fn with_max_groups<R>(cap: usize, f: impl FnOnce() -> R) -> R {
    TEST_MAX_GROUPS.with(|c| {
        let prev = c.replace(Some(cap));
        let result = std::panic::catch_unwind(std::panic::AssertUnwindSafe(f));
        c.set(prev);
        match result {
            Ok(v) => v,
            Err(p) => std::panic::resume_unwind(p),
        }
    })
}

/// Test hook: run `f` with a smaller intermediate-row cap so the error path
/// can fire without allocating a million rows. Restores the previous override
/// (including across panics).
#[cfg(test)]
pub(crate) fn with_max_intermediate_rows<R>(cap: usize, f: impl FnOnce() -> R) -> R {
    TEST_MAX_INTERMEDIATE_ROWS.with(|c| {
        let prev = c.replace(Some(cap));
        let result = std::panic::catch_unwind(std::panic::AssertUnwindSafe(f));
        c.set(prev);
        match result {
            Ok(v) => v,
            Err(p) => std::panic::resume_unwind(p),
        }
    })
}

/// Increment the test-only expand-row counter (no-op when counter is inactive).
#[cfg(test)]
fn record_expand_row() {
    TEST_EXPAND_PRODUCED.with(|c| {
        if let Some(prev) = c.get() {
            c.set(Some(prev + 1));
        }
    });
}

/// Run `f` while counting every row emitted by `exec_expand`.
/// Returns `(result_of_f, total_rows_produced)`.
///
/// Used by tests to assert early-termination: bounded execution must emit
/// ≤ `row_bound + ε` rows, while an equivalent unbounded run emits ≥ 100×.
#[cfg(test)]
pub(crate) fn with_expand_counter<R>(f: impl FnOnce() -> R) -> (R, usize) {
    TEST_EXPAND_PRODUCED.with(|c| c.set(Some(0)));
    let result = std::panic::catch_unwind(std::panic::AssertUnwindSafe(f));
    let count = TEST_EXPAND_PRODUCED.with(|c| c.get().unwrap_or(0));
    TEST_EXPAND_PRODUCED.with(|c| c.set(None));
    match result {
        Ok(v) => (v, count),
        Err(p) => std::panic::resume_unwind(p),
    }
}

fn row_cap_err(cap: usize) -> String {
    format!(
        "intermediate result exceeds {cap} rows; add a LIMIT or constrain patterns with shared variables"
    )
}

fn group_cap_err() -> String {
    format!(
        "group count exceeds {} distinct keys; add a WHERE clause or constrain the grouping key",
        max_groups()
    )
}

/// Convert a group-key value to a `ValueKey` with numeric unification.
///
/// `Int(n)` and `Float(n as f64)` produce the same `FloatBits` key so that
/// nodes with `score = 1` (Int) and `score = 1.0` (Float) land in the same
/// group, matching openCypher's `1 = 1.0` equality rule.
///
/// Note: integers whose magnitude exceeds 2^53 lose precision when cast to
/// `f64`, so two large integers that differ only beyond the float mantissa
/// width could be incorrectly unified. This is a known limitation documented
/// in `docs/site/query.md`.
fn group_key_normalize(v: &Value) -> Option<ValueKey> {
    match v {
        Value::Int(n) => Some(ValueKey::FloatBits((*n as f64).to_bits())),
        Value::Float(f) => Some(ValueKey::FloatBits(f.to_bits())),
        _ => ValueKey::from_value(v),
    }
}

/// Execute a plan against a view. Row order before OrderBy is deterministic
/// (scan order = dense ids; expand order = expand()'s sorted order).
///
/// Precondition: `OrderBy` items produced by `plan()` use `OrderTarget::Alias`
/// only; other variants are accepted defensively but non-standard.
///
/// When the plan has a `Limit` and no `OrderBy`, the executor switches to a
/// demand-driven (pull-based) strategy: all producer stages terminate as soon
/// as `SKIP + LIMIT` final rows have been collected.  No intermediate table is
/// ever fully materialised for the bounded path — the 1 M cap is the safety
/// net for unbounded queries only.
pub fn execute(view: &GraphView, plan: &[PlanOp], params: &Params) -> Result<ResultSet, String> {
    use crate::cypher::plan::row_bound;
    execute_inner(view, plan, params, row_bound(plan))
}

/// Execute a read query that may be a `UNION` / `UNION ALL` chain. Each part is
/// planned and executed against the same `view` (so masking applies uniformly),
/// then combined left-to-right: `UNION` dedups the accumulated rows, `UNION ALL`
/// concatenates. All parts must project the same column names.
pub fn execute_union(
    view: &GraphView,
    union: &crate::cypher::parser::UnionQuery,
    params: &Params,
) -> Result<ResultSet, String> {
    let mut acc: Option<ResultSet> = None;
    for (i, part) in union.parts.iter().enumerate() {
        let ops = crate::cypher::plan::plan(part)?;
        let rs = execute(view, &ops, params)?;
        acc = Some(match acc {
            None => rs,
            Some(prev) => {
                if prev.columns() != rs.columns() {
                    return Err(format!(
                        "UNION requires matching column names across all parts; \
                         got {:?} then {:?}",
                        prev.columns(),
                        rs.columns()
                    ));
                }
                // all_flags has one entry per boundary; index i-1 for part i.
                let all = union.all_flags.get(i - 1).copied().unwrap_or(false);
                combine_result_sets(prev, rs, all)
            }
        });
    }
    // `parse_read` guarantees at least one part.
    Ok(acc.expect("UNION query has at least one part"))
}

fn combine_result_sets(mut acc: ResultSet, other: ResultSet, all: bool) -> ResultSet {
    for i in 0..other.len() {
        acc.push_row(other.row(i).to_vec());
    }
    if !all {
        // UNION (distinct): keep first occurrence of each row.
        let mut seen: Vec<Vec<Option<Value>>> = Vec::new();
        for i in 0..acc.len() {
            let r = acc.row(i).to_vec();
            if !seen.contains(&r) {
                seen.push(r);
            }
        }
        let mut deduped = ResultSet::new(acc.columns().to_vec());
        for r in seen {
            deduped.push_row(r);
        }
        return deduped;
    }
    acc
}

/// Like `execute` but always disables LIMIT push-down (row_bound = None).
///
/// Used in tests as the reference implementation: the result must equal that of
/// `execute` (which applies push-down) modulo the subset selected by SKIP/LIMIT.
#[cfg(test)]
pub(crate) fn execute_unbounded(
    view: &GraphView,
    plan: &[PlanOp],
    params: &Params,
) -> Result<ResultSet, String> {
    execute_inner(view, plan, params, None)
}

fn execute_inner(
    view: &GraphView,
    plan: &[PlanOp],
    params: &Params,
    row_bound: Option<usize>,
) -> Result<ResultSet, String> {
    check_params(plan, params)?;

    // VarExpand / ShortestPath plans always take the staged path, even when
    // an Aggregate op is also present.  row_bound() already returns None for
    // these plans (so the pull path is never chosen), but the aggregate path
    // check below must also be skipped so that VarExpand+Aggregate falls
    // through to the staged path where both ops are handled.
    let has_var_expand = plan
        .iter()
        .any(|op| matches!(op, PlanOp::VarExpand { .. } | PlanOp::ShortestPath { .. }));

    // Pipeline plans (WITH / UNWIND / LeftOuterApply) always use the staged
    // path so that intermediate rows are correctly sequenced.
    //
    // So does a `GroupAggregate` followed by a `Filter` or a `Project`. Both
    // mean an aggregate `WITH` with something after it — a HAVING clause, or
    // a RETURN that projects the groups — and both have to see the
    // `Cell::Scalar`/`Cell::Node` rows `group_result_to_rows` produces. The
    // streaming `execute_group_aggregate` path emits the group table itself
    // and ignores every op after the aggregate except ORDER BY / SKIP /
    // LIMIT, so a `Project` there was silently dropped: `WITH c, count(t) AS
    // n RETURN c.name, n * 2 AS dbl` came back as columns `c` and `n`
    // carrying the group key and the raw count.
    //
    // A plain aggregate with no `WITH` (`RETURN c.name, count(*)`) has its
    // columns in the `GroupAggregate` op itself and no `Project` after it, so
    // it keeps the streaming path.
    let is_pipeline = plan.iter().any(|op| {
        matches!(
            op,
            PlanOp::With { .. } | PlanOp::Unwind { .. } | PlanOp::LeftOuterApply { .. }
        )
    }) || {
        let mut saw_gagg = false;
        plan.iter().any(|op| {
            if matches!(op, PlanOp::GroupAggregate { .. }) {
                saw_gagg = true;
                false
            } else {
                saw_gagg && matches!(op, PlanOp::Filter { .. } | PlanOp::Project { .. })
            }
        })
    };

    // GroupAggregate plans: streaming HashMap accumulator (O(groups) memory).
    // Checked before the bounded path and before single-aggregate so they are
    // never routed to pull.  Skipped for VarExpand and pipeline plans.
    if !has_var_expand
        && !is_pipeline
        && plan
            .iter()
            .any(|op| matches!(op, PlanOp::GroupAggregate { .. }))
    {
        return execute_group_aggregate(view, plan, params);
    }

    // Aggregate plans: streaming accumulator path (O(1) memory, no budget).
    // Checked before the bounded path so aggregates are never routed to pull.
    // Skipped for VarExpand and pipeline plans — they go to the staged path.
    if !has_var_expand
        && !is_pipeline
        && plan.iter().any(|op| matches!(op, PlanOp::Aggregate { .. }))
    {
        return execute_aggregate(view, plan, params);
    }

    // Bounded plans use the pull-based (demand-driven) executor so that ALL
    // producer stages terminate as soon as `bound` final rows are collected.
    // VarExpand plans have row_bound=None, so this branch never fires for them.
    if let Some(bound) = row_bound {
        return execute_pull(view, plan, params, bound);
    }

    // Staged (unbounded) path — full materialisation with the 1 M safety cap.
    let vars = collect_vars(plan);
    let mut rows: Vec<Row> = vec![vec![None; vars.names.len()]];
    let mut projected: Option<Projected> = None;
    // Tracks column names produced by a pipeline GroupAggregate so the staged
    // executor can emit the final ResultSet from `rows` when no Project follows.
    let mut pipeline_group_columns: Option<Vec<String>> = None;

    for op in plan {
        match op {
            PlanOp::ScanLabel { var, label } => {
                rows = scan_label(view, &vars, &rows, var, label.as_deref())?;
            }
            PlanOp::ScanKey { var, key, label } => {
                rows = scan_key(view, &vars, &rows, var, key, label.as_deref(), params)?;
            }
            PlanOp::IndexScan {
                var,
                label,
                field,
                value,
            } => {
                rows = scan_index(
                    view,
                    &vars,
                    &rows,
                    var,
                    label.as_deref(),
                    field,
                    value,
                    params,
                )?;
            }
            PlanOp::IndexIntersect {
                var,
                label,
                equalities,
            } => {
                rows = scan_intersect(
                    view,
                    &vars,
                    &rows,
                    var,
                    label.as_deref(),
                    equalities,
                    params,
                )?;
            }
            PlanOp::LookupProps { var, props } => {
                rows = retain_node(view, &vars, &rows, var, None, props, params)?;
            }
            PlanOp::JoinBound { var, label, props } => {
                rows = retain_node(view, &vars, &rows, var, label.as_deref(), props, params)?;
            }
            PlanOp::Expand { .. } => {
                rows = exec_expand(view, &vars, &rows, op, params)?;
            }
            PlanOp::VarExpand {
                from,
                rel_var,
                etypes,
                dir,
                to,
                min,
                max,
            } => {
                rows = exec_var_expand(
                    view, &vars, &rows, from, rel_var, etypes, *dir, to, *min, *max,
                )?;
            }
            PlanOp::ShortestPath {
                from,
                rel_var,
                etypes,
                dir,
                to,
                max_hops,
            } => {
                rows = exec_shortest_path(
                    view, &vars, &rows, from, rel_var, etypes, *dir, to, *max_hops,
                )?;
            }
            PlanOp::Filter { expr } => {
                rows = exec_filter(view, &vars, &rows, expr, params)?;
            }
            PlanOp::Project { items } => {
                projected = Some(exec_project(view, &vars, &rows, items, params)?);
            }
            PlanOp::Distinct => {
                if let Some(table) = projected.as_mut() {
                    exec_distinct(table)?;
                } else {
                    return Err("DISTINCT requires a Project".into());
                }
            }
            PlanOp::OrderBy { items } => {
                if let Some(table) = projected.as_mut() {
                    exec_order_by(table, items)?;
                } else {
                    // Pipeline mode: ORDER BY on raw rows (before any Project).
                    exec_order_by_rows(&vars, &mut rows, items, view);
                }
            }
            PlanOp::Skip(ls) => {
                let n = resolve_ls(ls, params)?;
                if let Some(table) = projected.as_mut() {
                    apply_skip(&mut table.rows, n);
                } else {
                    apply_skip(&mut rows, n);
                }
            }
            PlanOp::Limit(ls) => {
                let n = resolve_ls(ls, params)?;
                if let Some(table) = projected.as_mut() {
                    apply_limit(&mut table.rows, n);
                } else {
                    apply_limit(&mut rows, n);
                }
            }
            // GroupAggregate plans without VarExpand or pipeline are routed to
            // execute_group_aggregate() before reaching the staged path.
            // Plans that combine VarExpand with GroupAggregate, or pipeline plans
            // with an intermediate GroupAggregate, fall through to here.
            PlanOp::GroupAggregate { keys, aggs } => {
                let mut grp_groups: HashMap<GroupKey, GroupEntry> = HashMap::new();
                let mut grp_key_order: Vec<GroupKey> = Vec::new();
                // A grouping key that is a bare node variable stays a node
                // downstream: `WITH c, count(*) AS n WHERE n >= 2 RETURN c.key`
                // reads `c` as a node after the group, not as the key string a
                // display value would flatten it to. Without this the whole
                // pipeline after an aggregate loses node identity, and both
                // `c.key` and `key(c)` stop working.
                let mut grp_cells: HashMap<GroupKey, Vec<Option<Cell>>> = HashMap::new();
                for row in &rows {
                    let mut gk: GroupKey = Vec::with_capacity(keys.len());
                    let mut display_vals: Vec<Option<Value>> = Vec::with_capacity(keys.len());
                    for (_, item) in keys {
                        let val = project_item(view, &vars, row, item, params)?;
                        gk.push(val.as_ref().and_then(group_key_normalize));
                        display_vals.push(val);
                    }
                    if !grp_groups.contains_key(&gk) {
                        if grp_groups.len() >= max_groups() {
                            return Err(group_cap_err());
                        }
                        grp_key_order.push(gk.clone());
                        grp_cells.insert(gk.clone(), key_source_cells(&vars, row, keys));
                        grp_groups.insert(
                            gk.clone(),
                            (
                                display_vals,
                                aggs.iter().map(|(f, a, _)| AggAcc::for_arg(f, a)).collect(),
                            ),
                        );
                    }
                    let (_, accs) = grp_groups.get_mut(&gk).unwrap();
                    for (acc, (func, arg, _)) in accs.iter_mut().zip(aggs.iter()) {
                        update_acc(view, &vars, row, func, arg, acc)?;
                    }
                }
                // Same openCypher rule as execute_group_aggregate: no-key multi-agg on
                // empty input must emit exactly one row with zero/null accumulators.
                if keys.is_empty() && grp_key_order.is_empty() {
                    let empty_key: GroupKey = vec![];
                    grp_key_order.push(empty_key.clone());
                    grp_groups.insert(
                        empty_key,
                        (
                            vec![],
                            aggs.iter().map(|(f, a, _)| AggAcc::for_arg(f, a)).collect(),
                        ),
                    );
                }
                if is_pipeline {
                    // Pipeline mode: convert group results to raw rows with Cell::Scalar
                    // so that subsequent WITH / RETURN stages can consume them.
                    rows = group_result_to_rows(
                        keys,
                        aggs,
                        grp_key_order,
                        &mut grp_groups,
                        &mut grp_cells,
                        &vars,
                    );
                    projected = None;
                    // Record column order so the staged executor can build the final
                    // ResultSet from `rows` if no Project op follows.
                    let mut cols: Vec<String> = keys.iter().map(|(c, _)| c.clone()).collect();
                    cols.extend(aggs.iter().map(|(_, _, c)| c.clone()));
                    pipeline_group_columns = Some(cols);
                } else {
                    projected = Some(build_group_projected(
                        keys,
                        aggs,
                        grp_key_order,
                        &mut grp_groups,
                    ));
                    pipeline_group_columns = None;
                }
            }
            // Non-aggregate WITH: apply filter / order / skip / limit, then
            // project scalar aliases as Cell::Scalar while keeping node bindings.
            PlanOp::With {
                items,
                where_expr,
                order_by,
                skip,
                limit,
            } => {
                // Project each input row through the WITH items first so that
                // output aliases (e.g. `p.age AS age`) are available for the
                // subsequent WHERE filter, ORDER BY, SKIP, and LIMIT.
                let row_len = vars.names.len();
                let mut new_rows: Vec<Row> = Vec::with_capacity(rows.len());
                for row in &rows {
                    let mut new_row: Row = vec![None; row_len];
                    for item in items {
                        let col = column_name(item);
                        let Some(dst_slot) = vars.slot(&col) else {
                            continue;
                        };
                        match &item.value {
                            RetVal::Var(v) => {
                                // Carry the existing cell (node binding or scalar alias).
                                if let Some(src_slot) = vars.slot(v) {
                                    new_row[dst_slot] = row.get(src_slot).cloned().flatten();
                                }
                            }
                            RetVal::Prop { var, field } => {
                                let val = resolve_prop(view, &vars, row, var, field)?;
                                new_row[dst_slot] = val.map(Cell::Scalar);
                            }
                            RetVal::Agg { .. } => {} // aggregate WITH handled via GroupAggregate
                            RetVal::FuncCall { name, args } => {
                                let val = eval_func(name, args, view, &vars, row, params)?;
                                new_row[dst_slot] = val.map(Cell::Scalar);
                            }
                            RetVal::ScalarExpr(op) => {
                                let val = resolve_operand(view, &vars, row, op, params)?;
                                new_row[dst_slot] = val.map(Cell::Scalar);
                            }
                        }
                    }
                    new_rows.push(new_row);
                }
                rows = new_rows;
                // Apply WHERE (HAVING-like) filter on projected rows.
                if let Some(expr) = where_expr {
                    rows = exec_filter(view, &vars, &rows, expr, params)?;
                }
                // ORDER BY / SKIP / LIMIT on projected rows.
                if !order_by.is_empty() {
                    exec_order_by_rows(&vars, &mut rows, order_by, view);
                }
                if let Some(ls) = skip {
                    let n = resolve_ls(ls, params)?;
                    apply_skip(&mut rows, n);
                }
                if let Some(ls) = limit {
                    let n = resolve_ls(ls, params)?;
                    apply_limit(&mut rows, n);
                }
            }
            // UNWIND: expand each input row into N rows by iterating a list value.
            PlanOp::Unwind { expr, alias } => {
                let alias_slot = vars
                    .slot(alias)
                    .ok_or_else(|| format!("UNWIND alias `{alias}` not in VarTable"))?;
                let cap = max_intermediate_rows();
                let mut new_rows: Vec<Row> = Vec::new();
                for row in &rows {
                    let list_val: Option<Value> = match expr {
                        UnwindExpr::Lit(vals) => Some(Value::List(vals.clone())),
                        UnwindExpr::Prop { var, field } => {
                            resolve_prop(view, &vars, row, var, field)?
                        }
                        UnwindExpr::Var(name) => {
                            let slot = vars
                                .slot(name)
                                .ok_or_else(|| format!("UNWIND variable `{name}` is not bound"))?;
                            match row.get(slot).and_then(|c| c.as_ref()) {
                                Some(Cell::Scalar(v)) => Some(v.clone()),
                                Some(Cell::Node(_) | Cell::Rel(_) | Cell::Path(_)) => {
                                    return Err(format!(
                                        "UNWIND requires a list; `{name}` is bound to a graph element"
                                    ));
                                }
                                None => None,
                            }
                        }
                    };
                    match list_val {
                        None => {} // null → 0 rows (openCypher)
                        Some(Value::List(items_list)) => {
                            // Empty list → 0 rows (openCypher)
                            for item_val in items_list {
                                if new_rows.len() >= cap {
                                    return Err(row_cap_err(cap));
                                }
                                let mut new_row = row.clone();
                                new_row[alias_slot] = Some(Cell::Scalar(item_val));
                                new_rows.push(new_row);
                            }
                        }
                        Some(other) => {
                            let type_name = match &other {
                                Value::Int(_) => "Int",
                                Value::Float(_) => "Float",
                                Value::Str(_) => "Str",
                                Value::Bool(_) => "Bool",
                                Value::List(_) => unreachable!(),
                                Value::Map(_) => "Map",
                            };
                            return Err(format!(
                                "UNWIND requires a list; got {type_name} value for `{alias}`"
                            ));
                        }
                    }
                }
                rows = new_rows;
            }
            // Aggregate plans without VarExpand are routed to execute_aggregate()
            // before reaching the staged path.  Plans that combine VarExpand with
            // an Aggregate fall through to here; accumulate over the materialised
            // rows using the same agg_stream terminal logic (empty ops slice).
            PlanOp::Aggregate { func, arg, column } => {
                let ctx = AggStreamCtx {
                    view,
                    vars: &vars,
                    params,
                    func,
                    arg,
                };
                let mut acc = AggAcc::for_arg(func, arg);
                for row in &rows {
                    // agg_stream with empty ops hits the terminal branch (accumulate).
                    agg_stream(&ctx, &[], row, &mut acc)?;
                }
                let value = acc.finish();
                let mut rs = ResultSet::new(vec![column.clone()]);
                rs.push_row(vec![value]);
                return Ok(rs);
            }
            // OPTIONAL MATCH: left-outer-join apply.
            //
            // For each outer row, execute the inner plan in isolation.  If the
            // inner plan produces ≥1 row, those rows flow out.  If it produces
            // 0 rows, the outer row flows out with optional_vars nulled.
            PlanOp::LeftOuterApply {
                inner,
                optional_vars,
            } => {
                let cap = max_intermediate_rows();
                let mut new_rows: Vec<Row> = Vec::new();
                for outer_row in &rows {
                    // Seed the inner executor with just this one outer row.
                    let inner_seed: Vec<Row> = vec![outer_row.clone()];
                    // Execute the inner plan in "micro-staged" mode using the
                    // shared vars table (already contains all variable slots).
                    let inner_result =
                        exec_left_outer_inner(view, &vars, inner_seed, inner, params)?;
                    if inner_result.is_empty() {
                        // Left-outer fallback: null out optional vars.
                        let mut null_row = outer_row.clone();
                        for v in optional_vars {
                            if let Some(slot) = vars.slot(v) {
                                null_row[slot] = None;
                            }
                        }
                        if new_rows.len() >= cap {
                            return Err(row_cap_err(cap));
                        }
                        new_rows.push(null_row);
                    } else {
                        for r in inner_result {
                            if new_rows.len() >= cap {
                                return Err(row_cap_err(cap));
                            }
                            new_rows.push(r);
                        }
                    }
                }
                rows = new_rows;
            }
        }
    }

    Ok(match projected {
        Some(table) => finish(table),
        None => {
            // If a pipeline GroupAggregate ran and no Project followed, build the
            // final ResultSet from the rows that group_result_to_rows() produced.
            if let Some(cols) = pipeline_group_columns {
                let mut rs = ResultSet::new(cols.clone());
                for row in rows {
                    let vals: Vec<Option<Value>> = cols
                        .iter()
                        .map(|col| {
                            vars.slot(col)
                                .and_then(|s| row.get(s))
                                .and_then(|c| c.as_ref())
                                .and_then(|c| match c {
                                    Cell::Scalar(v) => Some(v.clone()),
                                    Cell::Node(id) => {
                                        view.ids.key_of(*id).map(|k| Value::Str(k.to_owned()))
                                    }
                                    Cell::Path(h) => Some(Value::Int(*h as i64)),
                                    Cell::Rel(_) => None,
                                })
                        })
                        .collect();
                    rs.push_row(vals);
                }
                rs
            } else {
                ResultSet::new(vec![])
            }
        }
    })
}

/// Execute the inner ops of a `LeftOuterApply` against a set of seed rows
/// (always a single-element vec in practice).  Returns the resulting rows.
///
/// This intentionally runs the same staged-path logic as `execute_inner` but
/// without the routing checks, aggregate fast-paths, or `check_params` (those
/// are handled by the outer call).  Only the ops that can appear inside an
/// OPTIONAL MATCH inner plan are needed: ScanLabel, JoinBound, LookupProps,
/// Expand, Filter.
fn exec_left_outer_inner(
    view: &GraphView,
    vars: &VarTable,
    mut rows: Vec<Row>,
    inner: &[PlanOp],
    params: &Params,
) -> Result<Vec<Row>, String> {
    for op in inner {
        match op {
            PlanOp::ScanLabel { var, label } => {
                rows = scan_label(view, vars, &rows, var, label.as_deref())?;
            }
            PlanOp::ScanKey { var, key, label } => {
                rows = scan_key(view, vars, &rows, var, key, label.as_deref(), params)?;
            }
            PlanOp::LookupProps { var, props } => {
                rows = retain_node(view, vars, &rows, var, None, props, params)?;
            }
            PlanOp::JoinBound { var, label, props } => {
                rows = retain_node(view, vars, &rows, var, label.as_deref(), props, params)?;
            }
            PlanOp::Expand { .. } => {
                rows = exec_expand(view, vars, &rows, op, params)?;
            }
            PlanOp::Filter { expr } => {
                rows = exec_filter(view, vars, &rows, expr, params)?;
            }
            other => {
                return Err(format!(
                    "unsupported op inside OPTIONAL MATCH inner plan: {other:?}"
                ));
            }
        }
    }
    Ok(rows)
}

fn finish(table: Projected) -> ResultSet {
    let mut rs = ResultSet::new(table.columns);
    for row in table.rows {
        rs.push_row(row);
    }
    rs
}

fn check_params(plan: &[PlanOp], params: &Params) -> Result<(), String> {
    let mut names = Vec::new();
    let mut seen = BTreeSet::new();
    collect_params_from_ops(plan, &mut names, &mut seen)?;
    for name in names {
        if !params.0.contains_key(&name) {
            return Err(format!("missing parameter `{name}`"));
        }
    }
    Ok(())
}

fn collect_params_from_ops(
    plan: &[PlanOp],
    names: &mut Vec<String>,
    seen: &mut BTreeSet<String>,
) -> Result<(), String> {
    for op in plan {
        match op {
            PlanOp::ScanKey { key, .. } => collect_operand(key, names, seen),
            PlanOp::LookupProps { props, .. }
            | PlanOp::JoinBound { props, .. }
            | PlanOp::Expand {
                to_props: props, ..
            } => {
                for (_, operand) in props {
                    collect_operand(operand, names, seen);
                }
            }
            PlanOp::Filter { expr } => collect_expr(expr, names, seen, 0)?,
            PlanOp::LeftOuterApply { inner, .. } => {
                collect_params_from_ops(inner, names, seen)?;
            }
            // Recurse into Project, With, and GroupAggregate to catch $param
            // references inside RETURN/WITH FuncCall args and group-key items.
            PlanOp::Project { items } => {
                for item in items {
                    collect_ret_item_params(item, names, seen);
                }
            }
            PlanOp::With {
                items, where_expr, ..
            } => {
                for item in items {
                    collect_ret_item_params(item, names, seen);
                }
                if let Some(expr) = where_expr {
                    let _ = collect_expr(expr, names, seen, 0);
                }
            }
            PlanOp::GroupAggregate { keys, aggs } => {
                for (_, item) in keys {
                    collect_ret_item_params(item, names, seen);
                }
                for (_, arg, _) in aggs {
                    match arg {
                        AggArg::Star => {}
                        AggArg::Var(_) => {}
                        AggArg::Prop { .. } => {}
                        AggArg::Distinct(_) => {}
                    }
                }
            }
            // SKIP/LIMIT $param: register the name so missing params are
            // caught at pre-flight time rather than execution time.
            PlanOp::Skip(LimitSkip::Param(n)) | PlanOp::Limit(LimitSkip::Param(n))
                if seen.insert(n.clone()) =>
            {
                names.push(n.clone());
            }
            _ => {}
        }
    }
    Ok(())
}

/// Collect `$param` names referenced inside a single RETURN/WITH item.
fn collect_ret_item_params(item: &RetItem, names: &mut Vec<String>, seen: &mut BTreeSet<String>) {
    match &item.value {
        RetVal::FuncCall { args, .. } => {
            for arg in args {
                collect_operand(arg, names, seen);
            }
        }
        RetVal::ScalarExpr(op) => {
            collect_operand(op, names, seen);
        }
        RetVal::Prop { .. } | RetVal::Var(_) | RetVal::Agg { .. } => {}
    }
}

fn collect_operand(op: &Operand, names: &mut Vec<String>, seen: &mut BTreeSet<String>) {
    match op {
        Operand::Param(n) => {
            if seen.insert(n.clone()) {
                names.push(n.clone());
            }
        }
        Operand::FuncCall { args, .. } => {
            for arg in args {
                collect_operand(arg, names, seen);
            }
        }
        Operand::BinArith { left, right, .. } => {
            collect_operand(left, names, seen);
            collect_operand(right, names, seen);
        }
        _ => {}
    }
}

fn collect_expr(
    expr: &Expr,
    names: &mut Vec<String>,
    seen: &mut BTreeSet<String>,
    depth: u32,
) -> Result<(), String> {
    if depth > 256 {
        return Err("expression nesting too deep".into());
    }
    match expr {
        Expr::And(lhs, rhs) | Expr::Or(lhs, rhs) => {
            collect_expr(lhs, names, seen, depth + 1)?;
            collect_expr(rhs, names, seen, depth + 1)
        }
        Expr::Not(inner) => collect_expr(inner, names, seen, depth + 1),
        Expr::Cmp { lhs, rhs, .. } => {
            collect_operand(lhs, names, seen);
            collect_operand(rhs, names, seen);
            Ok(())
        }
        Expr::Truthy(op) => {
            collect_operand(op, names, seen);
            Ok(())
        }
        Expr::IsNull(op) | Expr::IsNotNull(op) => {
            collect_operand(op, names, seen);
            Ok(())
        }
        Expr::In { expr, list } => {
            collect_operand(expr, names, seen);
            for item in list {
                collect_operand(item, names, seen);
            }
            Ok(())
        }
    }
}

fn collect_vars(plan: &[PlanOp]) -> VarTable {
    let mut vars = VarTable { names: Vec::new() };
    for op in plan {
        match op {
            PlanOp::ScanLabel { var, .. } => {
                vars.intern(var);
            }
            PlanOp::ScanKey { var, key, .. } => {
                vars.intern(var);
                intern_operand(&mut vars, key);
            }
            PlanOp::IndexScan { var, value, .. } => {
                vars.intern(var);
                intern_operand(&mut vars, value);
            }
            PlanOp::IndexIntersect {
                var, equalities, ..
            } => {
                vars.intern(var);
                for (_, operand) in equalities {
                    intern_operand(&mut vars, operand);
                }
            }
            PlanOp::LookupProps { var, props } | PlanOp::JoinBound { var, props, .. } => {
                vars.intern(var);
                for (_, operand) in props {
                    intern_operand(&mut vars, operand);
                }
            }
            PlanOp::Expand {
                from,
                rel_var,
                to,
                to_props,
                ..
            } => {
                vars.intern(from);
                vars.intern(to);
                if let Some(r) = rel_var {
                    vars.intern(r);
                }
                for (_, operand) in to_props {
                    intern_operand(&mut vars, operand);
                }
            }
            PlanOp::VarExpand {
                from, rel_var, to, ..
            } => {
                vars.intern(from);
                vars.intern(to);
                if let Some(r) = rel_var {
                    vars.intern(r);
                }
            }
            PlanOp::ShortestPath {
                from, rel_var, to, ..
            } => {
                vars.intern(from);
                vars.intern(to);
                if let Some(r) = rel_var {
                    vars.intern(r);
                }
            }
            PlanOp::Filter { expr } => intern_expr(&mut vars, expr),
            PlanOp::Project { items } => {
                for item in items {
                    match &item.value {
                        RetVal::Var(name) | RetVal::Prop { var: name, .. } => {
                            vars.intern(name);
                        }
                        RetVal::Agg { .. } => {} // handled by Aggregate op
                        RetVal::FuncCall { args, .. } => {
                            for arg in args {
                                intern_operand(&mut vars, arg);
                            }
                        }
                        RetVal::ScalarExpr(op) => {
                            intern_operand(&mut vars, op);
                        }
                    }
                }
            }
            PlanOp::Aggregate { arg, .. } => intern_agg_arg(&mut vars, arg),
            PlanOp::GroupAggregate { keys, aggs } => {
                // Intern input variable names (for the producer pass).
                for (col, item) in keys {
                    match &item.value {
                        RetVal::Var(name) | RetVal::Prop { var: name, .. } => {
                            vars.intern(name);
                        }
                        RetVal::Agg { .. } => {}
                        RetVal::FuncCall { args, .. } => {
                            for arg in args {
                                intern_operand(&mut vars, arg);
                            }
                        }
                        RetVal::ScalarExpr(op) => {
                            intern_operand(&mut vars, op);
                        }
                    }
                    // Intern output column name (for subsequent pipeline stages).
                    vars.intern(col);
                }
                for (_, arg, col) in aggs {
                    intern_agg_arg(&mut vars, arg);
                    // Intern output column name.
                    vars.intern(col);
                }
            }
            PlanOp::With {
                items,
                where_expr,
                order_by,
                ..
            } => {
                for item in items {
                    match &item.value {
                        RetVal::Var(name) | RetVal::Prop { var: name, .. } => {
                            vars.intern(name);
                        }
                        RetVal::Agg { .. } => {}
                        RetVal::FuncCall { args, .. } => {
                            for arg in args {
                                intern_operand(&mut vars, arg);
                            }
                        }
                        RetVal::ScalarExpr(op) => {
                            intern_operand(&mut vars, op);
                        }
                    }
                    // Intern output column name (alias or derived). The
                    // derived name must be the one `column_name` gives the
                    // same item, or the slot a later stage looks the alias up
                    // in does not exist.
                    if let Some(alias) = &item.alias {
                        vars.intern(alias);
                    } else if let Some(col) = ret_val_label(&item.value) {
                        vars.intern(&col);
                    }
                }
                if let Some(expr) = where_expr {
                    intern_expr(&mut vars, expr);
                }
                for oi in order_by {
                    match &oi.target {
                        OrderTarget::Alias(name) | OrderTarget::Var(name) => {
                            vars.intern(name);
                        }
                        OrderTarget::Prop { var, .. } => {
                            vars.intern(var);
                        }
                    }
                }
            }
            PlanOp::Unwind { expr, alias } => {
                vars.intern(alias);
                match expr {
                    UnwindExpr::Prop { var, .. } => {
                        vars.intern(var);
                    }
                    UnwindExpr::Var(name) => {
                        vars.intern(name);
                    }
                    UnwindExpr::Lit(_) => {}
                }
            }
            PlanOp::LeftOuterApply {
                inner,
                optional_vars,
            } => {
                // Intern all variables introduced by the inner plan.
                for op in inner {
                    // Recurse by treating inner ops through the same collect_vars logic.
                    // We re-use the same VarTable reference here, which is correct:
                    // inner vars are visible in the outer row after the apply.
                    match op {
                        PlanOp::ScanLabel { var, .. } => {
                            vars.intern(var);
                        }
                        PlanOp::ScanKey { var, key, .. } => {
                            vars.intern(var);
                            intern_operand(&mut vars, key);
                        }
                        PlanOp::IndexScan { var, value, .. } => {
                            vars.intern(var);
                            intern_operand(&mut vars, value);
                        }
                        PlanOp::IndexIntersect {
                            var, equalities, ..
                        } => {
                            vars.intern(var);
                            for (_, operand) in equalities {
                                intern_operand(&mut vars, operand);
                            }
                        }
                        PlanOp::Expand {
                            from, rel_var, to, ..
                        } => {
                            vars.intern(from);
                            vars.intern(to);
                            if let Some(r) = rel_var {
                                vars.intern(r);
                            }
                        }
                        PlanOp::JoinBound { var, .. } | PlanOp::LookupProps { var, .. } => {
                            vars.intern(var);
                        }
                        PlanOp::Filter { expr } => intern_expr(&mut vars, expr),
                        _ => {}
                    }
                }
                for v in optional_vars {
                    vars.intern(v);
                }
            }
            _ => {}
        }
    }
    vars
}

fn intern_operand(vars: &mut VarTable, operand: &Operand) {
    match operand {
        Operand::Prop { var, .. } | Operand::Var(var) => {
            vars.intern(var);
        }
        Operand::Lit(_) | Operand::Param(_) => {}
        Operand::BinArith { left, right, .. } => {
            intern_operand(vars, left);
            intern_operand(vars, right);
        }
        Operand::FuncCall { args, .. } => {
            for arg in args {
                intern_operand(vars, arg);
            }
        }
        Operand::Case { branches, default } => {
            for (cond, value) in branches {
                intern_expr(vars, cond);
                intern_operand(vars, value);
            }
            if let Some(d) = default {
                intern_operand(vars, d);
            }
        }
        Operand::Index { base, index } => {
            intern_operand(vars, base);
            intern_operand(vars, index);
        }
    }
}

/// Intern the variables an aggregate argument reads, through any `DISTINCT`.
fn intern_agg_arg(vars: &mut VarTable, arg: &AggArg) {
    match arg {
        AggArg::Star => {}
        AggArg::Var(v) => {
            vars.intern(v);
        }
        AggArg::Prop { var, .. } => {
            vars.intern(var);
        }
        AggArg::Distinct(inner) => intern_agg_arg(vars, inner),
    }
}

/// Render an aggregate argument the way it was written, for a column name.
fn agg_arg_label(arg: &AggArg) -> String {
    match arg {
        AggArg::Star => "*".to_string(),
        AggArg::Var(v) => v.clone(),
        AggArg::Prop { var, field } => format!("{var}.{field}"),
        AggArg::Distinct(inner) => format!("DISTINCT {}", agg_arg_label(inner)),
    }
}

fn intern_expr(vars: &mut VarTable, expr: &Expr) {
    match expr {
        Expr::And(lhs, rhs) | Expr::Or(lhs, rhs) => {
            intern_expr(vars, lhs);
            intern_expr(vars, rhs);
        }
        Expr::Not(inner) => intern_expr(vars, inner),
        Expr::Cmp { lhs, rhs, .. } => {
            intern_operand(vars, lhs);
            intern_operand(vars, rhs);
        }
        Expr::Truthy(op) => intern_operand(vars, op),
        Expr::IsNull(op) | Expr::IsNotNull(op) => intern_operand(vars, op),
        Expr::In { expr, list } => {
            intern_operand(vars, expr);
            for item in list {
                intern_operand(vars, item);
            }
        }
    }
}

fn scan_ids(view: &GraphView, label: Option<&str>) -> Vec<u32> {
    let ids = match label {
        Some(label) => view.nodes_with_label(label),
        // Real nodes always have labels; sentinel slots are gaps.
        None => (0..view.ids.len() as u32)
            .filter(|&id| view.label_of(id).is_some())
            .collect(),
    };
    if view.mask.is_some() {
        ids.into_iter().filter(|&id| view.visible(id)).collect()
    } else {
        ids
    }
}

/// Resolve a `ScanKey` operand to at most one dense id.
/// Missing key, non-string value, or label mismatch → `Ok(None)` (zero rows).
fn resolve_scan_key_id(
    view: &GraphView,
    vars: &VarTable,
    row: &Row,
    key: &Operand,
    label: Option<&str>,
    params: &Params,
) -> Result<Option<u32>, String> {
    #[cfg(test)]
    SCAN_KEY_FIRES.fetch_add(1, std::sync::atomic::Ordering::Relaxed);

    let Some(val) = resolve_operand(view, vars, row, key, params)? else {
        return Ok(None);
    };
    let Value::Str(s) = val else {
        return Ok(None);
    };
    let Some(id) = view.node_id(&s) else {
        return Ok(None);
    };
    if !view.visible(id) {
        return Ok(None);
    }
    if let Some(want) = label {
        match view.label_of(id) {
            Some(got) if got == want => {}
            _ => return Ok(None),
        }
    }
    Ok(Some(id))
}

fn scan_key(
    view: &GraphView,
    vars: &VarTable,
    rows: &[Row],
    var: &str,
    key: &Operand,
    label: Option<&str>,
    params: &Params,
) -> Result<Vec<Row>, String> {
    let slot = vars
        .slot(var)
        .ok_or_else(|| format!("unbound variable `{var}`"))?;
    let cap = max_intermediate_rows();
    let mut out = Vec::new();
    for row in rows {
        let Some(id) = resolve_scan_key_id(view, vars, row, key, label, params)? else {
            continue;
        };
        if out.len() >= cap {
            return Err(row_cap_err(cap));
        }
        let mut next = row.clone();
        next[slot] = Some(Cell::Node(id));
        out.push(next);
    }
    Ok(out)
}

fn scan_label(
    view: &GraphView,
    vars: &VarTable,
    rows: &[Row],
    var: &str,
    label: Option<&str>,
) -> Result<Vec<Row>, String> {
    let ids = scan_ids(view, label);
    let slot = vars
        .slot(var)
        .ok_or_else(|| format!("unbound variable `{var}`"))?;
    let cap = max_intermediate_rows();
    let mut out = Vec::with_capacity(rows.len().saturating_mul(ids.len()).min(cap));
    for row in rows {
        for &id in &ids {
            if out.len() >= cap {
                return Err(row_cap_err(cap));
            }
            let mut next = row.clone();
            next[slot] = Some(Cell::Node(id));
            out.push(next);
        }
    }
    Ok(out)
}

/// Resolve the matching node ids for an `IndexScan` of `label`/`field`/`value`.
///
/// Indexed path (`(label, field)` declared, concrete label, scalar value): the
/// property index answers directly. Otherwise a full label scan filtered by the
/// same `node_matches` equality the `LookupProps` plan uses — so the id set is
/// identical to `ScanLabel` + `LookupProps` regardless of whether an index
/// exists. `row` supplies `$param` bindings (the value is row-independent).
#[allow(clippy::too_many_arguments)]
fn index_scan_ids(
    view: &GraphView,
    vars: &VarTable,
    row: &Row,
    label: Option<&str>,
    field: &str,
    value: &Operand,
    params: &Params,
) -> Result<Vec<u32>, String> {
    if is_identity_eq_field(field) {
        return identity_eq_ids(view, vars, row, label, field, value, params);
    }
    let resolved = resolve_operand(view, vars, row, value, params)?;
    if let (Some(label_str), Some(val)) = (label, resolved.as_ref()) {
        if let Some(ids) = view.nodes_with_prop(label_str, field, val) {
            #[cfg(test)]
            INDEX_SCAN_FIRES.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
            return Ok(ids);
        }
    }
    // Fallback: full label scan + single-equality retain.
    let props = [(field.to_string(), value.clone())];
    let mut out = Vec::new();
    for id in scan_ids(view, label) {
        if node_matches(view, vars, row, id, None, &props, params)? {
            out.push(id);
        }
    }
    Ok(out)
}

/// Tagged `IndexScan` mode for `field ∈ {key, id}` (`IdentityEq`).
///
/// ScanKey the resolved string, keep the node only if stored-wins identity
/// equals the value, then union any other node whose stored `field` equals
/// the value (property index when declared, else a label/property filter).
#[allow(clippy::too_many_arguments)]
fn identity_eq_ids(
    view: &GraphView,
    vars: &VarTable,
    row: &Row,
    label: Option<&str>,
    field: &str,
    value: &Operand,
    params: &Params,
) -> Result<Vec<u32>, String> {
    let resolved = resolve_operand(view, vars, row, value, params)?;
    let mut out = Vec::new();
    let mut seen = HashSet::new();
    if let Some(id) = resolve_scan_key_id(view, vars, row, value, label, params)? {
        let props = [(field.to_string(), value.clone())];
        if node_matches(view, vars, row, id, None, &props, params)? {
            seen.insert(id);
            out.push(id);
        }
    }
    if let Some(val) = resolved.as_ref() {
        for id in stored_identity_hits(view, label, field, val) {
            if seen.insert(id) {
                out.push(id);
            }
        }
    }
    Ok(out)
}

/// Nodes whose *stored* `field` equals `val` (no identity fallback).
fn stored_identity_hits(
    view: &GraphView,
    label: Option<&str>,
    field: &str,
    val: &Value,
) -> Vec<u32> {
    if let Some(label_str) = label {
        if let Some(ids) = view.nodes_with_prop(label_str, field, val) {
            return ids;
        }
    }
    let mut out = Vec::new();
    for id in scan_ids(view, label) {
        if let Some(got) = view.prop(id, field).map(|vr| vr.into_value()) {
            if values_equal(&got, val) {
                out.push(id);
            }
        }
    }
    out
}

/// Execute an `IndexScan`: seed rows from nodes of `label` whose scalar
/// `field` equals `value` (see [`index_scan_ids`]).
#[allow(clippy::too_many_arguments)]
fn scan_index(
    view: &GraphView,
    vars: &VarTable,
    rows: &[Row],
    var: &str,
    label: Option<&str>,
    field: &str,
    value: &Operand,
    params: &Params,
) -> Result<Vec<Row>, String> {
    let Some(first) = rows.first() else {
        return Ok(Vec::new());
    };
    let ids = index_scan_ids(view, vars, first, label, field, value, params)?;
    let slot = vars
        .slot(var)
        .ok_or_else(|| format!("unbound variable `{var}`"))?;
    let cap = max_intermediate_rows();
    let mut out = Vec::new();
    for row in rows {
        for &id in &ids {
            if out.len() >= cap {
                return Err(row_cap_err(cap));
            }
            let mut next = row.clone();
            next[slot] = Some(Cell::Node(id));
            out.push(next);
        }
    }
    Ok(out)
}

/// Resolve candidate node ids for an `IndexIntersect` op.
///
/// Partitions `equalities` into indexed (label+field in `prop_index`) and
/// unindexed fields. If ALL fields are unindexed → full label scan filtered by
/// all equalities (`INDEX_INTERSECT_FIRES` does NOT advance). Otherwise:
/// intersects the sorted id-lists from indexed fields (smallest set first,
/// two-pointer merge), then applies unindexed fields as per-node post-filters
/// via `node_matches` (`INDEX_INTERSECT_FIRES` advances once).
#[allow(clippy::too_many_arguments)]
fn index_intersect_ids(
    view: &GraphView,
    vars: &VarTable,
    row: &Row,
    label: Option<&str>,
    equalities: &[(String, Operand)],
    params: &Params,
) -> Result<Vec<u32>, String> {
    // Resolve all operands to concrete Values (needed for both paths).
    let mut resolved: Vec<(String, Option<Value>)> = Vec::with_capacity(equalities.len());
    for (field, operand) in equalities {
        let val = resolve_operand(view, vars, row, operand, params)?;
        resolved.push((field.clone(), val));
    }

    // Partition into indexed and unindexed fields.
    let mut indexed_lists: Vec<Vec<u32>> = Vec::new();
    let mut unindexed_props: Vec<(String, Operand)> = Vec::new();

    for ((field, val_opt), (_, operand)) in resolved.iter().zip(equalities.iter()) {
        if let (Some(label_str), Some(val)) = (label, val_opt.as_ref()) {
            if let Some(ids) = view.nodes_with_prop(label_str, field, val) {
                indexed_lists.push(ids);
                continue;
            }
        }
        unindexed_props.push((field.clone(), operand.clone()));
    }

    if indexed_lists.is_empty() {
        // All-unindexed fallback: full label scan + filter. Counter does NOT advance.
        let mut out = Vec::new();
        for id in scan_ids(view, label) {
            if node_matches(view, vars, row, id, None, equalities, params)? {
                out.push(id);
            }
        }
        return Ok(out);
    }

    #[cfg(test)]
    INDEX_INTERSECT_FIRES.fetch_add(1, std::sync::atomic::Ordering::Relaxed);

    // Two-pointer intersection of sorted id lists — smallest list first.
    indexed_lists.sort_unstable_by_key(|v| v.len());
    let mut result = indexed_lists.remove(0);
    for other in indexed_lists {
        let mut merged = Vec::new();
        let (mut i, mut j) = (0, 0);
        while i < result.len() && j < other.len() {
            match result[i].cmp(&other[j]) {
                std::cmp::Ordering::Equal => {
                    merged.push(result[i]);
                    i += 1;
                    j += 1;
                }
                std::cmp::Ordering::Less => i += 1,
                std::cmp::Ordering::Greater => j += 1,
            }
        }
        result = merged;
    }

    // Apply unindexed fields as post-filter. Use a manual loop (not retain) so
    // that node_matches errors are propagated rather than silently dropped —
    // matching the all-unindexed fallback path which surfaces errors via `?`.
    if !unindexed_props.is_empty() {
        let mut filtered = Vec::with_capacity(result.len());
        for id in result {
            if node_matches(view, vars, row, id, None, &unindexed_props, params)? {
                filtered.push(id);
            }
        }
        result = filtered;
    }

    Ok(result)
}

/// Execute an `IndexIntersect`: seed rows from nodes satisfying all equalities
/// (see [`index_intersect_ids`]).
#[allow(clippy::too_many_arguments)]
fn scan_intersect(
    view: &GraphView,
    vars: &VarTable,
    rows: &[Row],
    var: &str,
    label: Option<&str>,
    equalities: &[(String, Operand)],
    params: &Params,
) -> Result<Vec<Row>, String> {
    let Some(first) = rows.first() else {
        return Ok(Vec::new());
    };
    let ids = index_intersect_ids(view, vars, first, label, equalities, params)?;
    let slot = vars
        .slot(var)
        .ok_or_else(|| format!("unbound variable `{var}`"))?;
    let cap = max_intermediate_rows();
    let mut out = Vec::new();
    for row in rows {
        for &id in &ids {
            if out.len() >= cap {
                return Err(row_cap_err(cap));
            }
            let mut next = row.clone();
            next[slot] = Some(Cell::Node(id));
            out.push(next);
        }
    }
    Ok(out)
}

fn require_cell<'a>(row: &'a Row, vars: &VarTable, var: &str) -> Result<&'a Cell, String> {
    let slot = vars
        .slot(var)
        .ok_or_else(|| format!("unbound variable `{var}`"))?;
    row.get(slot)
        .and_then(|c| c.as_ref())
        .ok_or_else(|| format!("unbound variable `{var}`"))
}

fn require_node(row: &Row, vars: &VarTable, var: &str) -> Result<u32, String> {
    match require_cell(row, vars, var)? {
        Cell::Node(id) => Ok(*id),
        Cell::Rel(_) => Err(format!("variable `{var}` is not a node")),
        Cell::Path(_) => Err(format!("variable `{var}` is a path, not a node")),
        Cell::Scalar(_) => Err(format!("variable `{var}` is a scalar value, not a node")),
    }
}

/// Supported scalar function names (case-insensitive).
const SCALAR_FUNCS: &[&str] = &[
    "toLower",
    "toUpper",
    "size",
    "coalesce",
    "type",
    "abs",
    "round",
    "textMatches",
    "contains",
    "startsWith",
    "endsWith",
    "toInteger",
    "toFloat",
    "toString",
    "decay",
    "key",
    "id",
    "labels",
];

/// Evaluate a two-string-argument predicate (`contains` / `startsWith` /
/// `endsWith`). Null in either argument → null; non-string args → null.
fn eval_string_predicate(
    name: &str,
    args: &[Operand],
    view: &GraphView,
    vars: &VarTable,
    row: &Row,
    params: &Params,
    f: impl Fn(&str, &str) -> bool,
) -> Result<Option<Value>, String> {
    if args.len() != 2 {
        return Err(format!(
            "{name}() requires exactly 2 arguments, got {}",
            args.len()
        ));
    }
    let a = resolve_operand(view, vars, row, &args[0], params)?;
    let b = resolve_operand(view, vars, row, &args[1], params)?;
    match (a, b) {
        (Some(Value::Str(s)), Some(Value::Str(sub))) => Ok(Some(Value::Bool(f(&s, &sub)))),
        (None, _) | (_, None) => Ok(None),
        _ => Ok(None),
    }
}

/// Evaluate one of the supported scalar functions.  Unknown function names
/// produce a named error listing the supported set.  Null propagation:
/// null in → null out, EXCEPT `coalesce` (which skips nulls).
fn eval_func(
    name: &str,
    args: &[Operand],
    view: &GraphView,
    vars: &VarTable,
    row: &Row,
    params: &Params,
) -> Result<Option<Value>, String> {
    let norm = name.to_ascii_lowercase();
    match norm.as_str() {
        "tolower" => {
            if args.len() != 1 {
                return Err(format!(
                    "toLower() requires exactly 1 argument, got {}",
                    args.len()
                ));
            }
            let v = resolve_operand(view, vars, row, &args[0], params)?;
            Ok(v.map(|val| match val {
                Value::Str(s) => Value::Str(s.to_ascii_lowercase()),
                other => other, // non-string: return unchanged (null already handled)
            }))
        }
        "toupper" => {
            if args.len() != 1 {
                return Err(format!(
                    "toUpper() requires exactly 1 argument, got {}",
                    args.len()
                ));
            }
            let v = resolve_operand(view, vars, row, &args[0], params)?;
            Ok(v.map(|val| match val {
                Value::Str(s) => Value::Str(s.to_ascii_uppercase()),
                other => other,
            }))
        }
        "size" => {
            if args.len() != 1 {
                return Err(format!(
                    "size() requires exactly 1 argument, got {}",
                    args.len()
                ));
            }
            let v = resolve_operand(view, vars, row, &args[0], params)?;
            match v {
                None => Ok(None), // null in → null out
                Some(Value::Str(s)) => Ok(Some(Value::Int(s.len() as i64))),
                Some(Value::List(items)) => Ok(Some(Value::Int(items.len() as i64))),
                Some(_) => Ok(None), // non-string/non-list → null (openCypher)
            }
        }
        "coalesce" => {
            // Returns first non-null arg; null if all args are null.
            for arg in args {
                if let Some(v) = resolve_operand(view, vars, row, arg, params)? {
                    return Ok(Some(v));
                }
            }
            Ok(None)
        }
        "type" => {
            if args.len() != 1 {
                return Err(format!(
                    "type() requires exactly 1 argument, got {}",
                    args.len()
                ));
            }
            // Argument must be a relationship variable (Cell::Rel).
            let Operand::Var(var_name) = &args[0] else {
                return Err(
                    "type() argument must be a relationship variable (e.g. type(r))".to_string(),
                );
            };
            let slot = vars
                .slot(var_name)
                .ok_or_else(|| format!("unbound variable `{var_name}` in type()"))?;
            match row.get(slot).and_then(|c| c.as_ref()) {
                Some(Cell::Rel(e)) => {
                    // Look up the etype string from the symbol table.
                    let etype = view.syms.resolve(e.etype).unwrap_or("").to_owned();
                    Ok(Some(Value::Str(etype)))
                }
                Some(Cell::Node(_)) => Err(format!(
                    "type() argument `{var_name}` is a node, not a relationship"
                )),
                Some(Cell::Scalar(_) | Cell::Path(_)) => Err(format!(
                    "type() argument `{var_name}` is not a relationship"
                )),
                None => Ok(None), // null binding → null (optional match scenario)
            }
        }
        "key" | "id" => {
            let fname = if norm == "id" { "id" } else { "key" };
            if args.len() != 1 {
                return Err(format!(
                    "{fname}() requires exactly 1 argument, got {}",
                    args.len()
                ));
            }
            let Operand::Var(var_name) = &args[0] else {
                return Err(format!(
                    "{fname}() argument must be a node variable (e.g. {fname}(n))"
                ));
            };
            let slot = vars
                .slot(var_name)
                .ok_or_else(|| format!("unbound variable `{var_name}` in {fname}()"))?;
            match row.get(slot).and_then(|c| c.as_ref()) {
                Some(Cell::Node(id)) => Ok(Some(Value::Str(view.key_of(*id).to_owned()))),
                Some(Cell::Rel(_)) => Err(format!(
                    "{fname}() argument `{var_name}` is a relationship, not a node"
                )),
                Some(Cell::Scalar(_) | Cell::Path(_)) => {
                    Err(format!("{fname}() argument `{var_name}` is not a node"))
                }
                None => Ok(None), // null binding → null (optional match scenario)
            }
        }
        "labels" => {
            if args.len() != 1 {
                return Err(format!(
                    "labels() requires exactly 1 argument, got {}",
                    args.len()
                ));
            }
            // openCypher returns a list because a node may carry several
            // labels; this store gives a node exactly one, so the list holds
            // one element. `n.label` is the scalar spelling of the same thing.
            let Operand::Var(var_name) = &args[0] else {
                return Err(
                    "labels() argument must be a node variable (e.g. labels(n))".to_string()
                );
            };
            let slot = vars
                .slot(var_name)
                .ok_or_else(|| format!("unbound variable `{var_name}` in labels()"))?;
            match row.get(slot).and_then(|c| c.as_ref()) {
                Some(Cell::Node(id)) => Ok(Some(Value::List(
                    view.label_of(*id)
                        .map(|l| vec![Value::Str(l.to_owned())])
                        .unwrap_or_default(),
                ))),
                Some(Cell::Rel(_)) => Err(format!(
                    "labels() argument `{var_name}` is a relationship, not a node"
                )),
                Some(Cell::Scalar(_) | Cell::Path(_)) => {
                    Err(format!("labels() argument `{var_name}` is not a node"))
                }
                None => Ok(None), // null binding → null (optional match scenario)
            }
        }
        "abs" => {
            if args.len() != 1 {
                return Err(format!(
                    "abs() requires exactly 1 argument, got {}",
                    args.len()
                ));
            }
            let v = resolve_operand(view, vars, row, &args[0], params)?;
            match v {
                None => Ok(None),
                Some(Value::Int(n)) => Ok(Some(Value::Int(n.abs()))),
                Some(Value::Float(f)) => Ok(Some(Value::Float(f.abs()))),
                Some(_) => Ok(None), // non-numeric → null
            }
        }
        "round" => {
            if args.len() != 1 {
                return Err(format!(
                    "round() requires exactly 1 argument, got {}",
                    args.len()
                ));
            }
            let v = resolve_operand(view, vars, row, &args[0], params)?;
            match v {
                None => Ok(None),
                Some(Value::Int(n)) => Ok(Some(Value::Int(n))), // int already rounded
                Some(Value::Float(f)) => Ok(Some(Value::Float(f.round()))),
                Some(_) => Ok(None), // non-numeric → null
            }
        }
        "textmatches" => {
            // Full-text predicate for use in WHERE clauses.
            // Signature: textMatches(field_value, "query string") → Bool
            //
            // Design choice: evaluated per-row via scratch tokenization, not via
            // the inverted index. The index provides performance for db.search();
            // this function provides a convenient WHERE-position predicate that
            // integrates with the planner's existing filter machinery without
            // requiring the index to be threaded into GraphView.
            //
            // Performance characteristic (documented): O(scan) per MATCH row.
            // For large result sets prefer db.search() + IN or a labeled filter.
            if args.len() != 2 {
                return Err(format!(
                    "textMatches() requires exactly 2 arguments (field_value, query), got {}",
                    args.len()
                ));
            }
            let field_val = resolve_operand(view, vars, row, &args[0], params)?;
            let query_val = resolve_operand(view, vars, row, &args[1], params)?;
            match (field_val, query_val) {
                (None, _) | (_, None) => Ok(None),
                (Some(Value::Str(s)), Some(Value::Str(q))) => {
                    // v2 grammar: phrase adjacency, negation, prefix, stemming.
                    Ok(Some(Value::Bool(core_storage::fulltext::eval_query_str(
                        &s, &q,
                    ))))
                }
                (Some(Value::List(items)), Some(Value::Str(q))) => Ok(Some(Value::Bool(
                    core_storage::fulltext::eval_query_str_list(&items, &q),
                ))),
                _ => Ok(Some(Value::Bool(false))), // non-string field → no match
            }
        }
        "contains" => eval_string_predicate("contains", args, view, vars, row, params, |s, sub| {
            s.contains(sub)
        }),
        "startswith" => {
            eval_string_predicate("startsWith", args, view, vars, row, params, |s, p| {
                s.starts_with(p)
            })
        }
        "endswith" => eval_string_predicate("endsWith", args, view, vars, row, params, |s, p| {
            s.ends_with(p)
        }),
        "tointeger" => {
            if args.len() != 1 {
                return Err(format!(
                    "toInteger() requires exactly 1 argument, got {}",
                    args.len()
                ));
            }
            let v = resolve_operand(view, vars, row, &args[0], params)?;
            Ok(match v {
                None => None,
                Some(Value::Int(n)) => Some(Value::Int(n)),
                Some(Value::Float(f)) => Some(Value::Int(f.trunc() as i64)),
                // Cypher parses an integer literal; a float-looking string
                // truncates via the float parse. Unparseable → null.
                Some(Value::Str(s)) => s
                    .trim()
                    .parse::<i64>()
                    .ok()
                    .or_else(|| s.trim().parse::<f64>().ok().map(|f| f.trunc() as i64))
                    .map(Value::Int),
                Some(_) => None,
            })
        }
        "tofloat" => {
            if args.len() != 1 {
                return Err(format!(
                    "toFloat() requires exactly 1 argument, got {}",
                    args.len()
                ));
            }
            let v = resolve_operand(view, vars, row, &args[0], params)?;
            Ok(match v {
                None => None,
                Some(Value::Float(f)) => Some(Value::Float(f)),
                Some(Value::Int(n)) => Some(Value::Float(n as f64)),
                Some(Value::Str(s)) => s.trim().parse::<f64>().ok().map(Value::Float),
                Some(_) => None,
            })
        }
        "tostring" => {
            if args.len() != 1 {
                return Err(format!(
                    "toString() requires exactly 1 argument, got {}",
                    args.len()
                ));
            }
            let v = resolve_operand(view, vars, row, &args[0], params)?;
            Ok(match v {
                None => None,
                Some(Value::Str(s)) => Some(Value::Str(s)),
                Some(Value::Int(n)) => Some(Value::Str(n.to_string())),
                Some(Value::Float(f)) => Some(Value::Str(f.to_string())),
                Some(Value::Bool(b)) => Some(Value::Str(b.to_string())),
                Some(_) => None, // list/map have no scalar string form
            })
        }
        "decay" => {
            // decay(base, age, halflife) = base * 0.5^(age / halflife).
            // Null in any argument → null; halflife <= 0 is an error, not a
            // silent null, since it signals a caller bug (divide-by-zero /
            // sign-flip), not an absent value.
            if args.len() != 3 {
                return Err(format!(
                    "decay() requires exactly 3 arguments, got {}",
                    args.len()
                ));
            }
            let base = resolve_operand(view, vars, row, &args[0], params)?;
            let age = resolve_operand(view, vars, row, &args[1], params)?;
            let halflife = resolve_operand(view, vars, row, &args[2], params)?;
            match (base, age, halflife) {
                (None, _, _) | (_, None, _) | (_, _, None) => Ok(None),
                (Some(b), Some(a), Some(h)) => {
                    let b = numeric_val(&b)
                        .ok_or_else(|| "decay() requires numeric arguments".to_string())?;
                    let a = numeric_val(&a)
                        .ok_or_else(|| "decay() requires numeric arguments".to_string())?;
                    let h = numeric_val(&h)
                        .ok_or_else(|| "decay() requires numeric arguments".to_string())?;
                    if h <= 0.0 {
                        return Err("decay() requires halflife > 0".to_string());
                    }
                    Ok(Some(Value::Float(b * 0.5f64.powf(a / h))))
                }
            }
        }
        _ => Err(format!(
            "unknown function `{name}`; supported: {}",
            SCALAR_FUNCS.join(", ")
        )),
    }
}

fn resolve_operand(
    view: &GraphView,
    vars: &VarTable,
    row: &Row,
    operand: &Operand,
    params: &Params,
) -> Result<Option<Value>, String> {
    match operand {
        Operand::Lit(v) => Ok(Some(v.clone())),
        Operand::Param(name) => match params.0.get(name) {
            Some(v) => Ok(Some(v.clone())),
            None => Err(format!("missing parameter `{name}`")),
        },
        Operand::Prop { var, field } => resolve_prop(view, vars, row, var, field),
        Operand::Var(name) => {
            // Bare variable reference: resolve from Cell (scalar alias or node key).
            match vars
                .slot(name)
                .and_then(|s| row.get(s))
                .and_then(|c| c.as_ref())
            {
                Some(Cell::Scalar(v)) => Ok(Some(v.clone())),
                Some(Cell::Node(id)) => match view.ids.key_of(*id) {
                    Some(key) => Ok(Some(Value::Str(key.to_owned()))),
                    None => Ok(None),
                },
                Some(Cell::Path(hops)) => Ok(Some(Value::Int(*hops as i64))),
                Some(Cell::Rel(_)) => Err(format!("variable `{name}` is a relationship")),
                None => Ok(None),
            }
        }
        Operand::FuncCall { name, args } => eval_func(name, args, view, vars, row, params),
        Operand::Index { base, index } => {
            let base_val = resolve_operand(view, vars, row, base, params)?;
            let idx_val = resolve_operand(view, vars, row, index, params)?;
            Ok(crate::value_ops::index_list(base_val, idx_val))
        }
        Operand::Case { branches, default } => {
            for (cond, value) in branches {
                if eval_expr(view, vars, row, cond, params, 0)? {
                    return resolve_operand(view, vars, row, value, params);
                }
            }
            match default {
                Some(d) => resolve_operand(view, vars, row, d, params),
                None => Ok(None),
            }
        }
        Operand::BinArith { op, left, right } => {
            use super::ast::ArithOp;
            let lv = resolve_operand(view, vars, row, left, params)?;
            let rv = resolve_operand(view, vars, row, right, params)?;
            match (lv, rv) {
                (None, _) | (_, None) => Ok(None), // null propagation
                (Some(Value::Int(a)), Some(Value::Int(b))) => {
                    let result = match op {
                        ArithOp::Sub => a.saturating_sub(b),
                        ArithOp::Mul => a.saturating_mul(b),
                        ArithOp::Add => a.saturating_add(b),
                        ArithOp::Div => {
                            if b == 0 {
                                return Err("division by zero".into());
                            }
                            a.checked_div(b).unwrap_or(i64::MAX)
                        }
                    };
                    Ok(Some(Value::Int(result)))
                }
                (Some(lv), Some(rv)) => {
                    let a = match &lv {
                        Value::Float(f) => *f,
                        Value::Int(i) => *i as f64,
                        _ => return Err(format!("arithmetic operand must be numeric, got {lv:?}")),
                    };
                    let b = match &rv {
                        Value::Float(f) => *f,
                        Value::Int(i) => *i as f64,
                        _ => return Err(format!("arithmetic operand must be numeric, got {rv:?}")),
                    };
                    let result = match op {
                        ArithOp::Sub => a - b,
                        ArithOp::Mul => a * b,
                        ArithOp::Add => a + b,
                        ArithOp::Div => {
                            if b == 0.0 {
                                return Err("division by zero".into());
                            }
                            a / b
                        }
                    };
                    Ok(Some(Value::Float(result)))
                }
            }
        }
    }
}

/// True for field names that read node identity rather than a stored
/// property. See [`node_identity_prop`].
fn is_identity_field(field: &str) -> bool {
    field == "key" || field == "id" || field == "label"
}

/// Identity equality fields that plan as a tagged `IndexScan` (`IdentityEq`):
/// `n.key` / `n.id` / `key(n)` / `id(n)`. `n.label` is not a key lookup.
fn is_identity_eq_field(field: &str) -> bool {
    field == "key" || field == "id"
}

/// Node identity exposed as a property: `n.key`, `n.id`, and `n.label`.
///
/// The key lives in the id map and the label in the interner — so `n.id`
/// used to read as null while `{id:}` already did a key lookup. A property
/// of the same name always wins, so a graph that really does store an `id`
/// or `key` field keeps it. `n.id` is an alias of `n.key`.
fn node_identity_prop(view: &GraphView, id: u32, field: &str) -> Option<Value> {
    match field {
        "key" | "id" => view.ids.key_of(id).map(|k| Value::Str(k.to_owned())),
        "label" => view.label_of(id).map(|l| Value::Str(l.to_owned())),
        _ => None,
    }
}

fn resolve_prop(
    view: &GraphView,
    vars: &VarTable,
    row: &Row,
    var: &str,
    field: &str,
) -> Result<Option<Value>, String> {
    // Distinguish "variable not in scope" (hard error) from "variable is null
    // because OPTIONAL MATCH found no match" (returns null, not an error).
    let slot = vars
        .slot(var)
        .ok_or_else(|| format!("unbound variable `{var}`"))?;
    let cell = match row.get(slot).and_then(|c| c.as_ref()) {
        Some(c) => c,
        // Slot exists but is None → OPTIONAL MATCH null binding; propagate null.
        None => return Ok(None),
    };
    match cell {
        Cell::Node(id) => Ok(view
            .prop(*id, field)
            .map(|vr| vr.into_value())
            .or_else(|| node_identity_prop(view, *id, field))),
        Cell::Rel(e) => Ok(view.edge_props.get(e.etype, e.src, e.dst, field)),
        // Virtual path cell: only `length` is exposed.
        Cell::Path(hops) => {
            if field == "length" {
                Ok(Some(Value::Int(*hops as i64)))
            } else {
                Ok(None)
            }
        }
        // Scalar alias: has no properties.
        Cell::Scalar(_) => Ok(None),
    }
}

fn node_matches(
    view: &GraphView,
    vars: &VarTable,
    row: &Row,
    id: u32,
    label: Option<&str>,
    props: &[(String, Operand)],
    params: &Params,
) -> Result<bool, String> {
    if let Some(want) = label {
        match view.label_of(id) {
            Some(got) if got == want => {}
            _ => return Ok(false),
        }
    }
    for (field, operand) in props {
        let Some(expected) = resolve_operand(view, vars, row, operand, params)? else {
            return Ok(false);
        };
        let got = view
            .prop(id, field)
            .map(|vr| vr.into_value())
            .or_else(|| node_identity_prop(view, id, field));
        match got {
            Some(got) if values_equal(&got, &expected) => {}
            _ => return Ok(false),
        }
    }
    Ok(true)
}

fn retain_node(
    view: &GraphView,
    vars: &VarTable,
    rows: &[Row],
    var: &str,
    label: Option<&str>,
    props: &[(String, Operand)],
    params: &Params,
) -> Result<Vec<Row>, String> {
    let mut out = Vec::with_capacity(rows.len());
    for row in rows {
        let id = require_node(row, vars, var)?;
        if node_matches(view, vars, row, id, label, props, params)? {
            out.push(row.clone());
        }
    }
    Ok(out)
}

fn map_dir(dir: RelDir) -> Dir {
    match dir {
        RelDir::Right => Dir::Out,
        RelDir::Left => Dir::In,
        RelDir::Undirected => Dir::Both,
    }
}

fn neighbor(from: u32, e: &EdgeRef, dir: RelDir) -> u32 {
    match dir {
        RelDir::Right => e.dst,
        RelDir::Left => e.src,
        RelDir::Undirected => {
            if e.src == from {
                e.dst
            } else {
                e.src
            }
        }
    }
}

fn row_has_edge(row: &Row, e: &EdgeRef) -> bool {
    row.iter()
        .any(|c| matches!(c, Some(Cell::Rel(existing)) if existing == e))
}

fn resolve_etypes(view: &GraphView, etypes: &[String]) -> Option<Vec<u32>> {
    if etypes.is_empty() {
        None // no type constraint → all edge types
    } else {
        // Resolve each named type to its symbol, skipping any not interned
        // (a named type with no edges contributes nothing).
        Some(etypes.iter().filter_map(|n| view.syms.get(n)).collect())
    }
}

fn exec_expand(
    view: &GraphView,
    vars: &VarTable,
    rows: &[Row],
    op: &PlanOp,
    params: &Params,
) -> Result<Vec<Row>, String> {
    let PlanOp::Expand {
        from,
        rel_var,
        etypes,
        dir,
        to,
        to_label,
        to_props,
    } = op
    else {
        return Err("internal: expected Expand".into());
    };
    let etypes = resolve_etypes(view, etypes);
    let exp_dir = map_dir(*dir);
    let to_slot = vars
        .slot(to)
        .ok_or_else(|| format!("unbound variable `{to}`"))?;
    let rel_slot = rel_var.as_ref().and_then(|rv| vars.slot(rv));
    let cap = max_intermediate_rows();
    let mut out = Vec::with_capacity(rows.len().saturating_mul(2).min(cap));
    for row in rows {
        let from_id = require_node(row, vars, from)?;
        let bound_to = match row.get(to_slot).and_then(|c| c.as_ref()) {
            Some(Cell::Node(id)) => Some(*id),
            Some(Cell::Rel(_) | Cell::Path(_) | Cell::Scalar(_)) => {
                return Err(format!("variable `{to}` is not a node"))
            }
            None => None,
        };
        for e in expand(view, from_id, etypes.as_deref(), exp_dir) {
            if row_has_edge(row, &e) {
                continue;
            }
            let nbr = neighbor(from_id, &e, *dir);
            if !view.visible(nbr) {
                continue;
            }
            if let Some(want) = bound_to {
                if nbr != want {
                    continue;
                }
            }
            if !node_matches(view, vars, row, nbr, to_label.as_deref(), to_props, params)? {
                continue;
            }
            if out.len() >= cap {
                return Err(row_cap_err(cap));
            }
            let mut next = row.clone();
            if let Some(slot) = rel_slot {
                next[slot] = Some(Cell::Rel(e));
            }
            if bound_to.is_none() {
                next[to_slot] = Some(Cell::Node(nbr));
            }
            out.push(next);
            #[cfg(test)]
            record_expand_row();
        }
    }
    Ok(out)
}

/// BFS variable-length expand with per-path edge-uniqueness (Cypher
/// relationship isomorphism).  A single path never reuses the same `EdgeRef`;
/// node revisits are allowed.
///
/// Emits one row per (start, end, depth) combination where `min ≤ depth ≤ max`.
/// The 1 M intermediate-row budget applies to `out.len()`.
#[allow(clippy::too_many_arguments)]
fn exec_var_expand(
    view: &GraphView,
    vars: &VarTable,
    rows: &[Row],
    from: &str,
    rel_var: &Option<String>,
    etypes: &[String],
    dir: RelDir,
    to: &str,
    min: u8,
    max: u8,
) -> Result<Vec<Row>, String> {
    let etypes = resolve_etypes(view, etypes);
    let exp_dir = map_dir(dir);
    let to_slot = vars
        .slot(to)
        .ok_or_else(|| format!("unbound variable `{to}`"))?;
    let rel_slot = rel_var.as_ref().and_then(|rv| vars.slot(rv));
    let cap = max_intermediate_rows();
    let mut out: Vec<Row> = Vec::new();

    for row in rows {
        let from_id = require_node(row, vars, from)?;
        let bound_to = match row.get(to_slot).and_then(|c| c.as_ref()) {
            Some(Cell::Node(id)) => Some(*id),
            Some(_) => return Err(format!("variable `{to}` is not a node")),
            None => None,
        };

        // BFS state: (current_node_id, edges_used_in_this_path).
        // Vec<EdgeRef> is cheap for paths ≤ 10 edges; contains() is O(max) = O(10).
        struct PathState {
            node: u32,
            edges: Vec<EdgeRef>,
        }

        let mut frontier: Vec<PathState> = vec![PathState {
            node: from_id,
            edges: Vec::new(),
        }];

        // Running count of all PathStates ever retained in the frontier (across all
        // depths and all input rows).  Counted even during the pre-emission phase
        // (depth < min) so that high-min queries on dense graphs cannot exhaust RAM
        // before the budget fires.
        let mut frontier_count: usize = 0;

        for depth in 1u8..=max {
            let mut next_frontier: Vec<PathState> = Vec::new();
            for state in &frontier {
                for e in expand(view, state.node, etypes.as_deref(), exp_dir) {
                    // Per-path edge-uniqueness: reject edges already used in this path.
                    if state.edges.contains(&e) {
                        continue;
                    }
                    let nbr = neighbor(state.node, &e, dir);
                    // Hidden nodes are non-existent under the mask — neither emit
                    // nor continue expanding through them.
                    if !view.visible(nbr) {
                        continue;
                    }

                    // Emit a result row if we're within the requested depth range
                    // and the destination matches any bound constraint.
                    if depth >= min {
                        let dest_matches = match bound_to {
                            Some(want) => nbr == want,
                            None => true,
                        };
                        if dest_matches {
                            if out.len() >= cap {
                                return Err(row_cap_err(cap));
                            }
                            let mut next = row.clone();
                            if let Some(slot) = rel_slot {
                                next[slot] = Some(Cell::Path(depth));
                            }
                            next[to_slot] = Some(Cell::Node(nbr));
                            out.push(next);
                        }
                    }

                    // Continue expanding if we haven't hit the max depth yet.
                    // Count each retained PathState against the budget so that
                    // high-min queries on dense graphs are caught before emission.
                    if depth < max {
                        frontier_count += 1;
                        if frontier_count >= cap {
                            return Err(row_cap_err(cap));
                        }
                        let mut new_edges = state.edges.clone();
                        new_edges.push(e);
                        next_frontier.push(PathState {
                            node: nbr,
                            edges: new_edges,
                        });
                    }
                }
            }
            frontier = next_frontier;
            if frontier.is_empty() {
                break;
            }
        }
    }
    Ok(out)
}

/// BFS shortest-path between two already-bound nodes.
///
/// Uses standard BFS with a visited-node set for efficiency.  Since BFS
/// expands nodes level-by-level, the first time `to` is reached is the
/// shortest path.  Emits exactly 0 or 1 rows.
#[allow(clippy::too_many_arguments)]
fn exec_shortest_path(
    view: &GraphView,
    vars: &VarTable,
    rows: &[Row],
    from: &str,
    rel_var: &Option<String>,
    etypes: &[String],
    dir: RelDir,
    to: &str,
    max_hops: u8,
) -> Result<Vec<Row>, String> {
    let etypes = resolve_etypes(view, etypes);
    let exp_dir = map_dir(dir);
    let rel_slot = rel_var.as_ref().and_then(|rv| vars.slot(rv));
    let mut out: Vec<Row> = Vec::new();

    for row in rows {
        let from_id = require_node(row, vars, from)?;
        let to_id = require_node(row, vars, to)?;

        // BFS with visited-node tracking.
        let mut visited = std::collections::BTreeSet::new();
        visited.insert(from_id);
        let mut frontier: Vec<u32> = vec![from_id];

        'bfs: for depth in 1u8..=max_hops {
            let mut next_frontier: Vec<u32> = Vec::new();
            for &node in &frontier {
                for e in expand(view, node, etypes.as_deref(), exp_dir) {
                    let nbr = neighbor(node, &e, dir);
                    // Hidden nodes are non-existent under the mask — do not traverse
                    // through them (paths through hidden nodes do not exist).
                    // `to_id` is guaranteed visible by its prior masked binding.
                    if !view.visible(nbr) {
                        continue;
                    }
                    if nbr == to_id {
                        // Found the shortest path at this depth — emit one row and stop.
                        let mut next = row.clone();
                        if let Some(slot) = rel_slot {
                            next[slot] = Some(Cell::Path(depth));
                        }
                        out.push(next);
                        break 'bfs;
                    }
                    if !visited.contains(&nbr) {
                        visited.insert(nbr);
                        next_frontier.push(nbr);
                    }
                }
            }
            frontier = next_frontier;
            if frontier.is_empty() {
                break;
            }
        }
    }
    Ok(out)
}

fn exec_filter(
    view: &GraphView,
    vars: &VarTable,
    rows: &[Row],
    expr: &Expr,
    params: &Params,
) -> Result<Vec<Row>, String> {
    let mut out = Vec::with_capacity(rows.len());
    for row in rows {
        if eval_expr(view, vars, row, expr, params, 0)? {
            out.push(row.clone());
        }
    }
    Ok(out)
}

/// Demand-driven executor for bounded plans (LIMIT without ORDER BY).
///
/// Locates the `Project` op, extracts the producer slice, and drives
/// `pull_rows` to collect up to `bound` (= SKIP + LIMIT) projected rows.
/// The 1 M intermediate-row cap is **not** applied here — it is the safety
/// net for the staged (unbounded) path only.
///
/// # PARTIAL-row caveat
///
/// `bound` counts *final projected* rows (post-filter, post-uniqueness).
/// A source row that is dropped by a Filter or by relationship-uniqueness
/// does **not** count toward the bound.  Execution may therefore visit
/// slightly more source rows than the bare LIMIT number, but will never
/// emit more than `bound` result rows.
/// Immutable context shared across all `pull_rows` recursive calls.
struct PullCtx<'a> {
    view: &'a GraphView<'a>,
    vars: &'a VarTable,
    project_items: &'a [RetItem],
    params: &'a Params<'a>,
    bound: usize,
}

fn execute_pull(
    view: &GraphView,
    plan: &[PlanOp],
    params: &Params,
    bound: usize,
) -> Result<ResultSet, String> {
    let proj_pos = match plan
        .iter()
        .position(|op| matches!(op, PlanOp::Project { .. }))
    {
        Some(p) => p,
        None => return Ok(ResultSet::new(vec![])),
    };
    let producers = &plan[..proj_pos];
    let project_items = match &plan[proj_pos] {
        PlanOp::Project { items } => items,
        _ => unreachable!(),
    };
    let columns: Vec<String> = project_items.iter().map(column_name).collect();
    let vars = collect_vars(plan);
    let ctx = PullCtx {
        view,
        vars: &vars,
        project_items,
        params,
        bound,
    };
    let mut initial_row: Row = vec![None; vars.names.len()];
    let mut result_rows: Vec<Vec<Option<Value>>> = Vec::with_capacity(bound);
    pull_rows(&ctx, producers, &mut initial_row, &mut result_rows)?;
    // SKIP: discard the leading rows (bound = SKIP+LIMIT ensures there are enough).
    // Resolve SKIP — params must have been validated by check_params already.
    let skip_n = plan[proj_pos + 1..]
        .iter()
        .find_map(|op| match op {
            PlanOp::Skip(ls) => Some(resolve_ls(ls, params)),
            _ => None,
        })
        .transpose()?
        .unwrap_or(0);
    let skip_n = usize::try_from(skip_n).unwrap_or(usize::MAX);
    let mut rs = ResultSet::new(columns);
    for row in result_rows.into_iter().skip(skip_n) {
        rs.push_row(row);
    }
    Ok(rs)
}

// ─── Aggregate execution path ────────────────────────────────────────────────
//
// Aggregate plans stream through ALL matching rows one at a time and maintain
// a single accumulator value.  Memory is O(1) regardless of graph size:
//
//   - No binding table is materialised.
//   - The 1 M intermediate-row budget does **not** apply.  Applying it would
//     produce wrong counts/sums on large graphs and is unnecessary here because
//     memory is bounded by the accumulator, not by the number of rows.
//
// v1 scope: exactly one aggregate function per query, no grouping keys.

/// Running state for a single aggregate accumulation.
enum AggAcc {
    Count(u64),
    /// `sum`: integers add exactly as `i64` (an overflow is an error); once a
    /// float is seen the result is a float (row 158).
    Sum {
        ints: i64,
        floats: f64,
        any_float: bool,
        has_value: bool,
    },
    Avg {
        sum: f64,
        n: u64,
    },
    Min(Option<Value>),
    Max(Option<Value>),
    Collect(Vec<Value>),
    /// `DISTINCT` wrapper: forwards to `inner` the first time a value is
    /// seen in this group and drops every repeat.
    Distinct {
        seen: HashSet<Option<ValueKey>>,
        inner: Box<AggAcc>,
    },
}

impl AggAcc {
    /// The accumulator an aggregate needs given the argument it was written
    /// with — a `DISTINCT` argument adds the seen-set around the plain one.
    fn for_arg(func: &AggFunc, arg: &AggArg) -> Self {
        match arg {
            AggArg::Distinct(_) => AggAcc::Distinct {
                seen: HashSet::new(),
                inner: Box::new(AggAcc::new(func)),
            },
            _ => AggAcc::new(func),
        }
    }

    fn new(func: &AggFunc) -> Self {
        match func {
            AggFunc::Count => AggAcc::Count(0),
            AggFunc::Sum => AggAcc::Sum {
                ints: 0,
                floats: 0.0,
                any_float: false,
                has_value: false,
            },
            AggFunc::Avg => AggAcc::Avg { sum: 0.0, n: 0 },
            AggFunc::Min => AggAcc::Min(None),
            AggFunc::Max => AggAcc::Max(None),
            AggFunc::Collect => AggAcc::Collect(Vec::new()),
        }
    }

    fn finish(self) -> Option<Value> {
        match self {
            // Saturating cast: a graph with >i64::MAX matched rows is not a
            // realistic concern today, but a silent wrapping cast would produce
            // a wrong (negative) count. Clamping to i64::MAX is the least-
            // surprising failure mode.
            AggAcc::Count(n) => Some(Value::Int(i64::try_from(n).unwrap_or(i64::MAX))),
            AggAcc::Sum {
                ints,
                floats,
                any_float,
                has_value,
            } => {
                if !has_value {
                    None
                } else if any_float {
                    Some(Value::Float(ints as f64 + floats))
                } else {
                    Some(Value::Int(ints))
                }
            }
            AggAcc::Avg { sum, n } => {
                if n > 0 {
                    Some(Value::Float(sum / n as f64))
                } else {
                    None
                }
            }
            AggAcc::Min(v) => v,
            AggAcc::Max(v) => v,
            // Empty collect() yields an empty list (never null), matching openCypher.
            AggAcc::Collect(items) => Some(Value::List(items)),
            AggAcc::Distinct { inner, .. } => inner.finish(),
        }
    }
}

/// Extract a numeric (f64) value from a `Value`, returning `None` for null /
/// non-numeric types.  Silently skipped per the aggregate contract.
fn numeric_val(v: &Value) -> Option<f64> {
    match v {
        Value::Int(n) => Some(*n as f64),
        Value::Float(f) => Some(*f),
        _ => None,
    }
}

/// Shared context for `agg_stream`.
struct AggStreamCtx<'a> {
    view: &'a GraphView<'a>,
    vars: &'a VarTable,
    params: &'a Params<'a>,
    func: &'a AggFunc,
    arg: &'a AggArg,
}

/// Streaming accumulator executor for a single aggregate.
///
/// Locates the `Aggregate` op, builds the producer slice (everything before
/// it), then calls `agg_stream` to walk all matching rows and accumulate.
fn execute_aggregate(
    view: &GraphView,
    plan: &[PlanOp],
    params: &Params,
) -> Result<ResultSet, String> {
    let agg_pos = match plan
        .iter()
        .position(|op| matches!(op, PlanOp::Aggregate { .. }))
    {
        Some(p) => p,
        None => return Ok(ResultSet::new(vec![])),
    };
    let producers = &plan[..agg_pos];
    let (func, arg, column) = match &plan[agg_pos] {
        PlanOp::Aggregate { func, arg, column } => (func, arg, column),
        _ => unreachable!(),
    };
    let vars = collect_vars(plan);
    let ctx = AggStreamCtx {
        view,
        vars: &vars,
        params,
        func,
        arg,
    };
    let initial_row: Row = vec![None; vars.names.len()];
    let mut acc = AggAcc::for_arg(func, arg);
    agg_stream(&ctx, producers, &initial_row, &mut acc)?;

    let value = acc.finish();
    let mut rs = ResultSet::new(vec![column.clone()]);
    rs.push_row(vec![value]);
    Ok(rs)
}

/// The value a `DISTINCT` aggregate argument takes on one row.
///
/// A node variable is identified by its key, a relationship variable by the
/// edge it is bound to, a scalar alias by its value, and a property by the
/// property's value. `DISTINCT *` is rejected at parse time, so `Star` is
/// unreachable and yields `None` defensively.
///
/// # A relationship is keyed on the edge, not dropped
///
/// A relationship cell has no key of its own, so this used to answer `None`
/// for one — and `None` means "contributes to no aggregate", so
/// `count(DISTINCT r)` counted nothing at all and returned `0` on a graph
/// full of edges. An edge *is* identified, by the triple the executor binds:
/// its type and its two endpoints. That is what is keyed here, so
/// `count(DISTINCT r)` equals `count(r)` on a graph where no pair is joined
/// twice by one type, and is smaller exactly where an alternation matched the
/// same edge through more than one pattern.
///
/// The key is only ever compared against other keys from this same function
/// within one `DISTINCT` gate, and an aggregate argument is one variable, so
/// a relationship's key never meets a node key or a property string: the
/// separator is there to keep the three ids apart, not to fence off a
/// collision that the grammar can produce.
fn distinct_value(
    view: &GraphView,
    vars: &VarTable,
    row: &Row,
    arg: &AggArg,
) -> Result<Option<Value>, String> {
    match arg {
        AggArg::Star => Ok(None),
        AggArg::Var(v) => {
            let Some(slot) = vars.slot(v) else {
                return Ok(None);
            };
            Ok(match row.get(slot).and_then(|c| c.as_ref()) {
                Some(Cell::Node(id)) => view.ids.key_of(*id).map(|k| Value::Str(k.to_owned())),
                Some(Cell::Scalar(val)) => Some(val.clone()),
                Some(Cell::Path(hops)) => Some(Value::Int(*hops as i64)),
                Some(Cell::Rel(e)) => Some(Value::Str(format!(
                    "{}\u{1}{}\u{1}{}",
                    e.etype, e.src, e.dst
                ))),
                None => None,
            })
        }
        AggArg::Prop { var, field } => resolve_prop(view, vars, row, var, field),
        // The parser never nests DISTINCT; treat a nested one as its inner arg.
        AggArg::Distinct(inner) => distinct_value(view, vars, row, inner),
    }
}

/// The value one row feeds a `sum` / `avg` / `min` / `max`: the property, or
/// the scalar a bare variable is bound to. Null, an unbound variable and a
/// node, relationship or path variable give `None` and contribute nothing, as
/// every non-property argument did before row 157.
fn scalar_agg_input(
    view: &GraphView,
    vars: &VarTable,
    row: &Row,
    arg: &AggArg,
) -> Result<Option<Value>, String> {
    Ok(match arg {
        AggArg::Prop { var, field } => resolve_prop(view, vars, row, var, field)?,
        AggArg::Var(v) => match vars
            .slot(v)
            .and_then(|s| row.get(s))
            .and_then(|c| c.as_ref())
        {
            Some(Cell::Scalar(x)) => Some(x.clone()),
            _ => None,
        },
        _ => None,
    })
}

/// Keep the smaller (`largest == false`) or larger of the running extreme and
/// the next value. A number or a boolean counts (row 160: `false` before
/// `true`); any other value is skipped.
fn fold_extreme(current: &mut Option<Value>, next: Value, largest: bool) {
    if numeric_val(&next).is_none() && !matches!(next, Value::Bool(_)) {
        return;
    }
    *current = Some(match current.take() {
        None => next,
        Some(prev) => {
            if cmp_optional(Some(&prev), Some(&next), largest) == std::cmp::Ordering::Greater {
                next
            } else {
                prev
            }
        }
    });
}

/// Update a single accumulator for one matched row.
///
/// Shared by `agg_stream` (single-aggregate path) and `group_stream` (grouped
/// aggregation).  The logic mirrors openCypher conventions: null / non-numeric
/// values are silently skipped for SUM / AVG / MIN / MAX.
fn update_acc(
    view: &GraphView,
    vars: &VarTable,
    row: &Row,
    func: &AggFunc,
    arg: &AggArg,
    acc: &mut AggAcc,
) -> Result<(), String> {
    // DISTINCT gate: resolve the row's value for the inner argument, drop the
    // row when this group has already seen it, and otherwise forward to the
    // plain accumulator underneath.
    if let AggArg::Distinct(inner_arg) = arg {
        let AggAcc::Distinct { seen, inner } = acc else {
            return Ok(());
        };
        let val = distinct_value(view, vars, row, inner_arg)?;
        // An unbound / null argument contributes to no aggregate, DISTINCT or
        // not: `count(DISTINCT x)` counts distinct non-null values.
        let Some(val) = val else {
            return Ok(());
        };
        if !seen.insert(group_key_normalize(&val)) {
            return Ok(());
        }
        return update_acc(view, vars, row, func, inner_arg, inner);
    }
    match (func, arg) {
        (AggFunc::Count, AggArg::Star) => {
            if let AggAcc::Count(n) = acc {
                *n += 1;
            }
        }
        (AggFunc::Count, AggArg::Var(v)) => {
            let slot = vars.slot(v);
            let is_bound = slot
                .and_then(|s| row.get(s))
                .and_then(|c| c.as_ref())
                .is_some();
            if is_bound {
                if let AggAcc::Count(n) = acc {
                    *n += 1;
                }
            }
        }
        (AggFunc::Count, AggArg::Prop { var, field }) => {
            let val = resolve_prop(view, vars, row, var, field)?;
            if val.is_some() {
                if let AggAcc::Count(n) = acc {
                    *n += 1;
                }
            }
        }
        // sum / avg / min / max read one value per row: a property or a bare
        // variable bound to a scalar (`UNWIND … AS x`, `WITH … AS x`). Row 157:
        // the bare-variable form used to fall through to the catch-all arm and
        // contribute nothing, so these returned null.
        (
            AggFunc::Sum | AggFunc::Avg | AggFunc::Min | AggFunc::Max,
            AggArg::Prop { .. } | AggArg::Var(_),
        ) => {
            let Some(v) = scalar_agg_input(view, vars, row, arg)? else {
                return Ok(());
            };
            match acc {
                AggAcc::Sum {
                    ints,
                    floats,
                    any_float,
                    has_value,
                } => match v {
                    Value::Int(n) => {
                        *ints = ints
                            .checked_add(n)
                            .ok_or_else(|| "sum(): integer overflow".to_string())?;
                        *has_value = true;
                    }
                    Value::Float(f) => {
                        *floats += f;
                        *any_float = true;
                        *has_value = true;
                    }
                    _ => {}
                },
                AggAcc::Avg { sum, n } => {
                    if let Some(num) = numeric_val(&v) {
                        *sum += num;
                        *n += 1;
                    }
                }
                AggAcc::Min(current) => fold_extreme(current, v, false),
                AggAcc::Max(current) => fold_extreme(current, v, true),
                _ => {}
            }
        }
        (AggFunc::Collect, AggArg::Var(v)) => {
            // Gather each row's value of `v` — node → key, scalar alias → value,
            // path → hop count. Nulls (unbound / relationship) are skipped.
            let val = vars
                .slot(v)
                .and_then(|s| row.get(s))
                .and_then(|c| c.as_ref())
                .and_then(|cell| match cell {
                    Cell::Scalar(x) => Some(x.clone()),
                    Cell::Node(id) => view.ids.key_of(*id).map(|k| Value::Str(k.to_owned())),
                    Cell::Path(h) => Some(Value::Int(*h as i64)),
                    Cell::Rel(_) => None,
                });
            if let Some(val) = val {
                if let AggAcc::Collect(items) = acc {
                    items.push(val);
                }
            }
        }
        (AggFunc::Collect, AggArg::Prop { var, field }) => {
            if let Some(val) = resolve_prop(view, vars, row, var, field)? {
                if let AggAcc::Collect(items) = acc {
                    items.push(val);
                }
            }
        }
        // Remaining combinations are rejected by the planner (e.g., SUM(*))
        // but handle defensively without panic.
        _ => {}
    }
    Ok(())
}

/// Recursively walk all matching rows through `ops`, updating `acc` for each
/// terminal row.  No bound — visits every row the producers can emit.
fn agg_stream(
    ctx: &AggStreamCtx<'_>,
    ops: &[PlanOp],
    row: &Row,
    acc: &mut AggAcc,
) -> Result<(), String> {
    let (op, rest) = match ops.split_first() {
        Some(pair) => pair,
        None => {
            // Terminal row — delegate to shared update_acc helper.
            update_acc(ctx.view, ctx.vars, row, ctx.func, ctx.arg, acc)?;
            return Ok(());
        }
    };

    match op {
        PlanOp::ScanLabel { var, label } => {
            let ids = scan_ids(ctx.view, label.as_deref());
            let slot = ctx
                .vars
                .slot(var)
                .ok_or_else(|| format!("unbound variable `{var}`"))?;
            for &id in &ids {
                let mut next = row.clone();
                next[slot] = Some(Cell::Node(id));
                agg_stream(ctx, rest, &next, acc)?;
            }
        }
        PlanOp::ScanKey { var, key, label } => {
            let slot = ctx
                .vars
                .slot(var)
                .ok_or_else(|| format!("unbound variable `{var}`"))?;
            if let Some(id) =
                resolve_scan_key_id(ctx.view, ctx.vars, row, key, label.as_deref(), ctx.params)?
            {
                let mut next = row.clone();
                next[slot] = Some(Cell::Node(id));
                agg_stream(ctx, rest, &next, acc)?;
            }
        }
        PlanOp::IndexScan {
            var,
            label,
            field,
            value,
        } => {
            let ids = index_scan_ids(
                ctx.view,
                ctx.vars,
                row,
                label.as_deref(),
                field,
                value,
                ctx.params,
            )?;
            let slot = ctx
                .vars
                .slot(var)
                .ok_or_else(|| format!("unbound variable `{var}`"))?;
            for &id in &ids {
                let mut next = row.clone();
                next[slot] = Some(Cell::Node(id));
                agg_stream(ctx, rest, &next, acc)?;
            }
        }
        PlanOp::IndexIntersect {
            var,
            label,
            equalities,
        } => {
            let ids = index_intersect_ids(
                ctx.view,
                ctx.vars,
                row,
                label.as_deref(),
                equalities,
                ctx.params,
            )?;
            let slot = ctx
                .vars
                .slot(var)
                .ok_or_else(|| format!("unbound variable `{var}`"))?;
            for id in ids {
                let mut next = row.clone();
                next[slot] = Some(Cell::Node(id));
                agg_stream(ctx, rest, &next, acc)?;
            }
        }
        PlanOp::Expand {
            from,
            rel_var,
            etypes,
            dir,
            to,
            to_label,
            to_props,
        } => {
            let etypes = resolve_etypes(ctx.view, etypes);
            let exp_dir = map_dir(*dir);
            let to_slot = ctx
                .vars
                .slot(to)
                .ok_or_else(|| format!("unbound variable `{to}`"))?;
            let rel_slot = rel_var.as_ref().and_then(|rv| ctx.vars.slot(rv));
            let from_id = require_node(row, ctx.vars, from)?;
            let bound_to = match row.get(to_slot).and_then(|c| c.as_ref()) {
                Some(Cell::Node(id)) => Some(*id),
                Some(Cell::Rel(_) | Cell::Path(_) | Cell::Scalar(_)) => {
                    return Err(format!("variable `{to}` is not a node"))
                }
                None => None,
            };
            for e in expand(ctx.view, from_id, etypes.as_deref(), exp_dir) {
                if row_has_edge(row, &e) {
                    continue;
                }
                let nbr = neighbor(from_id, &e, *dir);
                if !ctx.view.visible(nbr) {
                    continue;
                }
                if let Some(want) = bound_to {
                    if nbr != want {
                        continue;
                    }
                }
                if !node_matches(
                    ctx.view,
                    ctx.vars,
                    row,
                    nbr,
                    to_label.as_deref(),
                    to_props,
                    ctx.params,
                )? {
                    continue;
                }
                let mut next = row.clone();
                if let Some(slot) = rel_slot {
                    next[slot] = Some(Cell::Rel(e));
                }
                if bound_to.is_none() {
                    next[to_slot] = Some(Cell::Node(nbr));
                }
                agg_stream(ctx, rest, &next, acc)?;
            }
        }
        PlanOp::Filter { expr } => {
            if eval_expr(ctx.view, ctx.vars, row, expr, ctx.params, 0)? {
                agg_stream(ctx, rest, row, acc)?;
            }
        }
        PlanOp::LookupProps { var, props } => {
            let id = require_node(row, ctx.vars, var)?;
            if node_matches(ctx.view, ctx.vars, row, id, None, props, ctx.params)? {
                agg_stream(ctx, rest, row, acc)?;
            }
        }
        PlanOp::JoinBound { var, label, props } => {
            let id = require_node(row, ctx.vars, var)?;
            if node_matches(
                ctx.view,
                ctx.vars,
                row,
                id,
                label.as_deref(),
                props,
                ctx.params,
            )? {
                agg_stream(ctx, rest, row, acc)?;
            }
        }
        PlanOp::VarExpand {
            from,
            rel_var,
            etypes,
            dir,
            to,
            min,
            max,
        } => {
            let new_rows = exec_var_expand(
                ctx.view,
                ctx.vars,
                std::slice::from_ref(row),
                from,
                rel_var,
                etypes,
                *dir,
                to,
                *min,
                *max,
            )?;
            for nr in &new_rows {
                agg_stream(ctx, rest, nr, acc)?;
            }
        }
        PlanOp::ShortestPath {
            from,
            rel_var,
            etypes,
            dir,
            to,
            max_hops,
        } => {
            let new_rows = exec_shortest_path(
                ctx.view,
                ctx.vars,
                std::slice::from_ref(row),
                from,
                rel_var,
                etypes,
                *dir,
                to,
                *max_hops,
            )?;
            for nr in &new_rows {
                agg_stream(ctx, rest, nr, acc)?;
            }
        }
        // These must not appear in the producer slice of an aggregate plan.
        PlanOp::Project { .. } => {
            return Err(
                "agg executor: Project in producer slice — plan is structurally malformed"
                    .to_string(),
            );
        }
        PlanOp::Distinct => {
            return Err(
                "agg executor: Distinct in producer slice — plan is structurally malformed"
                    .to_string(),
            );
        }
        PlanOp::OrderBy { .. } => {
            return Err(
                "agg executor: OrderBy in producer slice — structurally malformed".to_string(),
            );
        }
        PlanOp::Skip(_) => {
            return Err(
                "agg executor: Skip in producer slice — structurally malformed".to_string(),
            );
        }
        PlanOp::Limit(_) => {
            return Err(
                "agg executor: Limit in producer slice — structurally malformed".to_string(),
            );
        }
        PlanOp::Aggregate { .. } => {
            return Err(
                "agg executor: nested Aggregate in producer slice — structurally malformed"
                    .to_string(),
            );
        }
        PlanOp::GroupAggregate { .. } => {
            return Err(
                "agg executor: GroupAggregate in producer slice — structurally malformed"
                    .to_string(),
            );
        }
        PlanOp::With { .. } => {
            return Err(
                "agg executor: With in producer slice — structurally malformed".to_string(),
            );
        }
        PlanOp::Unwind { .. } => {
            return Err(
                "agg executor: Unwind in producer slice — structurally malformed".to_string(),
            );
        }
        PlanOp::LeftOuterApply { .. } => {
            return Err(
                "agg executor: LeftOuterApply in producer slice — structurally malformed"
                    .to_string(),
            );
        }
    }
    Ok(())
}

// ─── GroupAggregate execution path ──────────────────────────────────────────
//
// GroupAggregate plans stream through ALL matching rows, computing a group-key
// tuple per row and maintaining per-group accumulators in a HashMap.  Memory
// is O(distinct groups); the 1 M intermediate-row budget does not apply.

/// Context shared across all `group_stream` recursive calls.
struct GroupStreamCtx<'a> {
    view: &'a GraphView<'a>,
    vars: &'a VarTable,
    params: &'a Params<'a>,
    keys: &'a [(String, RetItem)],
    aggs: &'a [(AggFunc, AggArg, String)],
}

/// Build the `Projected` output table from a finished group map.
fn build_group_projected(
    keys: &[(String, RetItem)],
    aggs: &[(AggFunc, AggArg, String)],
    key_order: Vec<GroupKey>,
    groups: &mut HashMap<GroupKey, GroupEntry>,
) -> Projected {
    let columns: Vec<String> = keys
        .iter()
        .map(|(col, _)| col.clone())
        .chain(aggs.iter().map(|(_, _, col)| col.clone()))
        .collect();
    let mut rows: Vec<Vec<Option<Value>>> = Vec::with_capacity(key_order.len());
    for gk in key_order {
        let (display_keys, accs) = groups.remove(&gk).unwrap_or_default();
        let mut row: Vec<Option<Value>> = Vec::with_capacity(columns.len());
        // Use the first-seen original values for display; Int(42) stays Int(42)
        // even though it was normalized to FloatBits for hashing.
        for display_val in display_keys {
            row.push(display_val);
        }
        for acc in accs {
            row.push(acc.finish());
        }
        rows.push(row);
    }
    Projected { columns, rows }
}

/// Convert a finished group map into raw `Row` vectors for pipeline consumption.
///
/// Used by the staged executor when a `GroupAggregate` appears in a WITH pipeline
/// (i.e., `is_pipeline = true`).  Each group becomes one `Row` with `Cell::Scalar`
/// values for key and aggregate columns.  Column names are interned in `vars` so
/// that subsequent pipeline stages can look them up by slot.
/// The original cell behind each grouping key that is a bare variable.
///
/// Only `RetVal::Var` keys carry a cell through: a computed key (`c.city`,
/// `toLower(c.name)`) has no node behind it and stays a scalar.
fn key_source_cells(vars: &VarTable, row: &Row, keys: &[(String, RetItem)]) -> Vec<Option<Cell>> {
    keys.iter()
        .map(|(_, item)| match &item.value {
            RetVal::Var(v) => vars
                .slot(v)
                .and_then(|s| row.get(s))
                .and_then(|c| c.clone()),
            _ => None,
        })
        .collect()
}

fn group_result_to_rows(
    keys: &[(String, RetItem)],
    aggs: &[(AggFunc, AggArg, String)],
    key_order: Vec<GroupKey>,
    groups: &mut HashMap<GroupKey, GroupEntry>,
    key_cells: &mut HashMap<GroupKey, Vec<Option<Cell>>>,
    vars: &VarTable,
) -> Vec<Row> {
    let row_len = vars.names.len();
    let mut out: Vec<Row> = Vec::with_capacity(key_order.len());
    for gk in key_order {
        let (display_keys, accs) = groups.remove(&gk).unwrap_or_default();
        let mut cells = key_cells.remove(&gk).unwrap_or_default();
        cells.resize(keys.len(), None);
        let mut row: Row = vec![None; row_len];
        for (((col, _), val), cell) in keys.iter().zip(display_keys).zip(cells) {
            if let Some(slot) = vars.slot(col) {
                row[slot] = cell.or_else(|| val.map(Cell::Scalar));
            }
        }
        for ((_, _, col), acc) in aggs.iter().zip(accs) {
            if let Some(slot) = vars.slot(col) {
                row[slot] = acc.finish().map(Cell::Scalar);
            }
        }
        out.push(row);
    }
    out
}

/// Sort raw rows by a list of `OrderItem`s (before projection).
///
/// Used by the staged path when an `OrderBy` op appears before `Project` —
/// for example, inside a non-aggregate WITH stage or after a pipeline
/// GroupAggregate.  `OrderTarget::Prop` is resolved by looking up the node
/// property; `Alias` / `Var` are resolved from `Cell::Scalar` in the row.
fn exec_order_by_rows(vars: &VarTable, rows: &mut Vec<Row>, items: &[OrderItem], view: &GraphView) {
    // Pre-compute sort-key values for all rows.  This avoids re-resolving
    // inside the comparator (the closure cannot return Err, so we pre-compute).
    let mut key_table: Vec<Vec<Option<Value>>> = Vec::with_capacity(rows.len());
    for row in rows.iter() {
        let mut row_key: Vec<Option<Value>> = Vec::with_capacity(items.len());
        for item in items {
            let val = match &item.target {
                OrderTarget::Alias(name) | OrderTarget::Var(name) => vars
                    .slot(name)
                    .and_then(|s| row.get(s))
                    .and_then(|c| c.as_ref())
                    .and_then(|c| match c {
                        Cell::Scalar(v) => Some(v.clone()),
                        Cell::Node(id) => view.ids.key_of(*id).map(|k| Value::Str(k.to_owned())),
                        Cell::Path(hops) => Some(Value::Int(*hops as i64)),
                        Cell::Rel(_) => None,
                    }),
                OrderTarget::Prop { var, field } => vars
                    .slot(var)
                    .and_then(|s| row.get(s))
                    .and_then(|c| c.as_ref())
                    .and_then(|c| match c {
                        Cell::Node(id) => view
                            .prop(*id, field)
                            .map(|vr| vr.into_value())
                            .or_else(|| node_identity_prop(view, *id, field)),
                        Cell::Rel(e) => view.edge_props.get(e.etype, e.src, e.dst, field),
                        _ => None,
                    }),
            };
            row_key.push(val);
        }
        key_table.push(row_key);
    }
    // Sort using a stable, index-based sort to keep equal rows in encounter order.
    let mut indices: Vec<usize> = (0..rows.len()).collect();
    indices.sort_by(|&a, &b| {
        for (ki, item) in items.iter().enumerate() {
            let c = cmp_optional(
                key_table[a].get(ki).and_then(|x| x.as_ref()),
                key_table[b].get(ki).and_then(|x| x.as_ref()),
                item.descending,
            );
            if c != std::cmp::Ordering::Equal {
                return c;
            }
        }
        std::cmp::Ordering::Equal
    });
    let sorted: Vec<Row> = indices.into_iter().map(|i| rows[i].clone()).collect();
    *rows = sorted;
}

/// Streaming grouped-aggregate executor.
///
/// Locates the `GroupAggregate` op, builds the producer slice (everything
/// before it), streams through all matching rows via `group_stream`, then
/// applies any `OrderBy` / `Skip` / `Limit` ops that follow.
fn execute_group_aggregate(
    view: &GraphView,
    plan: &[PlanOp],
    params: &Params,
) -> Result<ResultSet, String> {
    let gagg_pos = plan
        .iter()
        .position(|op| matches!(op, PlanOp::GroupAggregate { .. }))
        .ok_or_else(|| "internal: GroupAggregate op not found in plan".to_string())?;
    let producers = &plan[..gagg_pos];
    let (keys, aggs) = match &plan[gagg_pos] {
        PlanOp::GroupAggregate { keys, aggs } => (keys, aggs),
        _ => unreachable!(),
    };
    let tail = &plan[gagg_pos + 1..];
    let vars = collect_vars(plan);
    let initial_row: Row = vec![None; vars.names.len()];
    let mut groups: HashMap<GroupKey, GroupEntry> = HashMap::new();
    let mut key_order: Vec<GroupKey> = Vec::new();
    let ctx = GroupStreamCtx {
        view,
        vars: &vars,
        params,
        keys,
        aggs,
    };
    group_stream(&ctx, producers, &initial_row, &mut groups, &mut key_order)?;
    // openCypher: aggregates with no group-key items on empty input must
    // produce exactly one row (COUNT=0, SUM/AVG/etc.=null).  Seed the empty
    // key when no terminal rows arrived and there are no grouping keys.
    if keys.is_empty() && key_order.is_empty() {
        let empty_key: GroupKey = vec![];
        key_order.push(empty_key.clone());
        groups.insert(
            empty_key,
            (
                vec![],
                aggs.iter().map(|(f, a, _)| AggAcc::for_arg(f, a)).collect(),
            ),
        );
    }
    let mut projected = build_group_projected(keys, aggs, key_order, &mut groups);
    // Apply OrderBy / Skip / Limit from the tail of the plan.
    for op in tail {
        match op {
            PlanOp::OrderBy { items } => exec_order_by(&mut projected, items)?,
            PlanOp::Skip(ls) => {
                let n = resolve_ls(ls, params)?;
                apply_skip(&mut projected.rows, n);
            }
            PlanOp::Limit(ls) => {
                let n = resolve_ls(ls, params)?;
                apply_limit(&mut projected.rows, n);
            }
            _ => {} // Ignore unexpected ops defensively.
        }
    }
    Ok(finish(projected))
}

/// Recursively walk all matching rows through `ops`, computing the group key
/// and updating per-group accumulators at each terminal row.
///
/// `groups` maps group key → per-group accumulator vector.
/// `key_order` tracks insertion order so output rows are deterministic when
/// no ORDER BY is requested.
fn group_stream(
    ctx: &GroupStreamCtx<'_>,
    ops: &[PlanOp],
    row: &Row,
    groups: &mut HashMap<GroupKey, GroupEntry>,
    key_order: &mut Vec<GroupKey>,
) -> Result<(), String> {
    let (op, rest) = match ops.split_first() {
        Some(pair) => pair,
        None => {
            // Terminal row: compute normalized group key for equality/hashing
            // and capture original values for display (first-seen wins).
            let mut gk: GroupKey = Vec::with_capacity(ctx.keys.len());
            let mut display_vals: Vec<Option<Value>> = Vec::with_capacity(ctx.keys.len());
            for (_, item) in ctx.keys {
                let val = project_item(ctx.view, ctx.vars, row, item, ctx.params)?;
                gk.push(val.as_ref().and_then(group_key_normalize));
                display_vals.push(val);
            }
            if !groups.contains_key(&gk) {
                if groups.len() >= max_groups() {
                    return Err(group_cap_err());
                }
                key_order.push(gk.clone());
                let init: Vec<AggAcc> = ctx
                    .aggs
                    .iter()
                    .map(|(f, a, _)| AggAcc::for_arg(f, a))
                    .collect();
                groups.insert(gk.clone(), (display_vals, init));
            }
            let (_, accs) = groups.get_mut(&gk).unwrap();
            for (acc, (func, arg, _)) in accs.iter_mut().zip(ctx.aggs.iter()) {
                update_acc(ctx.view, ctx.vars, row, func, arg, acc)?;
            }
            return Ok(());
        }
    };

    match op {
        PlanOp::ScanLabel { var, label } => {
            let ids = scan_ids(ctx.view, label.as_deref());
            let slot = ctx
                .vars
                .slot(var)
                .ok_or_else(|| format!("unbound variable `{var}`"))?;
            for &id in &ids {
                let mut next = row.clone();
                next[slot] = Some(Cell::Node(id));
                group_stream(ctx, rest, &next, groups, key_order)?;
            }
        }
        PlanOp::ScanKey { var, key, label } => {
            let slot = ctx
                .vars
                .slot(var)
                .ok_or_else(|| format!("unbound variable `{var}`"))?;
            if let Some(id) =
                resolve_scan_key_id(ctx.view, ctx.vars, row, key, label.as_deref(), ctx.params)?
            {
                let mut next = row.clone();
                next[slot] = Some(Cell::Node(id));
                group_stream(ctx, rest, &next, groups, key_order)?;
            }
        }
        PlanOp::IndexScan {
            var,
            label,
            field,
            value,
        } => {
            let ids = index_scan_ids(
                ctx.view,
                ctx.vars,
                row,
                label.as_deref(),
                field,
                value,
                ctx.params,
            )?;
            let slot = ctx
                .vars
                .slot(var)
                .ok_or_else(|| format!("unbound variable `{var}`"))?;
            for &id in &ids {
                let mut next = row.clone();
                next[slot] = Some(Cell::Node(id));
                group_stream(ctx, rest, &next, groups, key_order)?;
            }
        }
        PlanOp::IndexIntersect {
            var,
            label,
            equalities,
        } => {
            let ids = index_intersect_ids(
                ctx.view,
                ctx.vars,
                row,
                label.as_deref(),
                equalities,
                ctx.params,
            )?;
            let slot = ctx
                .vars
                .slot(var)
                .ok_or_else(|| format!("unbound variable `{var}`"))?;
            for id in ids {
                let mut next = row.clone();
                next[slot] = Some(Cell::Node(id));
                group_stream(ctx, rest, &next, groups, key_order)?;
            }
        }
        PlanOp::Expand {
            from,
            rel_var,
            etypes,
            dir,
            to,
            to_label,
            to_props,
        } => {
            let etypes = resolve_etypes(ctx.view, etypes);
            let exp_dir = map_dir(*dir);
            let to_slot = ctx
                .vars
                .slot(to)
                .ok_or_else(|| format!("unbound variable `{to}`"))?;
            let rel_slot = rel_var.as_ref().and_then(|rv| ctx.vars.slot(rv));
            let from_id = require_node(row, ctx.vars, from)?;
            let bound_to = match row.get(to_slot).and_then(|c| c.as_ref()) {
                Some(Cell::Node(id)) => Some(*id),
                Some(Cell::Rel(_) | Cell::Path(_) | Cell::Scalar(_)) => {
                    return Err(format!("variable `{to}` is not a node"))
                }
                None => None,
            };
            for e in expand(ctx.view, from_id, etypes.as_deref(), exp_dir) {
                if row_has_edge(row, &e) {
                    continue;
                }
                let nbr = neighbor(from_id, &e, *dir);
                if !ctx.view.visible(nbr) {
                    continue;
                }
                if let Some(want) = bound_to {
                    if nbr != want {
                        continue;
                    }
                }
                if !node_matches(
                    ctx.view,
                    ctx.vars,
                    row,
                    nbr,
                    to_label.as_deref(),
                    to_props,
                    ctx.params,
                )? {
                    continue;
                }
                let mut next = row.clone();
                if let Some(slot) = rel_slot {
                    next[slot] = Some(Cell::Rel(e));
                }
                if bound_to.is_none() {
                    next[to_slot] = Some(Cell::Node(nbr));
                }
                group_stream(ctx, rest, &next, groups, key_order)?;
            }
        }
        PlanOp::Filter { expr } => {
            if eval_expr(ctx.view, ctx.vars, row, expr, ctx.params, 0)? {
                group_stream(ctx, rest, row, groups, key_order)?;
            }
        }
        PlanOp::LookupProps { var, props } => {
            let id = require_node(row, ctx.vars, var)?;
            if node_matches(ctx.view, ctx.vars, row, id, None, props, ctx.params)? {
                group_stream(ctx, rest, row, groups, key_order)?;
            }
        }
        PlanOp::JoinBound { var, label, props } => {
            let id = require_node(row, ctx.vars, var)?;
            if node_matches(
                ctx.view,
                ctx.vars,
                row,
                id,
                label.as_deref(),
                props,
                ctx.params,
            )? {
                group_stream(ctx, rest, row, groups, key_order)?;
            }
        }
        PlanOp::VarExpand {
            from,
            rel_var,
            etypes,
            dir,
            to,
            min,
            max,
        } => {
            let new_rows = exec_var_expand(
                ctx.view,
                ctx.vars,
                std::slice::from_ref(row),
                from,
                rel_var,
                etypes,
                *dir,
                to,
                *min,
                *max,
            )?;
            for nr in &new_rows {
                group_stream(ctx, rest, nr, groups, key_order)?;
            }
        }
        PlanOp::ShortestPath {
            from,
            rel_var,
            etypes,
            dir,
            to,
            max_hops,
        } => {
            let new_rows = exec_shortest_path(
                ctx.view,
                ctx.vars,
                std::slice::from_ref(row),
                from,
                rel_var,
                etypes,
                *dir,
                to,
                *max_hops,
            )?;
            for nr in &new_rows {
                group_stream(ctx, rest, nr, groups, key_order)?;
            }
        }
        // These ops must not appear in the producer slice of a GroupAggregate plan.
        PlanOp::Project { .. } => {
            return Err(
                "group executor: Project in producer slice — structurally malformed".to_string(),
            );
        }
        PlanOp::Distinct => {
            return Err(
                "group executor: Distinct in producer slice — structurally malformed".to_string(),
            );
        }
        PlanOp::OrderBy { .. } => {
            return Err(
                "group executor: OrderBy in producer slice — structurally malformed".to_string(),
            );
        }
        PlanOp::Skip(_) => {
            return Err(
                "group executor: Skip in producer slice — structurally malformed".to_string(),
            );
        }
        PlanOp::Limit(_) => {
            return Err(
                "group executor: Limit in producer slice — structurally malformed".to_string(),
            );
        }
        PlanOp::Aggregate { .. } => {
            return Err(
                "group executor: Aggregate in producer slice — structurally malformed".to_string(),
            );
        }
        PlanOp::GroupAggregate { .. } => {
            return Err(
                "group executor: nested GroupAggregate in producer slice — structurally malformed"
                    .to_string(),
            );
        }
        PlanOp::With { .. } => {
            return Err(
                "group executor: With in producer slice — structurally malformed".to_string(),
            );
        }
        PlanOp::Unwind { .. } => {
            return Err(
                "group executor: Unwind in producer slice — structurally malformed".to_string(),
            );
        }
        PlanOp::LeftOuterApply { .. } => {
            return Err(
                "group executor: LeftOuterApply in producer slice — structurally malformed"
                    .to_string(),
            );
        }
    }
    Ok(())
}

/// Recursively pull rows through `ops`, projecting into `result` until
/// `result.len() >= ctx.bound`.
///
/// Each arm binds one `PlanOp` and recurses on `rest`.  When `ops` is empty
/// (all producers consumed), the current `row` is projected and appended.
///
/// `row` is a mutable scratch buffer: each arm that assigns a slot saves and
/// restores it around the recursive call so the caller sees no net change.
/// This eliminates the per-row `Vec` clone that the staged path requires.
///
/// ScanLabel iterates `view.labels` directly (lazy, no intermediate `Vec<u32>`)
/// so the bound truncates the scan itself — scanning stops as soon as enough
/// result rows have been collected.
fn pull_rows(
    ctx: &PullCtx<'_>,
    ops: &[PlanOp],
    row: &mut Row,
    result: &mut Vec<Vec<Option<Value>>>,
) -> Result<(), String> {
    if result.len() >= ctx.bound {
        return Ok(());
    }
    let (op, rest) = match ops.split_first() {
        Some(pair) => pair,
        None => {
            // All producers consumed — project this final row.
            let mut cells = Vec::with_capacity(ctx.project_items.len());
            for item in ctx.project_items {
                cells.push(project_item(ctx.view, ctx.vars, row, item, ctx.params)?);
            }
            result.push(cells);
            return Ok(());
        }
    };
    match op {
        PlanOp::ScanLabel { var, label } => {
            let slot = ctx
                .vars
                .slot(var)
                .ok_or_else(|| format!("unbound variable `{var}`"))?;
            // Resolve the label to an interned symbol once. If the label is
            // specified but unknown, there are no matching nodes — return early.
            let want_sym = label.as_deref().and_then(|l| ctx.view.syms.get(l));
            if label.is_some() && want_sym.is_none() {
                return Ok(());
            }
            // Save the slot value so we can restore it after the loop.
            let prev = row[slot].clone();

            // Fast path: if the immediately following op is a simple
            // `Prop op Lit` comparison on this scan variable, fuse the filter
            // into the scan loop.  Pre-resolving the property column once
            // (hashing the field name once instead of per-node) eliminates the
            // outer HashMap string-hash on every candidate row.
            let fused_filter = rest.first().and_then(|next_op| {
                if let PlanOp::Filter {
                    expr:
                        Expr::Cmp {
                            lhs:
                                Operand::Prop {
                                    var: ref fv,
                                    field: ref f,
                                },
                            op: ref cmp_op_ref,
                            rhs: Operand::Lit(ref lit),
                        },
                } = *next_op
                {
                    if fv == var {
                        return Some((f.as_str(), cmp_op_ref, lit));
                    }
                }
                None
            });

            // `n.key` / `n.id` / `n.label` are not columns, so the fused path
            // would see an empty column and drop every row. Fall through to
            // the generic path, which goes through `resolve_prop`.
            let fused_filter = fused_filter.filter(|(f, _, _)| !is_identity_field(f));
            if let Some((field, cmp_op_ref, lit)) = fused_filter {
                // Fused scan+filter: column resolved once, comparison done
                // inline — no recursive call into pull_rows for the Filter arm.
                // The column reads overlay and snapshot base, as
                // `resolve_prop` does (ledger row 139).
                #[cfg(test)]
                FUSED_SCAN_FIRES.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
                let col = ctx.view.props.view_column(field);
                let rest_after_filter = &rest[1..];
                for (i, &sym) in ctx.view.labels.iter().enumerate() {
                    if result.len() >= ctx.bound {
                        break;
                    }
                    if sym == u32::MAX {
                        continue;
                    }
                    if let Some(ws) = want_sym {
                        if sym != ws {
                            continue;
                        }
                    }
                    let id = i as u32;
                    if !ctx.view.visible(id) {
                        continue;
                    }
                    if col.test(id, |v| eval_cmp(cmp_op_ref, v, lit)) {
                        row[slot] = Some(Cell::Node(id));
                        pull_rows(ctx, rest_after_filter, row, result)?;
                    }
                }
            } else {
                // Generic path: iterate label array lazily, no Vec allocation,
                // with early exit when the row bound is satisfied.
                for (i, &sym) in ctx.view.labels.iter().enumerate() {
                    if result.len() >= ctx.bound {
                        break;
                    }
                    // Skip tombstone / gap slots (u32::MAX sentinel).
                    if sym == u32::MAX {
                        continue;
                    }
                    // Filter by label symbol when a label is requested.
                    if let Some(ws) = want_sym {
                        if sym != ws {
                            continue;
                        }
                    }
                    let id = i as u32;
                    if !ctx.view.visible(id) {
                        continue;
                    }
                    row[slot] = Some(Cell::Node(id));
                    pull_rows(ctx, rest, row, result)?;
                }
            }
            // SAFETY: the `?` inside both scan loops propagates Err directly to
            // `execute_pull`, which drops `initial_row` — the skipped restore is
            // unobservable.  This invariant MUST stay true: no call site between
            // here and `execute_pull` may catch an Err and re-use the row.
            // Future refactors adding mid-stream error recovery must audit this site.
            row[slot] = prev;
        }
        PlanOp::ScanKey { var, key, label } => {
            let slot = ctx
                .vars
                .slot(var)
                .ok_or_else(|| format!("unbound variable `{var}`"))?;
            if let Some(id) =
                resolve_scan_key_id(ctx.view, ctx.vars, row, key, label.as_deref(), ctx.params)?
            {
                let prev = row[slot].clone();
                row[slot] = Some(Cell::Node(id));
                pull_rows(ctx, rest, row, result)?;
                row[slot] = prev;
            }
        }
        PlanOp::IndexScan {
            var,
            label,
            field,
            value,
        } => {
            let slot = ctx
                .vars
                .slot(var)
                .ok_or_else(|| format!("unbound variable `{var}`"))?;
            let ids = index_scan_ids(
                ctx.view,
                ctx.vars,
                row,
                label.as_deref(),
                field,
                value,
                ctx.params,
            )?;
            let prev = row[slot].clone();
            for id in ids {
                if result.len() >= ctx.bound {
                    break;
                }
                row[slot] = Some(Cell::Node(id));
                pull_rows(ctx, rest, row, result)?;
            }
            row[slot] = prev;
        }
        PlanOp::IndexIntersect {
            var,
            label,
            equalities,
        } => {
            let slot = ctx
                .vars
                .slot(var)
                .ok_or_else(|| format!("unbound variable `{var}`"))?;
            let ids = index_intersect_ids(
                ctx.view,
                ctx.vars,
                row,
                label.as_deref(),
                equalities,
                ctx.params,
            )?;
            let prev = row[slot].clone();
            for id in ids {
                if result.len() >= ctx.bound {
                    break;
                }
                row[slot] = Some(Cell::Node(id));
                pull_rows(ctx, rest, row, result)?;
            }
            row[slot] = prev;
        }
        PlanOp::Expand {
            from,
            rel_var,
            etypes,
            dir,
            to,
            to_label,
            to_props,
        } => {
            let etypes = resolve_etypes(ctx.view, etypes);
            let exp_dir = map_dir(*dir);
            let to_slot = ctx
                .vars
                .slot(to)
                .ok_or_else(|| format!("unbound variable `{to}`"))?;
            let rel_slot = rel_var.as_ref().and_then(|rv| ctx.vars.slot(rv));
            let from_id = require_node(row, ctx.vars, from)?;
            let bound_to = match row.get(to_slot).and_then(|c| c.as_ref()) {
                Some(Cell::Node(id)) => Some(*id),
                Some(Cell::Rel(_) | Cell::Path(_) | Cell::Scalar(_)) => {
                    return Err(format!("variable `{to}` is not a node"))
                }
                None => None,
            };
            for e in expand(ctx.view, from_id, etypes.as_deref(), exp_dir) {
                if result.len() >= ctx.bound {
                    break;
                }
                if row_has_edge(row, &e) {
                    continue;
                }
                let nbr = neighbor(from_id, &e, *dir);
                if !ctx.view.visible(nbr) {
                    continue;
                }
                if let Some(want) = bound_to {
                    if nbr != want {
                        continue;
                    }
                }
                if !node_matches(
                    ctx.view,
                    ctx.vars,
                    row,
                    nbr,
                    to_label.as_deref(),
                    to_props,
                    ctx.params,
                )? {
                    continue;
                }
                let mut next = row.clone();
                if let Some(slot) = rel_slot {
                    next[slot] = Some(Cell::Rel(e));
                }
                if bound_to.is_none() {
                    next[to_slot] = Some(Cell::Node(nbr));
                }
                #[cfg(test)]
                record_expand_row();
                pull_rows(ctx, rest, &mut next, result)?;
            }
        }
        PlanOp::Filter { expr } => {
            if eval_expr(ctx.view, ctx.vars, row, expr, ctx.params, 0)? {
                pull_rows(ctx, rest, row, result)?;
            }
        }
        PlanOp::LookupProps { var, props } => {
            let id = require_node(row, ctx.vars, var)?;
            if node_matches(ctx.view, ctx.vars, row, id, None, props, ctx.params)? {
                pull_rows(ctx, rest, row, result)?;
            }
        }
        PlanOp::JoinBound { var, label, props } => {
            let id = require_node(row, ctx.vars, var)?;
            if node_matches(
                ctx.view,
                ctx.vars,
                row,
                id,
                label.as_deref(),
                props,
                ctx.params,
            )? {
                pull_rows(ctx, rest, row, result)?;
            }
        }
        // The ops below must never appear inside the producers slice that
        // pull_rows receives.  Explicitly reject each so that adding a new
        // PlanOp variant to the enum forces a compile-time decision here
        // rather than silently falling through and producing wrong results.
        PlanOp::Project { .. } => {
            return Err(
                "pull executor: Project reached pull_rows — plan is structurally malformed"
                    .to_string(),
            );
        }
        PlanOp::Distinct => {
            return Err(
                "pull executor: Distinct reached pull_rows — DISTINCT queries must use \
                 the staged path (row_bound returns None)"
                    .to_string(),
            );
        }
        PlanOp::OrderBy { .. } => {
            return Err(
                "pull executor: OrderBy reached pull_rows — queries with ORDER BY \
                 must use the staged path (row_bound returns None)"
                    .to_string(),
            );
        }
        PlanOp::Skip(_) => {
            return Err(
                "pull executor: Skip reached pull_rows — Skip must appear after Project"
                    .to_string(),
            );
        }
        PlanOp::Limit(_) => {
            return Err(
                "pull executor: Limit reached pull_rows — Limit must appear after Project"
                    .to_string(),
            );
        }
        PlanOp::Aggregate { .. } => {
            return Err(
                "pull executor: Aggregate reached pull_rows — aggregate plans must use \
                 the execute_aggregate path (routed before pull in execute_inner)"
                    .to_string(),
            );
        }
        // VarExpand and ShortestPath always take the staged path (row_bound()
        // returns None for plans containing these ops, so pull_rows is never
        // called with them in the producer slice).  This arm exists so that
        // adding new variants to PlanOp forces a compile-time decision here.
        PlanOp::VarExpand { .. } => {
            return Err(
                "pull executor: VarExpand reached pull_rows — variable-length path \
                 plans must use the staged path (row_bound returns None)"
                    .to_string(),
            );
        }
        PlanOp::ShortestPath { .. } => {
            return Err(
                "pull executor: ShortestPath reached pull_rows — shortestPath plans \
                 must use the staged path (row_bound returns None)"
                    .to_string(),
            );
        }
        PlanOp::GroupAggregate { .. } => {
            return Err(
                "pull executor: GroupAggregate reached pull_rows — grouped aggregate plans \
                 must use the execute_group_aggregate path (routed before pull in execute_inner)"
                    .to_string(),
            );
        }
        PlanOp::With { .. } => {
            return Err(
                "pull executor: With reached pull_rows — pipeline plans must use the staged path \
                 (row_bound returns None for plans containing With)"
                    .to_string(),
            );
        }
        PlanOp::Unwind { .. } => {
            return Err(
                "pull executor: Unwind reached pull_rows — pipeline plans must use the staged path \
                 (row_bound returns None for plans containing Unwind)"
                    .to_string(),
            );
        }
        PlanOp::LeftOuterApply { .. } => {
            return Err(
                "pull executor: LeftOuterApply reached pull_rows — OPTIONAL MATCH plans must use \
                 the staged path (row_bound returns None for plans containing LeftOuterApply)"
                    .to_string(),
            );
        }
    }
    Ok(())
}

fn eval_expr(
    view: &GraphView,
    vars: &VarTable,
    row: &Row,
    expr: &Expr,
    params: &Params,
    depth: u32,
) -> Result<bool, String> {
    if depth > 256 {
        return Err("expression nesting too deep".into());
    }
    match expr {
        Expr::And(lhs, rhs) => {
            let l = eval_expr(view, vars, row, lhs, params, depth + 1)?;
            let r = eval_expr(view, vars, row, rhs, params, depth + 1)?;
            Ok(l && r)
        }
        Expr::Or(lhs, rhs) => {
            let l = eval_expr(view, vars, row, lhs, params, depth + 1)?;
            let r = eval_expr(view, vars, row, rhs, params, depth + 1)?;
            Ok(l || r)
        }
        Expr::Not(inner) => Ok(!eval_expr(view, vars, row, inner, params, depth + 1)?),
        Expr::Cmp { lhs, op, rhs } => {
            let l = resolve_operand(view, vars, row, lhs, params)?;
            let r = resolve_operand(view, vars, row, rhs, params)?;
            match (l, r) {
                (Some(a), Some(b)) => Ok(eval_cmp(op, &a, &b)),
                _ => Ok(false),
            }
        }
        Expr::Truthy(op) => {
            let val = resolve_operand(view, vars, row, op, params)?;
            Ok(match val {
                None => false,
                Some(Value::Bool(b)) => b,
                Some(Value::Int(n)) => n != 0,
                Some(Value::Float(f)) => f != 0.0,
                Some(Value::Str(s)) => !s.is_empty(),
                Some(Value::List(v)) => !v.is_empty(),
                Some(Value::Map(m)) => !m.is_empty(),
            })
        }
        Expr::IsNull(op) => {
            let val = resolve_operand(view, vars, row, op, params)?;
            Ok(val.is_none())
        }
        Expr::IsNotNull(op) => {
            let val = resolve_operand(view, vars, row, op, params)?;
            Ok(val.is_some())
        }
        Expr::In { expr, list } => eval_in(view, vars, row, expr, list, params),
    }
}

fn eval_in(
    view: &GraphView,
    vars: &VarTable,
    row: &Row,
    expr: &Operand,
    list: &[Operand],
    params: &Params,
) -> Result<bool, String> {
    let Some(needle) = resolve_operand(view, vars, row, expr, params)? else {
        return Ok(false);
    };
    for item_op in list {
        match resolve_operand(view, vars, row, item_op, params)? {
            None => {}
            Some(Value::List(items)) => {
                for item in items {
                    if crate::filter::eval_cmp(&crate::filter::CmpOp::Eq, &needle, &item) {
                        return Ok(true);
                    }
                }
            }
            Some(item) if crate::filter::eval_cmp(&crate::filter::CmpOp::Eq, &needle, &item) => {
                return Ok(true);
            }
            Some(_) => {}
        }
    }
    Ok(false)
}

/// Deduplicate projected rows. Numeric Int/Float unify matches grouping.
fn exec_distinct(table: &mut Projected) -> Result<(), String> {
    let cap = max_intermediate_rows();
    let mut seen: BTreeSet<Vec<Option<ValueKey>>> = BTreeSet::new();
    let mut out = Vec::with_capacity(table.rows.len().min(cap));
    for row in table.rows.drain(..) {
        let key: Vec<Option<ValueKey>> = row
            .iter()
            .map(|cell| cell.as_ref().and_then(group_key_normalize))
            .collect();
        if seen.insert(key) {
            if out.len() >= cap {
                return Err(row_cap_err(cap));
            }
            out.push(row);
        }
    }
    table.rows = out;
    Ok(())
}

fn column_name(item: &RetItem) -> String {
    if let Some(alias) = &item.alias {
        return alias.clone();
    }
    // Agg column names are computed by the planner and stored on the plan op;
    // this branch is unreachable for well-formed plans but needed for
    // exhaustiveness.
    ret_val_label(&item.value).unwrap_or_else(|| match &item.value {
        RetVal::Agg { func, arg } => {
            let f = match func {
                AggFunc::Count => "COUNT",
                AggFunc::Sum => "SUM",
                AggFunc::Avg => "AVG",
                AggFunc::Min => "MIN",
                AggFunc::Max => "MAX",
                AggFunc::Collect => "COLLECT",
            };
            format!("{f}({})", agg_arg_label(arg))
        }
        _ => unreachable!("ret_val_label names every non-aggregate item"),
    })
}

fn exec_project(
    view: &GraphView,
    vars: &VarTable,
    rows: &[Row],
    items: &[RetItem],
    params: &Params,
) -> Result<Projected, String> {
    let columns: Vec<String> = items.iter().map(column_name).collect();
    let mut out_rows = Vec::with_capacity(rows.len());
    for row in rows {
        let mut cells = Vec::with_capacity(items.len());
        for item in items {
            cells.push(project_item(view, vars, row, item, params)?);
        }
        out_rows.push(cells);
    }
    Ok(Projected {
        columns,
        rows: out_rows,
    })
}

fn project_item(
    view: &GraphView,
    vars: &VarTable,
    row: &Row,
    item: &RetItem,
    params: &Params,
) -> Result<Option<Value>, String> {
    match &item.value {
        RetVal::Var(v) => {
            // Look up the slot; missing from VarTable entirely is a hard error.
            let slot = vars.slot(v).ok_or_else(|| format!("unbound variable `{v}`"))?;
            match row.get(slot).and_then(|c| c.as_ref()) {
                // Cell is None — variable is optionally null (from OPTIONAL MATCH).
                None => Ok(None),
                Some(Cell::Node(id)) => match view.ids.key_of(*id) {
                    Some(key) => Ok(Some(Value::Str(key.to_owned()))),
                    None => Err(format!("unknown node id {id}")),
                },
                // Scalar alias produced by a prior WITH stage.
                Some(Cell::Scalar(val)) => Ok(Some(val.clone())),
                Some(Cell::Rel(_)) => Err(format!(
                    "variable `{v}` is a relationship; return its properties ({v}.field) instead"
                )),
                Some(Cell::Path(hops)) => Ok(Some(Value::Int(*hops as i64))),
            }
        }
        RetVal::Prop { var, field } => resolve_prop(view, vars, row, var, field),
        // Agg items are never projected by exec_project (aggregate plans have no
        // Project op). This arm exists solely to satisfy the exhaustive match.
        RetVal::Agg { .. } => Err(
            "project_item: Agg variant reached exec_project — aggregate plans must not contain Project"
                .to_string(),
        ),
        RetVal::FuncCall { name, args } => eval_func(name, args, view, vars, row, params),
        RetVal::ScalarExpr(op) => resolve_operand(view, vars, row, op, params),
    }
}

fn order_column(item: &OrderItem) -> String {
    match &item.target {
        OrderTarget::Alias(name) | OrderTarget::Var(name) => name.clone(),
        OrderTarget::Prop { var, field } => format!("{var}.{field}"),
    }
}

fn exec_order_by(table: &mut Projected, items: &[OrderItem]) -> Result<(), String> {
    let mut keys = Vec::with_capacity(items.len());
    for item in items {
        let name = order_column(item);
        let idx = table
            .columns
            .iter()
            .position(|c| c == &name)
            .ok_or_else(|| format!("ORDER BY target `{name}` is not a projected column"))?;
        keys.push((idx, item.descending));
    }
    table.rows.sort_by(|a, b| {
        for &(idx, desc) in &keys {
            let c = cmp_optional(
                a.get(idx).and_then(|x| x.as_ref()),
                b.get(idx).and_then(|x| x.as_ref()),
                desc,
            );
            if c != std::cmp::Ordering::Equal {
                return c;
            }
        }
        std::cmp::Ordering::Equal
    });
    Ok(())
}

/// Resolve a `LimitSkip` value to a concrete `u64` using the query params map.
///
/// `LimitSkip::Exact(n)` resolves immediately.  `LimitSkip::Param(name)` looks
/// up the named parameter and validates it is a non-negative integer.
fn resolve_ls(ls: &LimitSkip, params: &Params) -> Result<u64, String> {
    match ls {
        LimitSkip::Exact(n) => Ok(*n),
        LimitSkip::Param(name) => {
            let val = params
                .0
                .get(name)
                .ok_or_else(|| format!("missing parameter `{name}` (used in LIMIT/SKIP)"))?;
            match val {
                Value::Int(i) if *i >= 0 => Ok(*i as u64),
                Value::Int(i) => Err(format!(
                    "LIMIT/SKIP parameter `{name}` must be a non-negative integer, got {i}"
                )),
                other => Err(format!(
                    "LIMIT/SKIP parameter `{name}` must be an integer, got {other:?}"
                )),
            }
        }
    }
}

fn apply_skip<T>(rows: &mut Vec<T>, n: u64) {
    let n = usize::try_from(n).unwrap_or(usize::MAX);
    if n >= rows.len() {
        rows.clear();
    } else {
        rows.drain(0..n);
    }
}

fn apply_limit<T>(rows: &mut Vec<T>, n: u64) {
    let n = usize::try_from(n).unwrap_or(usize::MAX);
    rows.truncate(n);
}

#[cfg(test)]
mod tests {
    use super::{execute, resolve_operand, Params, Row, VarTable};
    use crate::cypher::ast::{
        ArithOp, LimitSkip, Operand, OrderItem, OrderTarget, RetItem, RetVal,
    };
    use crate::cypher::plan::{plan, PlanOp};
    use crate::cypher::{lex, parse, RelDir};
    use crate::result::ResultSet;
    use crate::view::GraphView;
    use core_storage::v8::seam::{ColumnsView, EdgePropsView, TopologyView};
    use core_storage::{ColumnStore, EdgeProps, IdMap, Interner, Topology, Value};
    use proptest::prelude::*;
    use std::collections::BTreeMap;

    struct Fx {
        ids: IdMap,
        syms: Interner,
        labels: Vec<u32>,
        props: ColumnStore,
        topo: Topology,
        eprops: EdgeProps,
    }

    impl Fx {
        fn new() -> Self {
            Fx {
                ids: IdMap::new(),
                syms: Interner::new(),
                labels: vec![],
                props: ColumnStore::new(),
                topo: Topology::new(),
                eprops: EdgeProps::new(),
            }
        }

        fn add(&mut self, label: &str, key: &str, props: Vec<(&str, Value)>) -> u32 {
            let id = self.ids.get_or_insert(key);
            let sym = self.syms.intern(label);
            self.labels.resize(id as usize + 1, u32::MAX);
            self.labels[id as usize] = sym;
            for (f, v) in props {
                self.props.set(id, f, v);
            }
            id
        }

        fn edge(&mut self, etype: &str, src: u32, dst: u32, props: Vec<(&str, Value)>) {
            let et = self.syms.intern(etype);
            self.topo.add_edge(et, src, dst);
            for (f, v) in props {
                self.eprops.set(et, src, dst, f, v);
            }
        }

        fn view(&self) -> GraphView<'_> {
            GraphView {
                ids: &self.ids,
                syms: &self.syms,
                labels: &self.labels,
                props: ColumnsView::owned(&self.props),
                topo: TopologyView::owned(&self.topo),
                edge_props: EdgePropsView::owned(&self.eprops),
                mask: None,
                prop_index: None,
            }
        }

        fn view_indexed<'a>(
            &'a self,
            index: &'a core_storage::property_index::PropertyIndex,
        ) -> GraphView<'a> {
            GraphView {
                prop_index: Some(index),
                ..self.view()
            }
        }
    }

    fn compile(src: &str) -> Vec<PlanOp> {
        plan(&parse(&lex(src).expect("lex")).expect("parse")).expect("plan")
    }

    fn run(
        view: &GraphView,
        src: &str,
        params: &BTreeMap<String, Value>,
    ) -> Result<ResultSet, String> {
        execute(view, &compile(src), &Params(params))
    }

    fn s(v: &str) -> Value {
        Value::Str(v.into())
    }

    fn f(v: f64) -> Value {
        Value::Float(v)
    }

    fn i(v: i64) -> Value {
        Value::Int(v)
    }

    fn rows_of(rs: &ResultSet) -> Vec<Vec<Option<Value>>> {
        (0..rs.len()).map(|i| rs.row(i).to_vec()).collect()
    }

    fn col(rs: &ResultSet, name: &str) -> Vec<Option<Value>> {
        (0..rs.len()).map(|i| rs.get(i, name).cloned()).collect()
    }

    fn hop_graph() -> Fx {
        let mut fx = Fx::new();
        let ada = fx.add("Person", "ada", vec![]);
        let bob = fx.add("Person", "bob", vec![]);
        let cam = fx.add("Person", "cam", vec![]);
        let acme = fx.add("Company", "acme", vec![]);
        fx.edge("KNOWS", ada, bob, vec![]);
        fx.edge("KNOWS", ada, cam, vec![]);
        fx.edge("KNOWS", bob, cam, vec![]);
        fx.edge("LIKES", ada, acme, vec![]);
        fx
    }

    fn undirected_graph() -> Fx {
        let mut fx = Fx::new();
        let a = fx.add("N", "a", vec![]);
        let b = fx.add("N", "b", vec![]);
        fx.edge("T", a, b, vec![("w", i(42))]);
        fx
    }

    fn triangle() -> Fx {
        let mut fx = Fx::new();
        let a = fx.add("N", "a", vec![]);
        let b = fx.add("N", "b", vec![]);
        let c = fx.add("N", "c", vec![]);
        fx.edge("T", a, b, vec![("eid", i(1))]);
        fx.edge("T", b, c, vec![("eid", i(2))]);
        fx.edge("T", c, a, vec![("eid", i(3))]);
        fx
    }

    fn single_edge() -> (Fx, u32, u32) {
        let mut fx = Fx::new();
        let a = fx.add("N", "a", vec![]);
        let b = fx.add("N", "b", vec![]);
        fx.edge("T", a, b, vec![]);
        (fx, a, b)
    }

    /// Companies with scored INDUSTRY_ALIGNMENT / SPECIALTY_MATCH edges to t1.
    fn dogfood_graph() -> Fx {
        let mut fx = Fx::new();
        let t1 = fx.add("Talent", "t1", vec![("id", s("t1"))]);
        let acme = fx.add("Company", "acme", vec![]);
        let beta = fx.add("Company", "beta", vec![]);
        let gamma = fx.add("Company", "gamma", vec![]);
        let delta = fx.add("Company", "delta", vec![]);
        let echo = fx.add("Company", "echo", vec![]);
        let foxtrot = fx.add("Company", "foxtrot", vec![]);
        let zeta = fx.add("Company", "zeta", vec![]);
        fx.edge("INDUSTRY_ALIGNMENT", acme, t1, vec![("score", f(0.9))]);
        fx.edge("SPECIALTY_MATCH", acme, t1, vec![("score", f(0.8))]);
        fx.edge("INDUSTRY_ALIGNMENT", beta, t1, vec![("score", f(0.6))]);
        fx.edge("SPECIALTY_MATCH", beta, t1, vec![("score", f(0.7))]);
        fx.edge("INDUSTRY_ALIGNMENT", gamma, t1, vec![("score", f(0.4))]);
        fx.edge("SPECIALTY_MATCH", gamma, t1, vec![("score", f(0.9))]);
        fx.edge("INDUSTRY_ALIGNMENT", delta, t1, vec![("score", f(0.8))]);
        fx.edge("SPECIALTY_MATCH", delta, t1, vec![("score", f(0.3))]);
        fx.edge("INDUSTRY_ALIGNMENT", echo, t1, vec![("score", f(0.5))]);
        fx.edge("SPECIALTY_MATCH", echo, t1, vec![("score", f(0.5))]);
        fx.edge("INDUSTRY_ALIGNMENT", foxtrot, t1, vec![("score", f(0.95))]);
        fx.edge("INDUSTRY_ALIGNMENT", zeta, t1, vec![("score", f(0.9))]);
        fx.edge("SPECIALTY_MATCH", zeta, t1, vec![("score", f(0.6))]);
        fx
    }

    const DOGFOOD: &str = "\
MATCH (t:Talent {id: $tid}) \
MATCH (c:Company)-[i:INDUSTRY_ALIGNMENT]->(t) \
MATCH (c)-[s:SPECIALTY_MATCH]->(t) \
WHERE i.score >= 0.5 AND s.score >= 0.5 \
RETURN c, i.score AS industry, s.score AS specialty \
ORDER BY industry DESC, specialty DESC \
LIMIT 10";

    fn tid_params() -> BTreeMap<String, Value> {
        let mut p = BTreeMap::new();
        p.insert("tid".into(), s("t1"));
        p
    }

    #[test]
    fn single_hop_match_label_and_etype_filters() {
        let fx = hop_graph();
        let v = fx.view();
        let rs = run(
            &v,
            "MATCH (a:Person)-[:KNOWS]->(b:Person) RETURN a, b",
            &BTreeMap::new(),
        )
        .expect("single-hop");
        assert_eq!(rs.columns(), &["a".to_string(), "b".to_string()]);
        assert_eq!(
            rows_of(&rs),
            vec![
                vec![Some(s("ada")), Some(s("bob"))],
                vec![Some(s("ada")), Some(s("cam"))],
                vec![Some(s("bob")), Some(s("cam"))],
            ]
        );
        // dest label filters out ada -LIKES-> acme; etype filters it too
        let likes = run(
            &v,
            "MATCH (a:Person)-[:LIKES]->(b:Company) RETURN a, b",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(rows_of(&likes), vec![vec![Some(s("ada")), Some(s("acme"))]]);
        let no_combo = run(
            &v,
            "MATCH (a:Person)-[:KNOWS]->(b:Company) RETURN a, b",
            &BTreeMap::new(),
        )
        .unwrap();
        assert!(no_combo.is_empty());
    }

    #[test]
    fn undirected_match_finds_both_orientations_and_binds_true_triple() {
        let fx = undirected_graph();
        let v = fx.view();
        // w lives only on the true (T, a, b) triple. A reversed (T, b, a)
        // binding would project None.
        let rs =
            run(&v, "MATCH (x)-[r:T]-(y) RETURN x, y, r.w", &BTreeMap::new()).expect("undirected");
        assert_eq!(
            rows_of(&rs),
            vec![
                vec![Some(s("a")), Some(s("b")), Some(i(42))],
                vec![Some(s("b")), Some(s("a")), Some(i(42))],
            ]
        );

        let left = run(
            &v,
            "MATCH (y)<-[r:T]-(x) RETURN x, y, r.w",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(
            rows_of(&left),
            vec![vec![Some(s("a")), Some(s("b")), Some(i(42))]]
        );
    }

    #[test]
    fn relationship_uniqueness_triangle_and_two_hop_cycle() {
        let tri = triangle();
        let v = tri.view();
        let rs = run(
            &v,
            "MATCH (x)-[r1:T]->(y)-[r2:T]->(z) RETURN x, y, z, r1.eid, r2.eid",
            &BTreeMap::new(),
        )
        .expect("triangle 2-hop");
        assert_eq!(
            rows_of(&rs),
            vec![
                vec![
                    Some(s("a")),
                    Some(s("b")),
                    Some(s("c")),
                    Some(i(1)),
                    Some(i(2))
                ],
                vec![
                    Some(s("b")),
                    Some(s("c")),
                    Some(s("a")),
                    Some(i(2)),
                    Some(i(3))
                ],
                vec![
                    Some(s("c")),
                    Some(s("a")),
                    Some(s("b")),
                    Some(i(3)),
                    Some(i(1))
                ],
            ]
        );
        for row in rows_of(&rs) {
            assert_ne!(row[3], row[4], "r1 must never bind the same edge as r2");
        }

        let (mut one, a, b) = single_edge();
        let v = one.view();
        let cycle = run(
            &v,
            "MATCH (x)-[:T]->(y)-[:T]->(x) RETURN x",
            &BTreeMap::new(),
        )
        .expect("2-hop cycle");
        assert!(
            cycle.is_empty(),
            "single directed edge cannot close a 2-hop cycle"
        );

        // Undirected 2-hop back would reuse the only EdgeRef without uniqueness.
        let undirected_cycle = run(&v, "MATCH (x)-[:T]-(y)-[:T]-(x) RETURN x", &BTreeMap::new())
            .expect("undirected uniqueness");
        assert!(
            undirected_cycle.is_empty(),
            "relationship uniqueness must reject walking the same triple back"
        );

        one.edge("T", b, a, vec![]);
        let v = one.view();
        let with_recip = run(
            &v,
            "MATCH (x)-[:T]->(y)-[:T]->(x) RETURN x",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(col(&with_recip, "x"), vec![Some(s("a")), Some(s("b"))]);
    }

    #[test]
    fn multi_match_join_bound_and_bound_destination_expand() {
        let fx = dogfood_graph();
        let v = fx.view();
        // No WHERE: any company with *both* edge types into the bound talent.
        // foxtrot has industry only → dropped by the second (bound-dest) expand.
        let rs = run(
            &v,
            "MATCH (t:Talent {id: $tid}) \
             MATCH (c:Company)-[i:INDUSTRY_ALIGNMENT]->(t) \
             MATCH (c)-[s:SPECIALTY_MATCH]->(t) \
             RETURN c",
            &tid_params(),
        )
        .expect("join + bound dest");
        assert_eq!(
            col(&rs, "c"),
            vec![
                Some(s("acme")),
                Some(s("beta")),
                Some(s("gamma")),
                Some(s("delta")),
                Some(s("echo")),
                Some(s("zeta")),
            ]
        );

        // Empty JoinBound (MATCH 2 start already bound, no label/props) keeps all.
        let keep = run(&v, "MATCH (c:Company) MATCH (c) RETURN c", &BTreeMap::new()).unwrap();
        assert_eq!(keep.len(), 7);
        // Label re-check on JoinBound drops everything.
        let drop = run(
            &v,
            "MATCH (c:Company) MATCH (c:Talent) RETURN c",
            &BTreeMap::new(),
        )
        .unwrap();
        assert!(drop.is_empty());
    }

    #[test]
    fn scan_key_exec_does_not_use_label_scan() {
        let mut fx = Fx::new();
        fx.add("Person", "p1", vec![]);
        fx.add("Person", "p2", vec![]);
        fx.add("Person", "p3", vec![]);
        fx.add("Company", "c1", vec![]);
        let v = fx.view();
        let params = BTreeMap::new();

        let fires_before = super::SCAN_KEY_FIRES.load(std::sync::atomic::Ordering::Relaxed);
        let rs = run(&v, "MATCH (n:Person {id: 'p2'}) RETURN n", &params).expect("scan key");
        let fires_after = super::SCAN_KEY_FIRES.load(std::sync::atomic::Ordering::Relaxed);
        assert!(
            fires_after > fires_before,
            "SCAN_KEY_FIRES must increment; before={fires_before} after={fires_after}"
        );
        assert_eq!(rows_of(&rs), vec![vec![Some(s("p2"))]]);

        let miss = run(&v, "MATCH (n:Person {id: 'nope'}) RETURN n", &params).unwrap();
        assert!(
            miss.is_empty(),
            "missing key must be zero rows, not an error"
        );

        let wrong = run(&v, "MATCH (n:Person {id: 'c1'}) RETURN n", &params).unwrap();
        assert!(
            wrong.is_empty(),
            "wrong label must be zero rows, not an error"
        );
    }

    #[test]
    fn rel_var_edge_prop_filter() {
        let mut fx = Fx::new();
        let a = fx.add("N", "a", vec![]);
        let b = fx.add("N", "b", vec![]);
        let c = fx.add("N", "c", vec![]);
        fx.edge("T", a, b, vec![("w", f(0.7))]);
        fx.edge("T", a, c, vec![("w", f(0.3))]);
        let v = fx.view();
        let rs = run(
            &v,
            "MATCH (x)-[r:T]->(y) WHERE r.w >= 0.5 RETURN y, r.w",
            &BTreeMap::new(),
        )
        .expect("edge-prop filter");
        assert_eq!(rows_of(&rs), vec![vec![Some(s("b")), Some(f(0.7))]]);
        let fail = run(
            &v,
            "MATCH (x)-[r:T]->(y) WHERE r.w >= 0.8 RETURN y",
            &BTreeMap::new(),
        )
        .unwrap();
        assert!(fail.is_empty());
    }

    #[test]
    fn params_present_resolve_missing_is_err_before_rows() {
        let fx = dogfood_graph();
        let v = fx.view();
        let hit =
            run(&v, "MATCH (t:Talent {id: $tid}) RETURN t", &tid_params()).expect("present param");
        assert_eq!(col(&hit, "t"), vec![Some(s("t1"))]);

        // Unknown label would yield Ok(empty) if params were not walked first.
        let err = run(
            &v,
            "MATCH (t:NoSuchLabel {id: $tid}) RETURN t",
            &BTreeMap::new(),
        )
        .expect_err("missing param must be Err, not Ok(empty)");
        assert!(
            err.contains("tid")
                && (err.contains("param") || err.contains("Param") || err.contains("missing")),
            "missing-param error must name the parameter, got: {err}"
        );

        let err = run(
            &v,
            "MATCH (t:Talent) WHERE t.id = $tid RETURN t",
            &BTreeMap::new(),
        )
        .expect_err("missing WHERE param");
        assert!(err.contains("tid"), "got: {err}");
    }

    #[test]
    fn order_by_none_last_then_skip_limit() {
        let mut fx = Fx::new();
        fx.add("Person", "ada", vec![("age", i(30))]);
        fx.add("Person", "bob", vec![]); // missing age → None
        fx.add("Person", "cam", vec![("age", i(10))]);
        fx.add("Person", "dan", vec![("age", i(20))]);
        let v = fx.view();

        let asc = run(
            &v,
            "MATCH (p:Person) RETURN p, p.age AS age ORDER BY age",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(
            rows_of(&asc),
            vec![
                vec![Some(s("cam")), Some(i(10))],
                vec![Some(s("dan")), Some(i(20))],
                vec![Some(s("ada")), Some(i(30))],
                vec![Some(s("bob")), None],
            ]
        );

        let desc = run(
            &v,
            "MATCH (p:Person) RETURN p, p.age AS age ORDER BY age DESC",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(
            rows_of(&desc),
            vec![
                vec![Some(s("ada")), Some(i(30))],
                vec![Some(s("dan")), Some(i(20))],
                vec![Some(s("cam")), Some(i(10))],
                vec![Some(s("bob")), None],
            ]
        );

        let skip_lim = run(
            &v,
            "MATCH (p:Person) RETURN p, p.age AS age ORDER BY age SKIP 1 LIMIT 2",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(
            rows_of(&skip_lim),
            vec![
                vec![Some(s("dan")), Some(i(20))],
                vec![Some(s("ada")), Some(i(30))],
            ]
        );

        let desc_sl = run(
            &v,
            "MATCH (p:Person) RETURN p, p.age AS age ORDER BY age DESC SKIP 1 LIMIT 2",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(
            rows_of(&desc_sl),
            vec![
                vec![Some(s("dan")), Some(i(20))],
                vec![Some(s("cam")), Some(i(10))],
            ]
        );
    }

    #[test]
    fn unknown_label_and_etype_are_ok_empty() {
        let fx = hop_graph();
        let v = fx.view();
        let lab = run(&v, "MATCH (x:Nope) RETURN x", &BTreeMap::new()).expect("unknown label");
        assert!(lab.is_empty());
        let et = run(
            &v,
            "MATCH (a)-[:NO_SUCH_ETYPE]->(b) RETURN a",
            &BTreeMap::new(),
        )
        .expect("unknown etype");
        assert!(et.is_empty());
    }

    #[test]
    fn execute_is_deterministic() {
        let fx = dogfood_graph();
        let v = fx.view();
        let p = tid_params();
        let a = run(&v, DOGFOOD, &p).expect("first");
        let b = run(&v, DOGFOOD, &p).expect("second");
        assert_eq!(a, b);
        let hop = hop_graph();
        let hv = hop.view();
        let q = "MATCH (a:Person)-[:KNOWS]->(b:Person) RETURN a, b";
        assert_eq!(run(&hv, q, &BTreeMap::new()), run(&hv, q, &BTreeMap::new()));
    }

    #[test]
    fn dogfood_pipeline_exact_rows() {
        let fx = dogfood_graph();
        let v = fx.view();
        let rs = run(&v, DOGFOOD, &tid_params()).expect("dogfood");
        assert_eq!(
            rs.columns(),
            &[
                "c".to_string(),
                "industry".to_string(),
                "specialty".to_string()
            ]
        );
        // industry DESC, specialty DESC; gamma (0.4) / delta (0.3) / foxtrot (no s) out.
        assert_eq!(
            rows_of(&rs),
            vec![
                vec![Some(s("acme")), Some(f(0.9)), Some(f(0.8))],
                vec![Some(s("zeta")), Some(f(0.9)), Some(f(0.6))],
                vec![Some(s("beta")), Some(f(0.6)), Some(f(0.7))],
                vec![Some(s("echo")), Some(f(0.5)), Some(f(0.5))],
            ]
        );
    }

    #[test]
    fn unknown_var_in_op_is_err_not_panic() {
        let fx = hop_graph();
        let v = fx.view();
        let plan = vec![
            PlanOp::ScanLabel {
                var: "a".into(),
                label: None,
            },
            PlanOp::Expand {
                from: "zzz".into(),
                rel_var: Some("r".into()),
                etypes: vec![],
                dir: RelDir::Right,
                to: "b".into(),
                to_label: None,
                to_props: vec![],
            },
            PlanOp::Project {
                items: vec![RetItem {
                    value: RetVal::Var("a".into()),
                    alias: None,
                }],
            },
        ];
        let params = BTreeMap::new();
        let caught = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
            execute(&v, &plan, &Params(&params))
        }));
        assert!(caught.is_ok(), "execute panicked on unknown var");
        let err = caught.unwrap().expect_err("unknown var must be Err");
        assert!(
            err.contains("zzz") && err.to_ascii_lowercase().contains("unbound"),
            "got: {err}"
        );

        let join = vec![
            PlanOp::JoinBound {
                var: "ghost".into(),
                label: None,
                props: vec![],
            },
            PlanOp::Project {
                items: vec![RetItem {
                    value: RetVal::Var("ghost".into()),
                    alias: None,
                }],
            },
        ];
        let caught = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
            execute(&v, &join, &Params(&params))
        }));
        assert!(caught.is_ok(), "execute panicked on JoinBound unknown var");
        assert!(caught.unwrap().is_err());
    }

    #[test]
    fn dest_props_on_bound_expand_are_applied() {
        let fx = dogfood_graph();
        let v = fx.view();
        let rs = run(
            &v,
            "MATCH (t:Talent {id: $tid}) \
             MATCH (c:Company)-[r:INDUSTRY_ALIGNMENT]->(t:Talent {id: $tid}) \
             RETURN c",
            &tid_params(),
        )
        .unwrap();
        // foxtrot included (has industry); companies without the edge are not.
        assert_eq!(
            col(&rs, "c"),
            vec![
                Some(s("acme")),
                Some(s("beta")),
                Some(s("gamma")),
                Some(s("delta")),
                Some(s("echo")),
                Some(s("foxtrot")),
                Some(s("zeta")),
            ]
        );
        let miss = run(
            &v,
            "MATCH (t:Talent {id: $tid}) \
             MATCH (c:Company)-[r:INDUSTRY_ALIGNMENT]->(t {id: 'nope'}) \
             RETURN c",
            &tid_params(),
        )
        .unwrap();
        assert!(miss.is_empty());
    }

    #[test]
    fn missing_node_prop_projects_none_and_pattern_misses() {
        let mut fx = Fx::new();
        fx.add("Person", "ada", vec![("age", i(30))]);
        fx.add("Person", "bob", vec![]);
        let v = fx.view();
        let rs = run(&v, "MATCH (p:Person) RETURN p.age", &BTreeMap::new()).unwrap();
        assert_eq!(col(&rs, "p.age"), vec![Some(i(30)), None]);
        let pat = run(&v, "MATCH (p:Person {age: 30}) RETURN p", &BTreeMap::new()).unwrap();
        assert_eq!(col(&pat, "p"), vec![Some(s("ada"))]);
    }

    #[test]
    fn unlabeled_match_does_not_project_sentinel_ghost_key() {
        let mut fx = Fx::new();
        fx.add("Person", "ada", vec![]);
        // Hostile fixture: IdMap slot exists (key_of would yield "ghost")
        // but the label is the gap sentinel.
        fx.ids.get_or_insert("ghost");
        fx.labels.resize(fx.ids.len(), u32::MAX);
        fx.add("Person", "bob", vec![]);
        let v = fx.view();
        let rs = run(&v, "MATCH (n) RETURN n", &BTreeMap::new()).expect("unlabeled scan");
        assert_eq!(col(&rs, "n"), vec![Some(s("ada")), Some(s("bob"))]);
        assert!(
            !rows_of(&rs)
                .iter()
                .any(|row| row.iter().any(|c| *c == Some(s("ghost")))),
            "sentinel slot must not project a ghost key"
        );
    }

    #[test]
    fn execute_never_panics_on_hostile_plans() {
        let fx = hop_graph();
        let v = fx.view();
        let params = BTreeMap::new();
        let hostile = vec![
            vec![],
            vec![PlanOp::Project { items: vec![] }],
            vec![PlanOp::OrderBy {
                items: vec![OrderItem {
                    target: OrderTarget::Alias("nope".into()),
                    descending: false,
                }],
            }],
            vec![PlanOp::Filter {
                expr: crate::cypher::ast::Expr::Cmp {
                    lhs: Operand::Prop {
                        var: "missing".into(),
                        field: "x".into(),
                    },
                    op: crate::filter::CmpOp::Eq,
                    rhs: Operand::Lit(i(1)),
                },
            }],
            vec![
                PlanOp::ScanLabel {
                    var: "a".into(),
                    label: None,
                },
                PlanOp::LookupProps {
                    var: "zzz".into(),
                    props: vec![("k".into(), Operand::Lit(i(1)))],
                },
            ],
            vec![
                PlanOp::Skip(LimitSkip::Exact(99)),
                PlanOp::Limit(LimitSkip::Exact(0)),
            ],
        ];
        for plan in hostile {
            let caught = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
                execute(&v, &plan, &Params(&params))
            }));
            assert!(caught.is_ok(), "execute panicked on hostile plan: {plan:?}");
        }
    }

    #[test]
    fn unlabeled_and_labeled_scans_skip_tombstoned_ids() {
        let mut fx = Fx::new();
        let ada = fx.add("Person", "ada", vec![]);
        let bob = fx.add("Person", "bob", vec![]);
        fx.edge("KNOWS", ada, bob, vec![]);
        // Same state delete_node leaves: id retired, label sentinel, edges gone.
        fx.ids.delete("ada");
        fx.labels[ada as usize] = u32::MAX;
        let knows = fx.syms.get("KNOWS").unwrap();
        fx.topo.remove_edge(knows, ada, bob);
        let v = fx.view();

        let labeled = run(&v, "MATCH (p:Person) RETURN p", &BTreeMap::new()).unwrap();
        assert_eq!(col(&labeled, "p"), vec![Some(s("bob"))]);
        let unlabeled = run(&v, "MATCH (n) RETURN n", &BTreeMap::new()).unwrap();
        assert_eq!(col(&unlabeled, "n"), vec![Some(s("bob"))]);
        let hop = run(&v, "MATCH (x)-[:KNOWS]->(y) RETURN x, y", &BTreeMap::new()).unwrap();
        assert!(
            hop.is_empty(),
            "expand cannot yield edges to a deleted node once topology is swept"
        );
    }

    // ──────────────────────────────────────────────────────────────────────────
    // Randomised property test (I2): bounded == unbounded[SKIP..SKIP+LIMIT]
    // ──────────────────────────────────────────────────────────────────────────

    proptest! {
        /// Randomised property test covering:
        ///
        /// - **Multi-hop** (1, 2, or 3 hops) — exercises nested pull_rows recursion.
        /// - **Optional WHERE filter** on a numeric prop `v` — bound must count
        ///   only post-filter rows.
        /// - **Cycle / shared-node topologies** — duplicate directed pairs and
        ///   reciprocal edges create paths where relationship-uniqueness rejects
        ///   some traversals; rejected rows must not count toward the bound.
        /// - **Randomized SKIP + LIMIT** — pull collects SKIP+LIMIT rows then
        ///   the wrapper discards the leading SKIP.
        ///
        /// Invariant: bounded[0..] == unbounded[skip..skip+limit] for all shapes.
        #[test]
        fn prop_bounded_equals_unbounded_slice(
            n_nodes     in 2u32..10u32,
            edge_pairs  in proptest::collection::vec(
                (any::<u32>(), any::<u32>()), 0..20usize
            ),
            // Extra reciprocal pairs to force uniqueness rejections on cycles.
            recip_pairs in proptest::collection::vec(
                (any::<u32>(), any::<u32>()), 0..8usize
            ),
            n_hops      in 1u32..4u32,   // 1, 2, or 3 hops
            use_filter  in any::<bool>(),
            threshold   in 0i64..8i64,   // filter: last-node.v > threshold
            limit       in 1u64..10u64,
            skip        in 0u64..4u64,
        ) {
            let mut fx = Fx::new();
            let mut node_ids = Vec::new();
            for idx in 0..n_nodes {
                // Every node carries a numeric prop `v` for the optional filter.
                let id = fx.add("N", &format!("n{idx}"), vec![("v", i(idx as i64 % 8))]);
                node_ids.push(id);
            }
            let n = node_ids.len();

            // Primary edges (forward direction).
            for (si, di) in &edge_pairs {
                let si = (*si as usize) % n;
                let di = (*di as usize) % n;
                if si != di {
                    fx.edge("T", node_ids[si], node_ids[di], vec![]);
                }
            }
            // Reciprocal edges — create A→B + B→A pairs so multi-hop paths
            // have uniqueness-rejected candidates (walking back the same edge).
            for (si, di) in &recip_pairs {
                let si = (*si as usize) % n;
                let di = (*di as usize) % n;
                if si != di {
                    fx.edge("T", node_ids[si], node_ids[di], vec![]);
                    fx.edge("T", node_ids[di], node_ids[si], vec![]);
                }
            }

            let v = fx.view();
            let params = BTreeMap::new();

            // Build query for `n_hops` hops with optional WHERE on the last node.
            // Variable names: a→b→c→d for hops 1/2/3.
            let var_names = ["a", "b", "c", "d"];
            let hop_count = n_hops as usize;
            let mut pattern = format!("({}:N)", var_names[0]);
            for h in 0..hop_count {
                pattern.push_str(&format!("-[:T]->({}", var_names[h + 1]));
                // Label the intermediate and last nodes as :N only for last.
                if h + 1 == hop_count {
                    pattern.push_str(":N)");
                } else {
                    pattern.push(')');
                }
            }
            let last_var = var_names[hop_count];
            let where_clause = if use_filter {
                format!(" WHERE {last_var}.v > {threshold}")
            } else {
                String::new()
            };
            let ret_vars: Vec<&str> = var_names[..=hop_count].to_vec();
            let ret_clause = ret_vars.join(", ");

            let full_q = format!("MATCH {pattern}{where_clause} RETURN {ret_clause}");
            let bounded_q = format!(
                "MATCH {pattern}{where_clause} RETURN {ret_clause} SKIP {skip} LIMIT {limit}"
            );

            let full_plan = compile(&full_q);
            // Use a generous cap for the reference run so it never errors on
            // small graphs, even with cycles.
            let unbounded = super::with_max_intermediate_rows(100_000, || {
                super::execute_unbounded(&v, &full_plan, &Params(&params))
            });
            // If the reference itself errors (shouldn't happen at this scale),
            // skip the proptest case rather than failing.
            let unbounded = match unbounded {
                Ok(rs) => rs,
                Err(_) => return Ok(()),
            };
            let total = unbounded.len();
            let full_rows = rows_of(&unbounded);

            let bounded = super::with_max_intermediate_rows(100_000, || {
                run(&v, &bounded_q, &params)
            }).expect("bounded must not error");

            let s = (skip as usize).min(total);
            let e = (skip as usize + limit as usize).min(total);
            prop_assert_eq!(
                rows_of(&bounded),
                full_rows[s..e].to_vec(),
                "hops={} filter={} threshold={} SKIP {} LIMIT {}: \
                 bounded != unbounded[{}..{}]",
                hop_count, use_filter, threshold, skip, limit, s, e
            );
        }
    }

    proptest! {
        /// Randomised property test pinning the fused ScanLabel+Filter fast path.
        ///
        /// Shape: `MATCH (n:N) WHERE n.v <op> <literal> RETURN n SKIP s LIMIT l`
        ///
        /// This shape places a `Filter { Cmp { Prop{n}, op, Lit } }` immediately
        /// after `ScanLabel { var: n }` in the plan, which triggers the fused
        /// detection in `pull_rows`.  The invariant is the same as
        /// `prop_bounded_equals_unbounded_slice`: bounded == unbounded[s..s+l].
        ///
        /// Coverage:
        ///  - All 6 CmpOps (=, <>, <, <=, >, >=)
        ///  - Nodes that have the prop (Int or Float) and nodes that are missing it
        ///  - Threshold values spanning match-all, match-none, and partial
        ///  - Randomised SKIP and LIMIT
        ///  - Boundary shape: compound AND (`WHERE n.v op lit AND n.v >= -999`)
        ///    which is semantically equivalent but bypasses fused detection
        ///    (Expr::And, not Expr::Cmp), exercising the generic fallback path
        #[test]
        fn prop_scan_filter_fused_equals_unbounded_slice(
            n_nodes     in 0u32..20u32,
            // Bit i set → node i has prop `v`.
            prop_mask   in any::<u32>(),
            // Bit i set → node i stores a Float value instead of Int.
            float_mask  in any::<u32>(),
            // Comparison threshold; range -1..8 spans all/none/partial matches
            // against node values in 0..7.
            threshold   in -1i64..8i64,
            // Index into the 6 CmpOps (Eq=0, Ne=1, Lt=2, Le=3, Gt=4, Ge=5).
            op_idx      in 0u32..6u32,
            skip        in 0u64..5u64,
            limit       in 1u64..8u64,
        ) {
            let op_str = match op_idx {
                0 => "=",
                1 => "<>",
                2 => "<",
                3 => "<=",
                4 => ">",
                _ => ">=",
            };

            let mut fx = Fx::new();
            for idx in 0..n_nodes {
                let has_prop = (prop_mask >> (idx % 32)) & 1 == 1;
                let use_float = (float_mask >> (idx % 32)) & 1 == 1;
                // Node values cycle in 0..7 so threshold spans all/none/partial.
                let props: Vec<(&str, Value)> = if has_prop {
                    if use_float {
                        vec![("v", f(idx as f64 % 7.0))]
                    } else {
                        vec![("v", i(idx as i64 % 7))]
                    }
                } else {
                    vec![]
                };
                fx.add("N", &format!("n{idx}"), props);
            }

            let v = fx.view();
            let params = BTreeMap::new();

            // ── Fused path ───────────────────────────────────────────────────
            let full_q = format!("MATCH (n:N) WHERE n.v {op_str} {threshold} RETURN n");
            let bounded_q = format!(
                "MATCH (n:N) WHERE n.v {op_str} {threshold} RETURN n SKIP {skip} LIMIT {limit}"
            );

            let full_plan = compile(&full_q);
            let unbounded = super::execute_unbounded(&v, &full_plan, &Params(&params));
            let unbounded = match unbounded {
                Ok(rs) => rs,
                Err(_) => return Ok(()),
            };
            let total = unbounded.len();
            let full_rows = rows_of(&unbounded);

            // Snapshot counter before executing the fused-shape query.
            let fires_before =
                super::FUSED_SCAN_FIRES.load(std::sync::atomic::Ordering::Relaxed);
            let bounded = run(&v, &bounded_q, &params).expect("fused bounded must not error");
            let fires_after =
                super::FUSED_SCAN_FIRES.load(std::sync::atomic::Ordering::Relaxed);

            // The planner must have emitted ScanLabel→Filter{Cmp} for this shape;
            // assert the fused arm actually executed (counter advanced).
            // Guard: with 0 nodes the label symbol is never interned, so pull_rows
            // exits before the fused detection — nothing to assert in that case.
            // Guard: op_idx==0 (Eq) is folded to IndexScan by the WHERE equality
            // fold pass — the fused ScanLabel+Filter shape is not produced, so the
            // FUSED counter does not advance; correctness is still verified below.
            if n_nodes > 0 && op_idx != 0 {
                prop_assert!(
                    fires_after > fires_before,
                    "fused arm did NOT fire for op={} threshold={} n_nodes={}: \
                     counter before={} after={}",
                    op_str,
                    threshold,
                    n_nodes,
                    fires_before,
                    fires_after
                );
            }

            let s = (skip as usize).min(total);
            let e = (skip as usize + limit as usize).min(total);
            prop_assert_eq!(
                rows_of(&bounded),
                full_rows[s..e].to_vec(),
                "fused path: op={} threshold={} n_nodes={} SKIP {} LIMIT {}: \
                 bounded != unbounded[{}..{}]",
                op_str, threshold, n_nodes, skip, limit, s, e
            );

            // ── Boundary: compound AND — misses fused detection → generic path ─
            // `AND n.v >= -999` is always true for our Int/Float range 0..7,
            // so the result set is identical — only the executor path differs.
            let compound_q = format!(
                "MATCH (n:N) WHERE n.v {} {} AND n.v >= -999 RETURN n SKIP {} LIMIT {}",
                op_str, threshold, skip, limit
            );
            let fires_before_compound =
                super::FUSED_SCAN_FIRES.load(std::sync::atomic::Ordering::Relaxed);
            let compound_bounded =
                run(&v, &compound_q, &params).expect("compound-AND bounded must not error");
            let fires_after_compound =
                super::FUSED_SCAN_FIRES.load(std::sync::atomic::Ordering::Relaxed);

            // Compound AND must NOT activate the fused arm.
            prop_assert_eq!(
                fires_after_compound,
                fires_before_compound,
                "fused arm fired for compound-AND shape (should use generic path): \
                 op={} threshold={} n_nodes={}",
                op_str,
                threshold,
                n_nodes
            );
            prop_assert_eq!(
                rows_of(&compound_bounded),
                full_rows[s..e].to_vec(),
                "compound-AND fallback: op={} threshold={} n_nodes={} SKIP {} LIMIT {}: \
                 result differs from fused",
                op_str, threshold, n_nodes, skip, limit
            );
        }
    }

    // ──────────────────────────────────────────────────────────────────────────
    // C1 / C2: dense hop-1 with downstream filter
    //
    // The actual failing shape: 1 source → N leaves (N > cap), and a downstream
    // Filter that passes only a small fraction.  The per-stage approach (Round 1)
    // errors because hop-1 Expand runs to the full cap before the Filter sees
    // anything.  The pull-based approach cascades the bound: once `limit` rows
    // have passed the Filter, all upstream loops stop.
    // ──────────────────────────────────────────────────────────────────────────

    #[test]
    fn dense_hop1_with_filter_survives_pull() {
        const LEAVES: usize = 120; // > CAP so staged always errors
        const CAP: usize = 100;

        let mut fx = Fx::new();
        let src = fx.add("Src", "src", vec![]);
        for idx in 0..LEAVES {
            let leaf = fx.add("Leaf", &format!("l{idx}"), vec![("v", i(idx as i64))]);
            fx.edge("T", src, leaf, vec![]);
        }
        let v = fx.view();
        let params = BTreeMap::new();

        // Staged (unbounded) path: expand produces 120 rows > cap=100 → error.
        let staged_err = super::with_max_intermediate_rows(CAP, || {
            super::execute_unbounded(
                &v,
                &compile("MATCH (s:Src)-[:T]->(l:Leaf) WHERE l.v >= 110 RETURN l, l.v"),
                &Params(&params),
            )
        });
        assert!(
            staged_err.is_err(),
            "staged path must error on 120 leaves with cap={CAP}"
        );
        assert!(
            staged_err
                .unwrap_err()
                .contains("intermediate result exceeds"),
            "wrong error message"
        );

        // Pull-based (LIMIT 5): never materialises more than 5 rows → survives.
        // WHERE l.v >= 110 means only leaves 110..119 pass (10 survivors), so
        // 5 results are found well before all 120 leaves are expanded.
        let ok = super::with_max_intermediate_rows(CAP, || {
            run(
                &v,
                "MATCH (s:Src)-[:T]->(l:Leaf) WHERE l.v >= 110 RETURN l, l.v LIMIT 5",
                &params,
            )
        });
        let rs = ok.expect("pull-based must survive despite dense hop-1 exceeding cap");
        assert_eq!(rs.len(), 5, "LIMIT 5 must return exactly 5 rows");

        // Verify all returned rows have l.v >= 110 (correct filter application).
        let vs: Vec<i64> = (0..rs.len())
            .filter_map(|i| match rs.get(i, "l.v") {
                Some(Value::Int(n)) => Some(*n),
                _ => None,
            })
            .collect();
        assert_eq!(vs.len(), 5, "all projected rows must have v");
        for v_val in &vs {
            assert!(*v_val >= 110, "filter must hold: v={v_val} is not >= 110");
        }

        // Early-termination proof: use a filter that passes early leaves (v < 10)
        // so pull stops after visiting just 5 leaves, while staged visits all 120.
        // Ratio: 120 / 5 = 24× — well above the 10× threshold.
        let (pull_result, pull_produced) = super::with_expand_counter(|| {
            super::with_max_intermediate_rows(1_000_000, || {
                run(
                    &v,
                    "MATCH (s:Src)-[:T]->(l:Leaf) WHERE l.v < 10 RETURN l LIMIT 5",
                    &params,
                )
            })
        });
        pull_result.expect("pull must succeed without cap");
        // Pull visits leaves 0..4 (all pass v < 10, LIMIT 5 satisfied immediately).
        assert!(
            pull_produced <= 5,
            "pull expand count {pull_produced} should be ≤ 5 (stops after 5 passing leaves)"
        );

        let (staged_result, staged_produced) = super::with_expand_counter(|| {
            super::with_max_intermediate_rows(1_000_000, || {
                super::execute_unbounded(
                    &v,
                    &compile("MATCH (s:Src)-[:T]->(l:Leaf) WHERE l.v < 10 RETURN l"),
                    &Params(&params),
                )
            })
        });
        staged_result.expect("staged must succeed with 1M cap");
        assert_eq!(
            staged_produced, LEAVES,
            "staged must expand all {LEAVES} leaves"
        );

        assert!(
            staged_produced >= pull_produced * 10,
            "staged ({staged_produced}) must be ≥ 10× pull ({pull_produced})"
        );
    }

    // ──────────────────────────────────────────────────────────────────────────
    // Harness-shape two-hop dense test
    //
    // Replicates the exact failure shape from the public benchmark table at
    // 1/100 scale (70 Talent + 20 Company with 3 industry categories).
    // IA edges are added directly (bypassing the rule engine) to reproduce the
    // dense edge structure that `Predicate::FieldEqual { field: "industry" }`
    // generates at full scale.
    //
    // Query: MATCH (t:Talent)-[:INDUSTRY_ALIGNMENT]->(c:Company)
    //               <-[:INDUSTRY_ALIGNMENT]-(t2:Talent)
    //        RETURN t, c, t2 LIMIT 10
    //
    // Before (staged with cap=100): hop-1 expands 70×~7=~466 rows > 100 → error
    // After  (pull-based):          finds 10 results, stops, returns correctly
    // ──────────────────────────────────────────────────────────────────────────

    #[test]
    fn harness_shape_two_hop_dense_survives_pull() {
        const N_TALENT: usize = 70;
        const N_COMPANY: usize = 20;
        const N_INDUSTRY: usize = 3;
        const CAP: usize = 100;
        const LIMIT: usize = 10;

        let mut fx = Fx::new();

        // Build Talent nodes with industry tag.
        let mut talent_ids: Vec<u32> = Vec::new();
        let mut talent_industry: Vec<usize> = Vec::new();
        for i in 0..N_TALENT {
            let ind = i % N_INDUSTRY;
            let id = fx.add(
                "Talent",
                &format!("t{i}"),
                vec![("industry", s(&ind.to_string()))],
            );
            talent_ids.push(id);
            talent_industry.push(ind);
        }

        // Build Company nodes with industry tag.
        let mut company_ids: Vec<u32> = Vec::new();
        let mut company_industry: Vec<usize> = Vec::new();
        for i in 0..N_COMPANY {
            let ind = i % N_INDUSTRY;
            let id = fx.add(
                "Company",
                &format!("c{i}"),
                vec![("industry", s(&ind.to_string()))],
            );
            company_ids.push(id);
            company_industry.push(ind);
        }

        // INDUSTRY_ALIGNMENT: Talent → Company when same industry.
        for (ti, &tid) in talent_ids.iter().enumerate() {
            for (ci, &cid) in company_ids.iter().enumerate() {
                if talent_industry[ti] == company_industry[ci] {
                    fx.edge("INDUSTRY_ALIGNMENT", tid, cid, vec![]);
                }
            }
        }

        let v = fx.view();
        let params = BTreeMap::new();
        let query = format!(
            "MATCH (t:Talent)-[:INDUSTRY_ALIGNMENT]->(c:Company)\
             <-[:INDUSTRY_ALIGNMENT]-(t2:Talent) RETURN t, c, t2 LIMIT {LIMIT}"
        );

        // Staged (unbounded) path: hop-1 expands 70×~7=~466 rows > cap=100 → error.
        const UNBOUNDED_Q: &str = "MATCH (t:Talent)-[:INDUSTRY_ALIGNMENT]->(c:Company)\
             <-[:INDUSTRY_ALIGNMENT]-(t2:Talent) RETURN t, c, t2";
        let staged_err = super::with_max_intermediate_rows(CAP, || {
            super::execute_unbounded(&v, &compile(UNBOUNDED_Q), &Params(&params))
        });
        assert!(
            staged_err.is_err(),
            "staged must error with cap={CAP} on harness-shape graph"
        );
        assert!(
            staged_err
                .unwrap_err()
                .contains("intermediate result exceeds"),
            "wrong error"
        );

        // Pull-based (LIMIT 10): cascades bound through both hops → completes.
        let ok = super::with_max_intermediate_rows(CAP, || run(&v, &query, &params));
        let rs = ok.expect("pull-based must complete on harness-shape with LIMIT 10");
        assert_eq!(rs.len(), LIMIT, "must return exactly {LIMIT} rows");

        // Each result row (t, c, t2) must have matching industry across all 3 variables.
        // t and c share an IA edge (same industry); c and t2 share an IA edge too.
        // Since we can't directly query industry from the ResultSet without projecting it,
        // just verify the result is semantically plausible: 3 non-None columns per row.
        for i in 0..rs.len() {
            let row = rs.row(i);
            assert_eq!(row.len(), 3, "each row must have 3 columns (t, c, t2)");
            assert!(row.iter().all(|c| c.is_some()), "all cells must be Some");
        }
    }

    #[test]
    fn intermediate_row_cap_errors_on_scan_and_expand() {
        let cap_msg = |n: usize| {
            format!(
                "intermediate result exceeds {n} rows; add a LIMIT or constrain patterns with shared variables"
            )
        };

        let mut scan_fx = Fx::new();
        scan_fx.add("N", "a", vec![]);
        scan_fx.add("N", "b", vec![]);
        scan_fx.add("N", "c", vec![]);
        let sv = scan_fx.view();
        let scan_err = super::with_max_intermediate_rows(2, || {
            run(&sv, "MATCH (n:N) RETURN n", &BTreeMap::new())
        })
        .expect_err("3-row scan must exceed cap 2");
        assert_eq!(scan_err, cap_msg(2));

        let mut exp_fx = Fx::new();
        let src = exp_fx.add("Src", "s", vec![]);
        let d1 = exp_fx.add("Dst", "d1", vec![]);
        let d2 = exp_fx.add("Dst", "d2", vec![]);
        exp_fx.edge("T", src, d1, vec![]);
        exp_fx.edge("T", src, d2, vec![]);
        let ev = exp_fx.view();
        // Scan of :Src is 1 row (under cap); expand to two dests would be 2.
        let exp_err = super::with_max_intermediate_rows(1, || {
            run(&ev, "MATCH (x:Src)-[:T]->(y) RETURN x, y", &BTreeMap::new())
        })
        .expect_err("2-row expand must exceed cap 1");
        assert_eq!(exp_err, cap_msg(1));
    }

    // ──────────────────────────────────────────────────────────────────────────
    // LIMIT push-down: semantics, early-termination proof, and budget survival
    // ──────────────────────────────────────────────────────────────────────────

    /// Property test: bounded execution (execute with push-down) produces the
    /// same rows as the reference unbounded path (execute_unbounded) sliced to
    /// the first LIMIT rows.  Exercises several LIMIT and SKIP+LIMIT values,
    /// including queries that have a Filter so the bound falls on Filter output,
    /// and queries with relationship-uniqueness rejections.
    #[test]
    fn bounded_matches_unbounded_slice_various_limits() {
        // ── single-hop, no Filter ──────────────────────────────────────────────
        let fx = hop_graph();
        let v = fx.view();
        let params = BTreeMap::new();

        // Full 3-row reference (ada→bob, ada→cam, bob→cam).
        let full_plan = compile("MATCH (a:Person)-[:KNOWS]->(b:Person) RETURN a, b");
        let full_rs = super::execute_unbounded(&v, &full_plan, &Params(&params)).unwrap();
        let full_rows = rows_of(&full_rs);
        assert_eq!(full_rows.len(), 3, "hop_graph has exactly 3 KNOWS paths");

        for limit in [1u64, 2, 3, 10] {
            let q = format!("MATCH (a:Person)-[:KNOWS]->(b:Person) RETURN a, b LIMIT {limit}");
            let rs = run(&v, &q, &params).unwrap();
            let expected_len = (limit as usize).min(full_rows.len());
            assert_eq!(
                rs.len(),
                expected_len,
                "LIMIT {limit}: expected {expected_len} rows, got {}",
                rs.len()
            );
            assert_eq!(
                rows_of(&rs),
                full_rows[..expected_len],
                "LIMIT {limit}: rows differ from reference slice"
            );
        }

        // SKIP + LIMIT: bound = SKIP + LIMIT so slicing is correct.
        let skip_rs = run(
            &v,
            "MATCH (a:Person)-[:KNOWS]->(b:Person) RETURN a, b SKIP 1 LIMIT 2",
            &params,
        )
        .unwrap();
        assert_eq!(rows_of(&skip_rs), full_rows[1..3]);

        // ── with WHERE (bound falls on Filter, not Expand) ────────────────────
        let mut fx2 = Fx::new();
        fx2.add("N", "a", vec![("v", i(1))]);
        fx2.add("N", "b", vec![("v", i(2))]);
        fx2.add("N", "c", vec![("v", i(3))]);
        let v2 = fx2.view();
        // Full result in order: a(1), b(2), c(3).  After WHERE v > 1: b, c.
        let filter_full = super::execute_unbounded(
            &v2,
            &compile("MATCH (n:N) WHERE n.v > 1 RETURN n"),
            &Params(&params),
        )
        .unwrap();
        assert_eq!(filter_full.len(), 2);

        let filter_lim = run(&v2, "MATCH (n:N) WHERE n.v > 1 RETURN n LIMIT 1", &params).unwrap();
        assert_eq!(filter_lim.len(), 1, "LIMIT 1 on filter query");
        assert_eq!(rows_of(&filter_lim), rows_of(&filter_full)[..1]);

        // ── relationship-uniqueness rejections do not count toward the bound ──
        let tri = triangle();
        let tv = tri.view();
        // Triangle has 3 two-hop paths; uniqueness rejects 6 (reversed pairs).
        let tri_full = super::execute_unbounded(
            &tv,
            &compile("MATCH (x)-[r1:T]->(y)-[r2:T]->(z) RETURN x, y, z"),
            &Params(&params),
        )
        .unwrap();
        assert_eq!(tri_full.len(), 3, "triangle has 3 unique two-hop paths");
        for limit in [1u64, 2, 3, 5] {
            let q = format!("MATCH (x)-[r1:T]->(y)-[r2:T]->(z) RETURN x, y, z LIMIT {limit}");
            let rs = run(&tv, &q, &params).unwrap();
            let expected_len = (limit as usize).min(3);
            assert_eq!(
                rs.len(),
                expected_len,
                "triangle LIMIT {limit}: got {} rows",
                rs.len()
            );
            assert_eq!(
                rows_of(&rs),
                rows_of(&tri_full)[..expected_len],
                "triangle LIMIT {limit}: rows differ"
            );
        }
    }

    /// Early-termination proof: on a star graph (1 hub → 500 leaves), a bounded
    /// query (LIMIT 5) must cause `exec_expand` to emit ≤ 5 rows, while an
    /// unbounded run emits all 500 (≥ 100× the bounded count).
    #[test]
    fn expand_terminates_early_with_row_bound() {
        const LEAVES: usize = 500;
        let mut fx = Fx::new();
        let hub = fx.add("Hub", "hub", vec![]);
        for i in 0..LEAVES {
            let leaf = fx.add("Leaf", &format!("leaf-{i}"), vec![]);
            fx.edge("T", hub, leaf, vec![]);
        }
        let v = fx.view();
        let params = BTreeMap::new();
        let full_plan = compile("MATCH (h:Hub)-[:T]->(x:Leaf) RETURN x");

        // Bounded path via the regular execute() entry point (LIMIT 5 → bound 5).
        let (bounded_result, bounded_produced) = super::with_expand_counter(|| {
            run(&v, "MATCH (h:Hub)-[:T]->(x:Leaf) RETURN x LIMIT 5", &params)
        });
        let bounded_rs = bounded_result.unwrap();
        assert_eq!(bounded_rs.len(), 5, "LIMIT 5 must return exactly 5 rows");
        assert!(
            bounded_produced <= 5,
            "bounded: exec_expand emitted {bounded_produced} rows, expected ≤ 5"
        );

        // Unbounded reference path via execute_unbounded (no push-down).
        let (unbounded_result, unbounded_produced) = super::with_expand_counter(|| {
            super::execute_unbounded(&v, &full_plan, &Params(&params))
        });
        let unbounded_rs = unbounded_result.unwrap();
        assert_eq!(
            unbounded_rs.len(),
            LEAVES,
            "unbounded must return all {LEAVES} rows"
        );
        assert_eq!(
            unbounded_produced, LEAVES,
            "unbounded: exec_expand must emit all {LEAVES} rows"
        );

        // Early-termination ratio ≥ 100×.
        assert!(
            unbounded_produced >= bounded_produced * 100,
            "unbounded ({unbounded_produced}) must be ≥ 100× bounded ({bounded_produced})"
        );
    }

    /// Budget-survival: a bounded query (LIMIT pushdown active) must complete
    /// without triggering the intermediate-row cap, even when the same query
    /// without pushdown (execute_unbounded) would exceed a reduced cap.
    ///
    /// This also verifies that the 1 M cap is still live for unbounded queries.
    #[test]
    fn bounded_query_survives_low_intermediate_row_cap() {
        // 20 leaf nodes: unbounded would emit 20 rows, exceeding a cap of 10.
        let mut fx = Fx::new();
        let src = fx.add("Src", "s", vec![]);
        for i in 0..20usize {
            let dst = fx.add("Dst", &format!("d{i}"), vec![]);
            fx.edge("T", src, dst, vec![]);
        }
        let v = fx.view();
        let params = BTreeMap::new();
        let full_plan = compile("MATCH (s:Src)-[:T]->(d:Dst) RETURN d");

        // Unbounded hits the cap — 1 M budget check must remain live.
        let cap_err = super::with_max_intermediate_rows(10, || {
            super::execute_unbounded(&v, &full_plan, &Params(&params))
        });
        assert!(
            cap_err.is_err(),
            "unbounded must hit the intermediate-row cap"
        );
        assert!(
            cap_err.unwrap_err().contains("intermediate result exceeds"),
            "cap error message must be the budget message"
        );

        // Bounded (LIMIT 5) stops before the cap — must complete successfully.
        let ok = super::with_max_intermediate_rows(10, || {
            run(&v, "MATCH (s:Src)-[:T]->(d:Dst) RETURN d LIMIT 5", &params)
        });
        assert_eq!(
            ok.unwrap().len(),
            5,
            "bounded (LIMIT 5) must complete with 5 rows, not a cap error"
        );

        // Bounded with LIMIT exactly at cap also survives.
        let at_cap = super::with_max_intermediate_rows(10, || {
            run(&v, "MATCH (s:Src)-[:T]->(d:Dst) RETURN d LIMIT 10", &params)
        });
        assert_eq!(
            at_cap.unwrap().len(),
            10,
            "bounded at LIMIT==cap must complete with 10 rows"
        );
    }

    // ──────────────────────────────────────────────────────────────────────────
    // Aggregate execution tests
    // ──────────────────────────────────────────────────────────────────────────

    #[test]
    fn count_star_returns_total_node_count() {
        let mut fx = Fx::new();
        fx.add("Person", "ada", vec![]);
        fx.add("Person", "bob", vec![]);
        fx.add("Person", "cam", vec![]);
        let v = fx.view();
        let params = BTreeMap::new();

        let rs = run(&v, "MATCH (n:Person) RETURN COUNT(*)", &params).expect("COUNT(*)");
        assert_eq!(rs.columns(), &["COUNT(*)".to_string()]);
        assert_eq!(rs.len(), 1);
        assert_eq!(rs.row(0), &[Some(i(3))]);

        // Empty graph: COUNT(*) should return 0.
        let rs_empty = run(&v, "MATCH (n:Ghost) RETURN COUNT(*)", &params).expect("COUNT(*) empty");
        assert_eq!(rs_empty.row(0), &[Some(i(0))]);
    }

    #[test]
    fn count_star_alias_sets_column_name() {
        let mut fx = Fx::new();
        fx.add("N", "a", vec![]);
        let v = fx.view();
        let params = BTreeMap::new();
        let rs = run(&v, "MATCH (n:N) RETURN COUNT(*) AS total", &params).expect("COUNT AS");
        assert_eq!(rs.columns(), &["total".to_string()]);
        assert_eq!(rs.row(0), &[Some(i(1))]);
    }

    #[test]
    fn count_var_skips_null_node_bindings() {
        // COUNT(n) counts rows where n is bound. Since ScanLabel always binds n,
        // this matches COUNT(*) for nodes. Primarily documents the semantics.
        let mut fx = Fx::new();
        fx.add("N", "a", vec![]);
        fx.add("N", "b", vec![]);
        let v = fx.view();
        let params = BTreeMap::new();
        let rs = run(&v, "MATCH (n:N) RETURN COUNT(n)", &params).expect("COUNT(n)");
        assert_eq!(rs.columns(), &["COUNT(n)".to_string()]);
        assert_eq!(rs.row(0), &[Some(i(2))]);
    }

    #[test]
    fn sum_numeric_prop_ignores_null_and_non_numeric() {
        let mut fx = Fx::new();
        fx.add("N", "a", vec![("v", i(10))]);
        fx.add("N", "b", vec![("v", i(20))]);
        fx.add("N", "c", vec![]); // missing prop — skipped
        let v = fx.view();
        let params = BTreeMap::new();
        let rs = run(&v, "MATCH (n:N) RETURN SUM(n.v)", &params).expect("SUM");
        assert_eq!(rs.columns(), &["SUM(n.v)".to_string()]);
        // 10 + 20 = 30 (null skipped); integers sum to an integer (row 158)
        assert_eq!(rs.row(0), &[Some(i(30))]);

        // All props null → result is null.
        let rs_null = run(&v, "MATCH (n:N) RETURN SUM(n.missing)", &params).expect("SUM null");
        assert_eq!(rs_null.row(0), &[None]);
    }

    #[test]
    fn avg_numeric_prop() {
        let mut fx = Fx::new();
        fx.add("N", "a", vec![("v", i(10))]);
        fx.add("N", "b", vec![("v", i(30))]);
        let v = fx.view();
        let params = BTreeMap::new();
        let rs = run(&v, "MATCH (n:N) RETURN AVG(n.v) AS avg_v", &params).expect("AVG");
        assert_eq!(rs.columns(), &["avg_v".to_string()]);
        // (10 + 30) / 2 = 20.0
        assert_eq!(rs.row(0), &[Some(f(20.0))]);

        // Empty graph: AVG returns null.
        let rs_empty = run(&v, "MATCH (n:Ghost) RETURN AVG(n.v)", &params).expect("AVG empty");
        assert_eq!(rs_empty.row(0), &[None]);
    }

    #[test]
    fn min_max_numeric_prop() {
        let mut fx = Fx::new();
        fx.add("N", "a", vec![("v", i(5))]);
        fx.add("N", "b", vec![("v", i(1))]);
        fx.add("N", "c", vec![("v", i(9))]);
        fx.add("N", "d", vec![]); // null skipped
        let v = fx.view();
        let params = BTreeMap::new();

        let min_rs = run(&v, "MATCH (n:N) RETURN MIN(n.v)", &params).expect("MIN");
        assert_eq!(min_rs.row(0), &[Some(i(1))]);

        let max_rs = run(&v, "MATCH (n:N) RETURN MAX(n.v)", &params).expect("MAX");
        assert_eq!(max_rs.row(0), &[Some(i(9))]);
    }

    /// M-2: MIN/MAX with mixed Int and Float props.  The cmp_optional ordering
    /// places Int and Float by numeric value (cross-variant numeric comparison).
    #[test]
    fn min_max_mixed_int_float_props() {
        let mut fx = Fx::new();
        // Int 3, Float 1.5, Int 7, Float 2.0 — min=1.5 (Float), max=7 (Int).
        fx.add("N", "a", vec![("v", i(3))]);
        fx.add("N", "b", vec![("v", f(1.5))]);
        fx.add("N", "c", vec![("v", i(7))]);
        fx.add("N", "d", vec![("v", f(2.0))]);
        let v = fx.view();
        let params = BTreeMap::new();

        let min_rs = run(&v, "MATCH (n:N) RETURN MIN(n.v)", &params).expect("MIN mixed");
        // 1.5 < 2.0 < 3 < 7 — minimum is Float(1.5).
        assert_eq!(min_rs.row(0), &[Some(f(1.5))]);

        let max_rs = run(&v, "MATCH (n:N) RETURN MAX(n.v)", &params).expect("MAX mixed");
        // Maximum is Int(7).
        assert_eq!(max_rs.row(0), &[Some(i(7))]);
    }

    /// I-1: LIMIT, SKIP, and ORDER BY are silently dropped for aggregate
    /// queries (always one result row).  Pin both boundary values.
    #[test]
    fn aggregate_limit_skip_order_by_are_no_ops() {
        let mut fx = Fx::new();
        fx.add("N", "a", vec![]);
        fx.add("N", "b", vec![]);
        fx.add("N", "c", vec![]);
        let v = fx.view();
        let params = BTreeMap::new();

        // LIMIT 5 — aggregate always returns exactly 1 row regardless.
        let rs_lim5 =
            run(&v, "MATCH (n:N) RETURN COUNT(*) LIMIT 5", &params).expect("COUNT(*) LIMIT 5");
        assert_eq!(
            rs_lim5.len(),
            1,
            "aggregate with LIMIT 5 must still return 1 row"
        );
        assert_eq!(rs_lim5.row(0), &[Some(i(3))]);

        // LIMIT 0 — even LIMIT 0 does not suppress the aggregate row.
        let rs_lim0 =
            run(&v, "MATCH (n:N) RETURN COUNT(*) LIMIT 0", &params).expect("COUNT(*) LIMIT 0");
        assert_eq!(
            rs_lim0.len(),
            1,
            "aggregate with LIMIT 0 must still return 1 row"
        );
        assert_eq!(rs_lim0.row(0), &[Some(i(3))]);

        // SKIP 100 — does not discard the single result row.
        let rs_skip =
            run(&v, "MATCH (n:N) RETURN COUNT(*) SKIP 100", &params).expect("COUNT(*) SKIP 100");
        assert_eq!(
            rs_skip.len(),
            1,
            "aggregate with large SKIP must still return 1 row"
        );

        // ORDER BY is a no-op on a single-row result (but must not panic).
        // Note: the planner drops ORDER BY for aggregates; verify that the plan
        // compiles without error and returns the correct count.
        let rs_ord = plan_src("MATCH (n:N) RETURN COUNT(*) ORDER BY n");
        // ORDER BY on aggregate: planner drops ORDER BY, so this should plan OK.
        // (The planner exits early after emitting Aggregate, so ORDER BY is ignored.)
        assert!(
            rs_ord.is_ok(),
            "COUNT(*) ORDER BY should plan without error (ORDER BY dropped)"
        );
        let plan_ops = rs_ord.unwrap();
        // Must not contain an OrderBy op — it was dropped.
        assert!(
            !plan_ops
                .iter()
                .any(|op| matches!(op, crate::cypher::plan::PlanOp::OrderBy { .. })),
            "aggregate plan must not contain OrderBy"
        );
    }

    #[test]
    fn count_star_no_budget_cap_applies() {
        // COUNT(*) with a dense graph that would error the staged path.
        // The aggregate path must complete without hitting the cap.
        let mut fx = Fx::new();
        let src = fx.add("Src", "s", vec![]);
        for i in 0..30usize {
            let dst = fx.add("Dst", &format!("d{i}"), vec![]);
            fx.edge("T", src, dst, vec![]);
        }
        let v = fx.view();
        let params = BTreeMap::new();

        // Staged path errors on 30 nodes > cap 10.
        let cap_err =
            super::with_max_intermediate_rows(10, || run(&v, "MATCH (n:Dst) RETURN n", &params));
        assert!(
            cap_err.is_err(),
            "staged path must error on 30 nodes with cap=10"
        );

        // COUNT(*) does not apply the cap — must complete and return 30.
        let agg_ok = super::with_max_intermediate_rows(10, || {
            run(&v, "MATCH (n:Dst) RETURN COUNT(*)", &params)
        })
        .expect("aggregate must not hit the intermediate-row cap");
        assert_eq!(
            agg_ok.row(0),
            &[Some(i(30))],
            "COUNT(*) must count all 30 nodes regardless of cap"
        );
    }

    fn plan_src(src: &str) -> Result<Vec<crate::cypher::plan::PlanOp>, String> {
        use crate::cypher::{lex, parse, plan};
        let toks = lex(src).map_err(|e| format!("lex: {e}"))?;
        let ast = parse(&toks).map_err(|e| format!("parse: {e}"))?;
        plan(&ast).map_err(|e| format!("plan: {e}"))
    }

    /// Updated from the v1 pin: grouped aggregation is now supported.
    /// Verifies plan routing and that the only remaining plan-error is SUM(*).
    #[test]
    fn grouped_aggregation_plan_routing() {
        use crate::cypher::plan::PlanOp;

        // RETURN a, COUNT(*) — grouped aggregation now routes to GroupAggregate.
        let ops = plan_src("MATCH (a:N) RETURN a, COUNT(*)")
            .expect("grouped aggregation must now succeed");
        assert!(
            ops.iter()
                .any(|op| matches!(op, PlanOp::GroupAggregate { .. })),
            "grouped aggregation plan must contain GroupAggregate op, got: {ops:?}"
        );

        // Multiple aggregates without group keys also routes to GroupAggregate.
        let ops2 = plan_src("MATCH (a:N) RETURN COUNT(*), COUNT(a)")
            .expect("multi-aggregate must now succeed");
        assert!(
            ops2.iter()
                .any(|op| matches!(op, PlanOp::GroupAggregate { .. })),
            "multi-aggregate plan must contain GroupAggregate op, got: {ops2:?}"
        );

        // SUM(*) is still a plan error: Star is invalid for SUM.
        let err3 = plan_src("MATCH (a:N) RETURN SUM(*)").expect_err("SUM(*) must be plan error");
        assert!(
            err3.to_ascii_lowercase().contains("sum") || err3.to_ascii_lowercase().contains("*"),
            "error must mention SUM or *, got: {err3}"
        );
    }

    // ─── Grouped aggregation execution tests ─────────────────────────────────

    #[test]
    fn grouped_single_key_count() {
        // Graph: 3 nodes with "t" prop — two "X", one "Y".
        let mut fx = Fx::new();
        fx.add("N", "a", vec![("t", s("X"))]);
        fx.add("N", "b", vec![("t", s("X"))]);
        fx.add("N", "c", vec![("t", s("Y"))]);
        let v = fx.view();
        let params = BTreeMap::new();

        let rs = run(&v, "MATCH (n:N) RETURN n.t, COUNT(*) AS cnt", &params)
            .expect("single-key grouped COUNT must succeed");
        assert_eq!(
            rs.columns(),
            &["n.t".to_string(), "cnt".to_string()],
            "columns must match RETURN clause"
        );
        assert_eq!(rs.len(), 2, "must produce exactly 2 groups (X and Y)");

        // Find each group regardless of row order.
        let find = |label: &Value| (0..rs.len()).find(|&i| rs.row(i)[0].as_ref() == Some(label));
        let xi = find(&s("X")).expect("group X must exist");
        let yi = find(&s("Y")).expect("group Y must exist");
        assert_eq!(rs.row(xi)[1], Some(i(2)), "X group count must be 2");
        assert_eq!(rs.row(yi)[1], Some(i(1)), "Y group count must be 1");
    }

    #[test]
    fn grouped_two_keys_sum_avg() {
        // Four nodes with two categorical props and a numeric value.
        let mut fx = Fx::new();
        fx.add(
            "N",
            "a",
            vec![("cat", s("A")), ("sub", s("1")), ("v", i(10))],
        );
        fx.add(
            "N",
            "b",
            vec![("cat", s("A")), ("sub", s("1")), ("v", i(20))],
        );
        fx.add(
            "N",
            "c",
            vec![("cat", s("A")), ("sub", s("2")), ("v", i(5))],
        );
        fx.add(
            "N",
            "d",
            vec![("cat", s("B")), ("sub", s("1")), ("v", i(100))],
        );
        let v = fx.view();
        let params = BTreeMap::new();

        let rs = run(
            &v,
            "MATCH (n:N) RETURN n.cat, n.sub, SUM(n.v) AS total, AVG(n.v) AS avg_v",
            &params,
        )
        .expect("two-key SUM + AVG must succeed");
        assert_eq!(
            rs.columns(),
            &[
                "n.cat".to_string(),
                "n.sub".to_string(),
                "total".to_string(),
                "avg_v".to_string()
            ]
        );
        assert_eq!(rs.len(), 3, "must produce 3 groups: (A,1), (A,2), (B,1)");

        let find = |cat: &Value, sub: &Value| {
            (0..rs.len())
                .find(|&i| rs.row(i)[0].as_ref() == Some(cat) && rs.row(i)[1].as_ref() == Some(sub))
        };
        let a1 = find(&s("A"), &s("1")).expect("group (A,1) must exist");
        assert_eq!(rs.row(a1)[2], Some(i(30)), "(A,1) SUM must be 30");
        assert_eq!(rs.row(a1)[3], Some(f(15.0)), "(A,1) AVG must be 15.0");

        let a2 = find(&s("A"), &s("2")).expect("group (A,2) must exist");
        assert_eq!(rs.row(a2)[2], Some(i(5)), "(A,2) SUM must be 5");

        let b1 = find(&s("B"), &s("1")).expect("group (B,1) must exist");
        assert_eq!(rs.row(b1)[2], Some(i(100)), "(B,1) SUM must be 100");
        assert_eq!(rs.row(b1)[3], Some(f(100.0)), "(B,1) AVG must be 100.0");
    }

    #[test]
    fn grouped_order_by_count_desc_limit() {
        // 5 categories with different node counts: D=4, A=3, B=2, C=1, E=1.
        let mut fx = Fx::new();
        fx.add("N", "a1", vec![("cat", s("A"))]);
        fx.add("N", "a2", vec![("cat", s("A"))]);
        fx.add("N", "a3", vec![("cat", s("A"))]);
        fx.add("N", "b1", vec![("cat", s("B"))]);
        fx.add("N", "b2", vec![("cat", s("B"))]);
        fx.add("N", "c1", vec![("cat", s("C"))]);
        fx.add("N", "d1", vec![("cat", s("D"))]);
        fx.add("N", "d2", vec![("cat", s("D"))]);
        fx.add("N", "d3", vec![("cat", s("D"))]);
        fx.add("N", "d4", vec![("cat", s("D"))]);
        fx.add("N", "e1", vec![("cat", s("E"))]);
        let v = fx.view();
        let params = BTreeMap::new();

        let rs = run(
            &v,
            "MATCH (n:N) RETURN n.cat, COUNT(*) AS cnt ORDER BY cnt DESC LIMIT 3",
            &params,
        )
        .expect("ORDER BY count DESC LIMIT 3 must succeed");
        assert_eq!(rs.len(), 3, "LIMIT 3 must return exactly 3 groups");
        // Top 3: D(4), A(3), B(2) — descending order.
        assert_eq!(rs.row(0)[1], Some(i(4)), "row 0 must be count 4");
        assert_eq!(rs.row(0)[0], Some(s("D")), "row 0 must be category D");
        assert_eq!(rs.row(1)[1], Some(i(3)), "row 1 must be count 3");
        assert_eq!(rs.row(1)[0], Some(s("A")), "row 1 must be category A");
        assert_eq!(rs.row(2)[1], Some(i(2)), "row 2 must be count 2");
        assert_eq!(rs.row(2)[0], Some(s("B")), "row 2 must be category B");

        // Also verify row_bound is None for this plan (LIMIT must not be pushed).
        let plan_ops =
            plan_src("MATCH (n:N) RETURN n.cat, COUNT(*) AS cnt ORDER BY cnt DESC LIMIT 3")
                .expect("plan must succeed");
        assert_eq!(
            crate::cypher::plan::row_bound(&plan_ops),
            None,
            "GroupAggregate plan with LIMIT must have row_bound = None"
        );
    }

    #[test]
    fn grouped_empty_input_yields_zero_groups() {
        let fx = Fx::new(); // empty graph
        let v = fx.view();
        let params = BTreeMap::new();

        let rs = run(&v, "MATCH (n:N) RETURN n.t, COUNT(*) AS cnt", &params)
            .expect("grouped aggregate on empty graph must succeed");
        assert_eq!(rs.len(), 0, "empty input must yield zero groups");
        assert_eq!(
            rs.columns(),
            &["n.t".to_string(), "cnt".to_string()],
            "columns must still be present even with zero rows"
        );
    }

    #[test]
    fn grouped_null_key_groups_together() {
        // openCypher semantics: NULL group keys group together.
        let mut fx = Fx::new();
        fx.add("N", "a", vec![("t", s("X"))]);
        fx.add("N", "b", vec![]); // no "t" prop → null key
        fx.add("N", "c", vec![]); // null key — groups with b
        fx.add("N", "d", vec![("t", s("Y"))]);
        let v = fx.view();
        let params = BTreeMap::new();

        let rs = run(&v, "MATCH (n:N) RETURN n.t, COUNT(*) AS cnt", &params)
            .expect("null-key grouped aggregate must succeed");
        assert_eq!(rs.len(), 3, "must produce 3 groups: X, null, Y");

        // Locate the null group: row where n.t column is None (null).
        let null_row = (0..rs.len())
            .find(|&i| rs.row(i)[0].is_none())
            .expect("null group must be present");
        assert_eq!(
            rs.row(null_row)[1],
            Some(i(2)),
            "null group must count 2 rows (b and c)"
        );

        let x_row = (0..rs.len())
            .find(|&i| rs.row(i)[0] == Some(s("X")))
            .expect("X group must exist");
        assert_eq!(rs.row(x_row)[1], Some(i(1)));

        let y_row = (0..rs.len())
            .find(|&i| rs.row(i)[0] == Some(s("Y")))
            .expect("Y group must exist");
        assert_eq!(rs.row(y_row)[1], Some(i(1)));
    }

    #[test]
    fn grouped_cap_error_on_high_cardinality() {
        // Use with_max_groups to cap at 2 groups, then run a query that would
        // produce 3 groups → must return the named cap error.
        let mut fx = Fx::new();
        fx.add("N", "a", vec![("t", s("A"))]);
        fx.add("N", "b", vec![("t", s("B"))]);
        fx.add("N", "c", vec![("t", s("C"))]);
        let v = fx.view();
        let params = BTreeMap::new();

        let err = super::with_max_groups(2, || {
            run(&v, "MATCH (n:N) RETURN n.t, COUNT(*) AS cnt", &params)
        })
        .expect_err("must error when group count exceeds cap");
        assert!(
            err.to_ascii_lowercase().contains("group count"),
            "error must mention group count, got: {err}"
        );
    }

    #[test]
    fn multi_aggregate_no_keys() {
        // RETURN COUNT(*), COUNT(n) — multiple aggregates, no group keys.
        // Routes to GroupAggregate with empty keys.
        let mut fx = Fx::new();
        fx.add("N", "a", vec![]);
        fx.add("N", "b", vec![]);
        let v = fx.view();
        let params = BTreeMap::new();

        let rs = run(&v, "MATCH (n:N) RETURN COUNT(*), COUNT(n)", &params)
            .expect("multi-aggregate no keys must succeed");
        assert_eq!(rs.len(), 1, "must produce exactly one result row");
        assert_eq!(rs.row(0)[0], Some(i(2)), "COUNT(*) must be 2");
        assert_eq!(rs.row(0)[1], Some(i(2)), "COUNT(n) must be 2");
    }

    #[test]
    fn aggregate_over_hop_counts_edges() {
        let fx = hop_graph();
        let v = fx.view();
        let params = BTreeMap::new();
        // hop_graph has 3 KNOWS edges
        let rs = run(
            &v,
            "MATCH (a:Person)-[:KNOWS]->(b:Person) RETURN COUNT(*)",
            &params,
        )
        .expect("COUNT(*) on hop graph");
        assert_eq!(rs.row(0), &[Some(i(3))]);
    }

    // ──────────────────────────────────────────────────────────────────────────
    // C-1 fix tests: empty-input no-keys multi-aggregate must return 1 row
    // ──────────────────────────────────────────────────────────────────────────

    #[test]
    fn multi_aggregate_no_keys_empty_graph() {
        // C-1: RETURN COUNT(*), COUNT(n) on an empty graph must produce exactly
        // one row (COUNT=0, COUNT=0), not zero rows.  Routes to GroupAggregate
        // (keys=[], aggs=[Count, Count]).  Verifies parity with the single-agg
        // fast path, which also returns one row on empty input.
        let fx = Fx::new(); // empty — no nodes
        let v = fx.view();
        let params = BTreeMap::new();

        let rs = run(&v, "MATCH (n:N) RETURN COUNT(*), COUNT(n)", &params)
            .expect("empty-graph multi-agg must succeed");
        assert_eq!(rs.len(), 1, "must produce exactly 1 row on empty input");
        assert_eq!(
            rs.row(0)[0],
            Some(i(0)),
            "COUNT(*) on empty graph must be 0"
        );
        assert_eq!(
            rs.row(0)[1],
            Some(i(0)),
            "COUNT(n) on empty graph must be 0"
        );
    }

    #[test]
    fn fast_path_and_group_path_agree_on_empty_input() {
        // Equivalence pin: the single-agg fast path (Aggregate) and the
        // multi-agg GroupAggregate path must agree on COUNT for empty input.
        // Single-agg fast path: RETURN COUNT(*) → execute_aggregate.
        // Group path: RETURN COUNT(*), COUNT(n) → execute_group_aggregate.
        // Both must report COUNT = 0 on an empty graph.
        let fx = Fx::new();
        let v = fx.view();
        let params = BTreeMap::new();

        let fast = run(&v, "MATCH (n:N) RETURN COUNT(*)", &params)
            .expect("fast-path COUNT on empty graph");
        assert_eq!(fast.len(), 1, "fast path: 1 row on empty input");
        let fast_count = fast.row(0)[0].clone();

        let grouped = run(&v, "MATCH (n:N) RETURN COUNT(*), COUNT(n)", &params)
            .expect("group-path COUNT on empty graph");
        assert_eq!(grouped.len(), 1, "group path: 1 row on empty input");
        let group_count = grouped.row(0)[0].clone();

        assert_eq!(
            fast_count, group_count,
            "fast path and group path must agree on COUNT(*) for empty input"
        );

        // Same check on non-empty input: both paths should see count = 2.
        let mut fx2 = Fx::new();
        fx2.add("N", "x", vec![]);
        fx2.add("N", "y", vec![]);
        let v2 = fx2.view();

        let fast2 = run(&v2, "MATCH (n:N) RETURN COUNT(*)", &params)
            .expect("fast-path COUNT on 2-node graph");
        assert_eq!(fast2.row(0)[0], Some(i(2)), "fast path: COUNT(*) = 2");

        let grouped2 = run(&v2, "MATCH (n:N) RETURN COUNT(*), COUNT(n)", &params)
            .expect("group-path COUNT on 2-node graph");
        assert_eq!(grouped2.row(0)[0], Some(i(2)), "group path: COUNT(*) = 2");
        assert_eq!(grouped2.row(0)[1], Some(i(2)), "group path: COUNT(n) = 2");
    }

    // ──────────────────────────────────────────────────────────────────────────
    // I-1 fix tests: Int/Float group-key unification
    // ──────────────────────────────────────────────────────────────────────────

    #[test]
    fn grouped_int_float_key_unification() {
        // openCypher equality: 1 = 1.0.  Nodes with score=1 (Int) and score=1.0
        // (Float) must land in the same group after group_key_normalize maps
        // both to FloatBits.  Result: one group with count=2.
        //
        // Display value: first-seen wins.  Node "a" (Int(1)) is scanned first
        // (lower id), so the key column must display as Int(1), not Float(1.0).
        let mut fx = Fx::new();
        fx.add("N", "a", vec![("score", i(1))]); // scanned first → first-seen
        fx.add("N", "b", vec![("score", f(1.0))]);
        let v = fx.view();
        let params = BTreeMap::new();

        let rs = run(&v, "MATCH (n:N) RETURN n.score, COUNT(*) AS cnt", &params)
            .expect("int/float unification must succeed");
        assert_eq!(
            rs.len(),
            1,
            "Int(1) and Float(1.0) must group together into 1 group"
        );
        // Key column must display the first-seen original value (Int), not Float.
        assert_eq!(
            rs.row(0)[0],
            Some(i(1)),
            "key column must display Int(1) (first-seen)"
        );
        assert_eq!(rs.row(0)[1], Some(i(2)), "unified group must have count=2");
    }

    #[test]
    fn distinct_collapses_duplicate_projected_values() {
        let mut fx = Fx::new();
        fx.add("N", "a1", vec![("city", s("Austin"))]);
        fx.add("N", "a2", vec![("city", s("Austin"))]);
        let v = fx.view();
        let params = BTreeMap::new();
        let rs = run(&v, "MATCH (n:N) RETURN DISTINCT n.city", &params).unwrap();
        assert_eq!(rs.len(), 1);
        assert_eq!(rs.row(0)[0], Some(s("Austin")));
    }

    #[test]
    fn distinct_int_float_unify() {
        let mut fx = Fx::new();
        fx.add("N", "a", vec![("score", i(1))]);
        fx.add("N", "b", vec![("score", f(1.0))]);
        let v = fx.view();
        let params = BTreeMap::new();
        let rs = run(&v, "MATCH (n:N) RETURN DISTINCT n.score", &params).unwrap();
        assert_eq!(
            rs.len(),
            1,
            "Int(1) and Float(1.0) must DISTINCT as one row"
        );
        assert_eq!(rs.row(0)[0], Some(i(1)), "first-seen Int(1) wins display");
    }

    #[test]
    fn distinct_caps_at_intermediate_row_budget() {
        let mut fx = Fx::new();
        fx.add("N", "a", vec![("city", s("Austin"))]);
        fx.add("N", "b", vec![("city", s("Paris"))]);
        let v = fx.view();
        let params = BTreeMap::new();
        let err = super::with_max_intermediate_rows(1, || {
            run(&v, "MATCH (n:N) RETURN DISTINCT n.city", &params)
        })
        .expect_err("two distinct cities must exceed cap=1");
        assert!(
            err.contains("1") || err.contains("row"),
            "cap error must mention the budget, got: {err}"
        );
    }

    #[test]
    fn where_in_list_filters_rows() {
        let mut fx = Fx::new();
        fx.add("N", "a", vec![("city", s("Austin"))]);
        fx.add("N", "p", vec![("city", s("Paris"))]);
        fx.add("N", "l", vec![("city", s("London"))]);
        let v = fx.view();
        let mut params = BTreeMap::new();
        params.insert("c".into(), s("Paris"));
        let rs = run(
            &v,
            "MATCH (n:N) WHERE n.city IN ['Austin', $c] RETURN n.city",
            &params,
        )
        .unwrap();
        assert_eq!(rs.len(), 2);
    }

    #[test]
    fn pure_int_keys_display_as_int() {
        // N-1 regression: group keys that are integers must not be upcast to
        // Float in the output row.  Previously group_key_normalize converted
        // Int→FloatBits and value_key_to_value reconstructed Float(42.0) from
        // the key; now the original Value is stored and reused for display.
        let mut fx = Fx::new();
        fx.add("N", "a", vec![("age", i(10))]);
        fx.add("N", "b", vec![("age", i(20))]);
        fx.add("N", "c", vec![("age", i(10))]);
        let v = fx.view();
        let params = BTreeMap::new();

        let rs = run(
            &v,
            "MATCH (n:N) RETURN n.age, COUNT(*) AS cnt ORDER BY n.age",
            &params,
        )
        .expect("pure-Int group keys must succeed");
        assert_eq!(rs.len(), 2, "must have 2 groups: age=10 and age=20");
        // Key column must be Int, not Float.
        assert_eq!(
            rs.row(0)[0],
            Some(i(10)),
            "age=10 key column must display as Int(10)"
        );
        assert_eq!(rs.row(0)[1], Some(i(2)), "age=10 group has 2 nodes");
        assert_eq!(
            rs.row(1)[0],
            Some(i(20)),
            "age=20 key column must display as Int(20)"
        );
        assert_eq!(rs.row(1)[1], Some(i(1)), "age=20 group has 1 node");
    }

    // ──────────────────────────────────────────────────────────────────────────
    // M-1: VarExpand + GroupAggregate staged-path integration test
    // ──────────────────────────────────────────────────────────────────────────

    #[test]
    fn var_expand_group_aggregate_staged_path() {
        // Combines variable-length MATCH with a grouped RETURN — exercises the
        // staged-path GroupAggregate arm that groups over materialised VarExpand
        // rows.
        //
        // Graph: chain a -T-> b -T-> c  (3 nodes, 2 edges)
        // MATCH (x)-[*1..2]->(y) RETURN y.key, COUNT(*)
        //   1-hop results: (x=a, y=b), (x=b, y=c)       → b gets 1, c gets 1
        //   2-hop results: (x=a, y=c)                    → c gets 1 more
        //   So: b → 1, c → 2.
        let mut fx = Fx::new();
        let a = fx.add("N", "a", vec![("key", s("a"))]);
        let b = fx.add("N", "b", vec![("key", s("b"))]);
        let c = fx.add("N", "c", vec![("key", s("c"))]);
        fx.edge("T", a, b, vec![]);
        fx.edge("T", b, c, vec![]);
        let v = fx.view();
        let params = BTreeMap::new();

        let rs = run(
            &v,
            "MATCH (x:N)-[*1..2]->(y:N) RETURN y.key, COUNT(*) AS cnt ORDER BY y.key",
            &params,
        )
        .expect("VarExpand + GroupAggregate must succeed");

        assert_eq!(rs.len(), 2, "must have 2 destination groups: b and c");
        assert_eq!(rs.row(0)[0], Some(s("b")), "first group key must be 'b'");
        assert_eq!(rs.row(0)[1], Some(i(1)), "b is reached via 1 path");
        assert_eq!(rs.row(1)[0], Some(s("c")), "second group key must be 'c'");
        assert_eq!(
            rs.row(1)[1],
            Some(i(2)),
            "c is reached via 2 paths (1-hop and 2-hop)"
        );
    }

    // ──────────────────────────────────────────────────────────────────────────
    // pull_rows defense-in-depth: VarExpand and ShortestPath Err arms
    // These arms exist so that adding PlanOp variants forces a compile-time
    // decision in pull_rows.  Routing prevention is tested separately; these
    // tests verify the arms themselves are not dead code.
    // ──────────────────────────────────────────────────────────────────────────

    #[test]
    fn pull_rows_var_expand_arm_returns_named_err() {
        let fx = Fx::new();
        let view = fx.view();
        let vars = super::VarTable {
            names: vec!["a".into(), "b".into()],
        };
        let project_items: Vec<crate::cypher::ast::RetItem> = vec![];
        let empty_params = BTreeMap::new();
        let params = super::Params(&empty_params);
        let ctx = super::PullCtx {
            view: &view,
            vars: &vars,
            project_items: &project_items,
            params: &params,
            bound: 100,
        };
        let ops = vec![PlanOp::VarExpand {
            from: "a".into(),
            rel_var: None,
            etypes: vec![],
            dir: crate::cypher::RelDir::Right,
            to: "b".into(),
            min: 1,
            max: 3,
        }];
        let mut row = vec![None; vars.names.len()];
        let mut result = Vec::new();
        let err = super::pull_rows(&ctx, &ops, &mut row, &mut result)
            .expect_err("VarExpand must Err in pull_rows");
        assert!(
            err.contains("VarExpand") && err.contains("pull executor"),
            "error must name VarExpand and pull executor, got: {err}"
        );
    }

    #[test]
    fn pull_rows_shortest_path_arm_returns_named_err() {
        let fx = Fx::new();
        let view = fx.view();
        let vars = super::VarTable {
            names: vec!["a".into(), "b".into()],
        };
        let project_items: Vec<crate::cypher::ast::RetItem> = vec![];
        let empty_params = BTreeMap::new();
        let params = super::Params(&empty_params);
        let ctx = super::PullCtx {
            view: &view,
            vars: &vars,
            project_items: &project_items,
            params: &params,
            bound: 100,
        };
        let ops = vec![PlanOp::ShortestPath {
            from: "a".into(),
            rel_var: None,
            etypes: vec![],
            dir: crate::cypher::RelDir::Right,
            to: "b".into(),
            max_hops: 5,
        }];
        let mut row = vec![None; vars.names.len()];
        let mut result = Vec::new();
        let err = super::pull_rows(&ctx, &ops, &mut row, &mut result)
            .expect_err("ShortestPath must Err in pull_rows");
        assert!(
            err.contains("ShortestPath") && err.contains("pull executor"),
            "error must name ShortestPath and pull executor, got: {err}"
        );
    }

    // ── WITH / UNWIND pipeline tests ─────────────────────────────────────────

    /// Helper: build a graph with cities for pipeline tests.
    fn city_graph() -> Fx {
        let mut fx = Fx::new();
        fx.add(
            "Person",
            "alice",
            vec![("city", s("Boston")), ("age", i(30))],
        );
        fx.add("Person", "bob", vec![("city", s("Boston")), ("age", i(25))]);
        fx.add(
            "Person",
            "carol",
            vec![("city", s("Austin")), ("age", i(35))],
        );
        fx.add(
            "Person",
            "dave",
            vec![("city", s("Austin")), ("age", i(28))],
        );
        fx.add("Person", "eve", vec![("city", s("Boston")), ("age", i(22))]);
        fx
    }

    /// HAVING idiom: WITH a, COUNT(*) AS c WHERE c > 2.
    /// Boston has 3 people, Austin has 2. WHERE c > 2 keeps only Boston.
    #[test]
    fn with_aggregate_having_filters_groups() {
        let fx = city_graph();
        let view = fx.view();
        let params = BTreeMap::new();
        let rs = run(
            &view,
            "MATCH (p:Person) WITH p.city AS city, COUNT(*) AS cnt WHERE cnt > 2 RETURN city, cnt",
            &params,
        )
        .expect("WITH HAVING query must succeed");
        assert_eq!(rs.len(), 1, "only Boston group has cnt > 2");
        assert_eq!(rs.get(0, "city"), Some(&s("Boston")));
        assert_eq!(rs.get(0, "cnt"), Some(&i(3)));
    }

    /// WITH ORDER BY / LIMIT then RETURN.
    #[test]
    fn with_order_limit_then_return() {
        let fx = city_graph();
        let view = fx.view();
        let params = BTreeMap::new();
        // Get top 2 oldest people via WITH … ORDER BY age DESC LIMIT 2 RETURN name.
        let rs = run(&view, "MATCH (p:Person) WITH p, p.age AS age ORDER BY age DESC LIMIT 2 RETURN p.city AS city, age", &params)
            .expect("WITH ORDER LIMIT must succeed");
        assert_eq!(rs.len(), 2, "LIMIT 2");
        // carol=35 is first, alice=30 is second (or bob=25, dave=28 depending on tie)
        let ages: Vec<Option<Value>> = (0..rs.len()).map(|i| rs.get(i, "age").cloned()).collect();
        assert_eq!(ages[0], Some(i(35)), "oldest person first");
        assert_eq!(ages[1], Some(i(30)), "second oldest");
    }

    /// Chained WITH stages.
    #[test]
    fn chained_with_stages() {
        let fx = city_graph();
        let view = fx.view();
        let params = BTreeMap::new();
        // First WITH: project city alias. Second WITH: filter by city.
        let rs = run(&view,
            "MATCH (p:Person) WITH p.city AS city, p.age AS age WITH city, age WHERE age > 25 RETURN city, age",
            &params).expect("chained WITH must succeed");
        // alice=30, carol=35, dave=28 have age > 25; bob=25, eve=22 excluded
        assert_eq!(rs.len(), 3, "3 people with age > 25");
    }

    /// MATCH re-entry after WITH using a bound variable.
    #[test]
    fn with_then_match_reentry() {
        let mut fx = Fx::new();
        let alice = fx.add("Person", "alice", vec![]);
        let bob = fx.add("Person", "bob", vec![]);
        let corp = fx.add("Company", "acme", vec![]);
        fx.edge("WORKS_AT", alice, corp, vec![("years", i(5))]);
        fx.edge("WORKS_AT", bob, corp, vec![("years", i(3))]);
        let view = fx.view();
        let params = BTreeMap::new();
        // Find person, re-enter MATCH via their company.
        let rs = run(&view,
            "MATCH (p:Person)-[r:WORKS_AT]->(c:Company) WITH p, c MATCH (c)-[r2:WORKS_AT]-(colleague:Person) RETURN p, colleague",
            &params).expect("WITH MATCH re-entry must succeed");
        // alice→acme→bob, bob→acme→alice each appear (2 people × 2 colleagues)
        assert!(rs.len() >= 2, "cross join via company: got {}", rs.len());
    }

    /// UNWIND literal list produces one row per element.
    #[test]
    fn unwind_literal_list_produces_rows() {
        let fx = city_graph();
        let view = fx.view();
        let params = BTreeMap::new();
        let rs = run(
            &view,
            "MATCH (p:Person) WHERE p.city = 'Boston' UNWIND [1, 2, 3] AS x RETURN p, x",
            &params,
        )
        .expect("UNWIND literal must succeed");
        // 3 Boston people × 3 elements = 9 rows
        assert_eq!(rs.len(), 9, "UNWIND [1,2,3] × 3 Boston people = 9 rows");
        // All x values should include 1, 2, and 3.
        let xs: Vec<Option<Value>> = (0..rs.len())
            .map(|row_i| rs.get(row_i, "x").cloned())
            .collect();
        assert!(xs.contains(&Some(i(1))));
        assert!(xs.contains(&Some(i(2))));
        assert!(xs.contains(&Some(i(3))));
    }

    /// UNWIND of a list-valued property.
    #[test]
    fn unwind_list_property() {
        let mut fx = Fx::new();
        fx.add(
            "Tag",
            "post1",
            vec![("tags", Value::List(vec![s("rust"), s("graph")]))],
        );
        fx.add("Tag", "post2", vec![("tags", Value::List(vec![s("db")]))]);
        let view = fx.view();
        let params = BTreeMap::new();
        let rs = run(
            &view,
            "MATCH (p:Tag) UNWIND p.tags AS tag RETURN p, tag",
            &params,
        )
        .expect("UNWIND property must succeed");
        assert_eq!(rs.len(), 3, "2+1 tag elements");
    }

    /// UNWIND null / empty list → 0 rows (openCypher).
    #[test]
    fn unwind_empty_list_yields_zero_rows() {
        let fx = city_graph();
        let view = fx.view();
        let params = BTreeMap::new();
        let rs = run(
            &view,
            "MATCH (p:Person) WHERE p.city = 'Boston' UNWIND [] AS x RETURN x",
            &params,
        )
        .expect("UNWIND [] must succeed with 0 rows");
        assert_eq!(rs.len(), 0, "UNWIND [] should produce 0 rows");
    }

    /// UNWIND of a non-list → named error.
    #[test]
    fn unwind_non_list_is_named_error() {
        let fx = city_graph();
        let view = fx.view();
        let params = BTreeMap::new();
        // p.city is a Str, not a List.
        let err = run(
            &view,
            "MATCH (p:Person) WHERE p.city = 'Boston' UNWIND p.city AS x RETURN x",
            &params,
        )
        .expect_err("UNWIND non-list must error");
        assert!(
            err.contains("UNWIND") && err.contains("list"),
            "error must mention UNWIND and list: {err}"
        );
    }

    /// UNWIND cross-product that trips the intermediate-row budget.
    #[test]
    fn unwind_cross_product_trips_budget() {
        let mut fx = Fx::new();
        // 10 nodes, each UNWIND 10 elements → 100 rows; set cap to 5.
        for i in 0..10 {
            fx.add("N", &format!("n{i}"), vec![]);
        }
        let view = fx.view();
        let params = BTreeMap::new();
        let err = super::with_max_intermediate_rows(5, || {
            run(
                &view,
                "MATCH (n:N) UNWIND [1, 2, 3, 4, 5, 6, 7, 8, 9, 10] AS x RETURN n, x",
                &params,
            )
        })
        .expect_err("must hit budget");
        assert!(
            err.contains("intermediate result exceeds"),
            "error must mention intermediate result limit: {err}"
        );
    }

    /// Routing pins: plans with WITH take staged path (row_bound returns None);
    /// non-WITH plans are unchanged.
    #[test]
    fn routing_with_plan_uses_staged_path() {
        use crate::cypher::plan::row_bound;
        let ops_with = compile("MATCH (a) WITH a RETURN a");
        assert_eq!(
            row_bound(&ops_with),
            None,
            "WITH plan must have row_bound=None"
        );

        let ops_unwind = compile("MATCH (a) UNWIND [1, 2] AS x RETURN a");
        assert_eq!(
            row_bound(&ops_unwind),
            None,
            "UNWIND plan must have row_bound=None"
        );

        // Non-WITH plan with LIMIT still uses pull path.
        let ops_plain = compile("MATCH (a) RETURN a LIMIT 5");
        assert!(
            row_bound(&ops_plain).is_some(),
            "plain LIMIT plan must have a row bound"
        );
    }

    /// I-1: UNWIND of a null/absent property → 0 rows for that node; other nodes
    /// with valid lists still expand normally.
    #[test]
    fn unwind_null_property_yields_zero_rows_for_that_node() {
        let mut fx = Fx::new();
        // post1 has a list property; post2 has no `tags` property at all (null).
        fx.add(
            "Post",
            "post1",
            vec![("tags", Value::List(vec![s("rust"), s("graph")]))],
        );
        fx.add("Post", "post2", vec![("other", Value::Str("hello".into()))]);
        let view = fx.view();
        let params = BTreeMap::new();
        let rs = run(
            &view,
            "MATCH (p:Post) UNWIND p.tags AS tag RETURN tag",
            &params,
        )
        .expect("UNWIND null property must succeed (not error)");
        // post1 contributes 2 rows; post2 contributes 0 rows (null → skip).
        assert_eq!(
            rs.len(),
            2,
            "null property must yield 0 rows; list property yields N rows"
        );
        let tag0 = rs.get(0, "tag");
        let tag1 = rs.get(1, "tag");
        let tags = [tag0, tag1];
        assert!(tags.contains(&Some(&s("rust"))), "rust tag present");
        assert!(tags.contains(&Some(&s("graph"))), "graph tag present");
    }

    /// I-2a: WHERE before UNWIND (pre-expansion filter — existing grammar).
    #[test]
    fn where_before_unwind_filters_nodes() {
        let mut fx = Fx::new();
        fx.add(
            "P",
            "a",
            vec![
                ("group", Value::Str("keep".into())),
                ("items", Value::List(vec![Value::Int(1), Value::Int(2)])),
            ],
        );
        fx.add(
            "P",
            "b",
            vec![
                ("group", Value::Str("drop".into())),
                ("items", Value::List(vec![Value::Int(3), Value::Int(4)])),
            ],
        );
        let view = fx.view();
        let params = BTreeMap::new();
        // WHERE filters before UNWIND: only node `a` (group=keep) expands.
        let rs = run(
            &view,
            "MATCH (n:P) WHERE n.group = 'keep' UNWIND n.items AS x RETURN x",
            &params,
        )
        .expect("WHERE before UNWIND must succeed");
        assert_eq!(rs.len(), 2, "only matching node expands; 2 items");
    }

    /// I-2b: UNWIND then WHERE (post-expansion filter — new grammar support).
    #[test]
    fn where_after_unwind_filters_expanded_rows() {
        let mut fx = Fx::new();
        fx.add(
            "N",
            "n1",
            vec![(
                "vals",
                Value::List(vec![Value::Int(1), Value::Int(3), Value::Int(5)]),
            )],
        );
        let view = fx.view();
        let params = BTreeMap::new();
        // WHERE after UNWIND: filters by the alias `x`; keeps only x > 2.
        let rs = run(
            &view,
            "MATCH (n:N) UNWIND n.vals AS x WHERE x > 2 RETURN x",
            &params,
        )
        .expect("WHERE after UNWIND must succeed");
        assert_eq!(rs.len(), 2, "x=3 and x=5 pass; x=1 filtered out");
        let v0 = rs.get(0, "x").cloned();
        let v1 = rs.get(1, "x").cloned();
        let vals = [v0, v1];
        assert!(vals.contains(&Some(Value::Int(3))), "x=3 present");
        assert!(vals.contains(&Some(Value::Int(5))), "x=5 present");
    }

    /// I-3: ORDER BY with a typo'd alias in aggregate WITH → named error at plan time.
    #[test]
    fn aggregate_with_order_by_unknown_alias_is_named_error() {
        use crate::cypher::{lex, parser::parse, plan::plan};
        // `cntt` is a typo; the real column is `cnt` — must fail at planning.
        let src = "MATCH (p:Person) WITH p.city AS city, COUNT(*) AS cnt ORDER BY cntt DESC RETURN city, cnt";
        let plan_result = plan(&parse(&lex(src).expect("lex")).expect("parse"));
        let err = plan_result
            .expect_err("typo'd ORDER BY alias in aggregate WITH must be a named plan error");
        assert!(
            err.contains("cntt") || err.contains("unbound"),
            "error must mention the unknown alias: {err}"
        );
    }

    /// I-3 non-aggregate path: ORDER BY with unknown alias → named error at plan time.
    #[test]
    fn non_aggregate_with_order_by_unknown_alias_is_named_error() {
        use crate::cypher::{lex, parser::parse, plan::plan};
        let src = "MATCH (p:Person) WITH p, p.age AS age ORDER BY nope DESC RETURN p.city AS city";
        let plan_result = plan(&parse(&lex(src).expect("lex")).expect("parse"));
        let err = plan_result
            .expect_err("typo'd ORDER BY alias in non-aggregate WITH must be a named plan error");
        assert!(
            err.contains("nope") || err.contains("unbound"),
            "error must mention the unknown alias: {err}"
        );
    }

    /// Round-2 pin 1: post-UNWIND WHERE referencing a pre-MATCH-bound variable.
    /// Verifies scope is correctly available across the UNWIND boundary —
    /// a regression that drops `n` from scope would either error or return all rows.
    #[test]
    fn post_unwind_where_references_pre_match_variable() {
        let mut fx = Fx::new();
        // n1: threshold=3, list=[1,2,3,4,5] → keep 4,5 (x > threshold)
        fx.add(
            "N",
            "n1",
            vec![
                ("threshold", Value::Int(3)),
                (
                    "xs",
                    Value::List(vec![
                        Value::Int(1),
                        Value::Int(2),
                        Value::Int(3),
                        Value::Int(4),
                        Value::Int(5),
                    ]),
                ),
            ],
        );
        // n2: threshold=10, list=[1,2] → keep nothing
        fx.add(
            "N",
            "n2",
            vec![
                ("threshold", Value::Int(10)),
                ("xs", Value::List(vec![Value::Int(1), Value::Int(2)])),
            ],
        );
        let view = fx.view();
        let params = BTreeMap::new();
        let rs = run(
            &view,
            "MATCH (n:N) UNWIND n.xs AS x WHERE x > n.threshold RETURN x",
            &params,
        )
        .expect("post-UNWIND WHERE referencing MATCH variable must succeed");
        // n1 contributes x=4 and x=5; n2 contributes nothing.
        assert_eq!(rs.len(), 2, "exactly 2 rows: x=4 and x=5 from n1");
        let v0 = rs.get(0, "x").cloned();
        let v1 = rs.get(1, "x").cloned();
        let got = [v0, v1];
        assert!(got.contains(&Some(Value::Int(4))), "x=4 present");
        assert!(got.contains(&Some(Value::Int(5))), "x=5 present");
    }

    /// Round-2 pin 2: ORDER BY on an aggregate alias INSIDE a WITH stage.
    /// Verifies compile_with_stage emits OrderBy that exec_order_by_rows
    /// applies to Cell::Scalar group rows — the outer q.order_by path is separate.
    #[test]
    fn with_stage_aggregate_order_by_descending() {
        // Boston: 3 people (alice age=30, bob age=25, eve age=22)
        // Austin: 2 people (carol age=35, dave age=28)
        let fx = city_graph();
        let view = fx.view();
        let params = BTreeMap::new();
        let rs = run(
            &view,
            "MATCH (p:Person) WITH p.city AS city, COUNT(*) AS cnt ORDER BY cnt DESC RETURN city, cnt",
            &params,
        )
        .expect("aggregate WITH ORDER BY must succeed");
        assert_eq!(rs.len(), 2, "two city groups");
        // With ORDER BY cnt DESC: Boston (3) first, Austin (2) second.
        assert_eq!(
            rs.get(0, "cnt"),
            Some(&Value::Int(3)),
            "first row must be the group with cnt=3 (Boston)"
        );
        assert_eq!(
            rs.get(1, "cnt"),
            Some(&Value::Int(2)),
            "second row must be the group with cnt=2 (Austin)"
        );
    }

    /// Round-2 pin 3: UNWIND-then-WITH composition.
    /// Verifies the pipeline correctly threads UNWIND-expanded rows into a
    /// downstream WITH stage (both simple pass-through and aggregation).
    #[test]
    fn unwind_then_with_composition() {
        let mut fx = Fx::new();
        fx.add(
            "Doc",
            "d1",
            vec![(
                "scores",
                Value::List(vec![Value::Int(10), Value::Int(20), Value::Int(30)]),
            )],
        );
        fx.add(
            "Doc",
            "d2",
            vec![("scores", Value::List(vec![Value::Int(5), Value::Int(15)]))],
        );
        let view = fx.view();
        let params = BTreeMap::new();

        // Simple pass-through: UNWIND then WITH x (non-aggregate).
        let rs = run(
            &view,
            "MATCH (n:Doc) UNWIND n.scores AS x WITH x RETURN x",
            &params,
        )
        .expect("UNWIND then WITH pass-through must succeed");
        assert_eq!(rs.len(), 5, "3 + 2 = 5 expanded rows carried through WITH");

        // Aggregation over UNWIND: COUNT all expanded values.
        let rs2 = run(
            &view,
            "MATCH (n:Doc) UNWIND n.scores AS x WITH COUNT(*) AS total RETURN total",
            &params,
        )
        .expect("UNWIND then aggregate WITH must succeed");
        assert_eq!(rs2.len(), 1, "single aggregate row");
        assert_eq!(
            rs2.get(0, "total"),
            Some(&Value::Int(5)),
            "total must be 5 (3+2 expanded rows)"
        );
    }

    /// Regression: i64::MIN / -1 must not panic (checked_div saturates to i64::MAX).
    /// Before the fix, the BinArith Div branch used plain `a / b` which panics on overflow
    /// in both debug and release. ArithOp::Div is not reachable via the current parser
    /// (the lexer rejects '/' as an illegal character), so this test calls resolve_operand
    /// directly to pin the exact overflow case.
    #[test]
    fn binarith_div_min_over_neg1_does_not_panic() {
        let fx = Fx::new();
        let view = fx.view();
        let vars = VarTable { names: vec![] };
        let row: Row = vec![];
        let params = BTreeMap::new();

        // Construct: i64::MIN / -1  (overflows — checked_div returns None → saturates to i64::MAX)
        let operand = Operand::BinArith {
            op: ArithOp::Div,
            left: Box::new(Operand::Lit(Value::Int(i64::MIN))),
            right: Box::new(Operand::Lit(Value::Int(-1))),
        };

        let result = resolve_operand(&view, &vars, &row, &operand, &Params(&params));
        assert!(
            result.is_ok(),
            "overflow division must not return Err: {result:?}"
        );
        assert_eq!(
            result.unwrap(),
            Some(Value::Int(i64::MAX)),
            "i64::MIN / -1 must saturate to i64::MAX, not panic"
        );
    }

    /// Regression: collect_operand was missing the BinArith arm, so $params inside
    /// arithmetic expressions were invisible to pre-flight validation.  A query like
    /// `RETURN abs($missing - 1)` with no $missing supplied must fail at pre-flight
    /// with a named-param error, not silently return null.
    #[test]
    fn binarith_missing_param_caught_at_preflight() {
        let fx = Fx::new();
        let view = fx.view();
        // Parser requires MATCH; bare RETURN is not supported.
        // The param is inside a BinArith sub-expression wrapped in a scalar
        // function call (`abs`): collect_operand must recurse into BinArith
        // left/right to surface $missing at pre-flight time.
        let err = run(
            &view,
            "MATCH (n:X) RETURN abs($missing - 1)",
            &BTreeMap::new(),
        )
        .expect_err("missing param inside BinArith must be caught at preflight");
        assert!(
            err.contains("missing"),
            "error must name the param 'missing', got: {err}"
        );
    }

    // ── IS NULL / IS NOT NULL evaluation ─────────────────────────────────────

    /// `WHERE n.missing IS NULL` must return nodes that lack the property.
    #[test]
    fn is_null_filters_absent_prop() {
        let mut fx = Fx::new();
        fx.add("Person", "alice", vec![("age", Value::Int(30))]);
        fx.add("Person", "bob", vec![]); // no `age` prop
        let view = fx.view();
        let rs = run(
            &view,
            "MATCH (n:Person) WHERE n.age IS NULL RETURN n",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(rs.len(), 1, "only bob lacks age");
        assert_eq!(rs.get(0, "n"), Some(&s("bob")));
    }

    /// `WHERE n.prop IS NOT NULL` must return nodes that have the property.
    #[test]
    fn is_not_null_filters_present_prop() {
        let mut fx = Fx::new();
        fx.add("Person", "alice", vec![("age", Value::Int(30))]);
        fx.add("Person", "bob", vec![]); // no `age` prop
        let view = fx.view();
        let rs = run(
            &view,
            "MATCH (n:Person) WHERE n.age IS NOT NULL RETURN n",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(rs.len(), 1, "only alice has age");
        assert_eq!(rs.get(0, "n"), Some(&s("alice")));
    }

    /// Anti-join idiom: OPTIONAL MATCH (a)-[:T]->(b) WHERE b IS NULL
    /// returns nodes that have no outgoing T edge.
    #[test]
    fn optional_match_is_null_anti_join() {
        let mut fx = Fx::new();
        let alice = fx.add("Person", "alice", vec![]);
        let bob = fx.add("Person", "bob", vec![]);
        let carol = fx.add("Person", "carol", vec![]);
        fx.edge("KNOWS", alice, carol, vec![]); // alice → carol
        fx.edge("KNOWS", carol, bob, vec![]); // carol → bob
                                              // bob has no outgoing KNOWS edge; alice and carol both do.
        let view = fx.view();
        let rs = run(
            &view,
            "MATCH (a:Person) OPTIONAL MATCH (a)-[:KNOWS]->(b) WITH a, b WHERE b IS NULL RETURN a",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(rs.len(), 1, "only bob has no outgoing KNOWS edge");
        assert_eq!(rs.get(0, "a"), Some(&s("bob")));
    }

    /// IS NULL composes with AND correctly.
    #[test]
    fn is_null_combined_with_and_exec() {
        let mut fx = Fx::new();
        fx.add("N", "a", vec![("x", Value::Int(1))]);
        fx.add("N", "b", vec![("x", Value::Int(2)), ("y", Value::Int(9))]);
        fx.add("N", "c", vec![]); // no x, no y
        let view = fx.view();
        // WHERE n.y IS NULL AND n.x IS NOT NULL → only a (x=1, no y)
        let rs = run(
            &view,
            "MATCH (n:N) WHERE n.y IS NULL AND n.x IS NOT NULL RETURN n",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(rs.len(), 1);
        assert_eq!(rs.get(0, "n"), Some(&s("a")));
    }

    // ── Arithmetic evaluation ─────────────────────────────────────────────────

    /// `n.age + 1` in RETURN produces ScalarExpr result per row.
    #[test]
    fn arithmetic_add_in_return() {
        let mut fx = Fx::new();
        fx.add("Person", "alice", vec![("age", Value::Int(29))]);
        let view = fx.view();
        let rs = run(
            &view,
            "MATCH (n:Person) RETURN n.age + 1 AS adjusted",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(rs.len(), 1);
        assert_eq!(rs.get(0, "adjusted"), Some(&Value::Int(30)));
    }

    /// Parentheses override default precedence: (1+2)*3 = 9, not 1+(2*3)=7.
    #[test]
    fn arithmetic_precedence_parens_pin() {
        let mut fx = Fx::new();
        fx.add("N", "x", vec![]);
        let view = fx.view();
        let rs = run(
            &view,
            "MATCH (n:N) RETURN (1 + 2) * 3 AS r",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(rs.len(), 1);
        assert_eq!(rs.get(0, "r"), Some(&Value::Int(9)));
    }

    /// Without parens, 1+2*3 = 7 (multiplication over addition).
    #[test]
    fn arithmetic_precedence_mul_over_add_pin() {
        let mut fx = Fx::new();
        fx.add("N", "x", vec![]);
        let view = fx.view();
        let rs = run(&view, "MATCH (n:N) RETURN 1 + 2 * 3 AS r", &BTreeMap::new()).unwrap();
        assert_eq!(rs.len(), 1);
        assert_eq!(rs.get(0, "r"), Some(&Value::Int(7)));
    }

    /// Division by zero in an arithmetic expression returns an error (not panic).
    #[test]
    fn arithmetic_div_by_zero_returns_error() {
        let mut fx = Fx::new();
        fx.add("N", "x", vec![]);
        let view = fx.view();
        let err = run(
            &view,
            "MATCH (n:N) WHERE 1 / 0 > 0 RETURN n",
            &BTreeMap::new(),
        )
        .expect_err("division by zero must error");
        assert!(
            err.contains("division by zero"),
            "error must mention division by zero, got: {err}"
        );
    }

    /// Null propagation: arithmetic on a null operand yields null (filter excludes row).
    #[test]
    fn arithmetic_null_propagates() {
        let mut fx = Fx::new();
        fx.add("N", "x", vec![]); // no `val` prop → null
        let view = fx.view();
        // WHERE n.val + 1 > 0 — null propagates → false → no rows
        let rs = run(
            &view,
            "MATCH (n:N) WHERE n.val + 1 > 0 RETURN n",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(rs.len(), 0, "null arithmetic must not match");
    }

    /// Arithmetic in a WHERE comparison: `n.age + 1 > 5` filters correctly.
    #[test]
    fn case_when_expression_in_return() {
        let mut fx = Fx::new();
        fx.add("N", "a", vec![("id", s("a")), ("age", i(20))]);
        fx.add("N", "b", vec![("id", s("b")), ("age", i(65))]);
        let v = fx.view();
        let rs = run(
            &v,
            "MATCH (n:N) RETURN n.id AS id, \
             CASE WHEN n.age >= 65 THEN 'senior' ELSE 'other' END AS band",
            &BTreeMap::new(),
        )
        .unwrap();
        let band_of = |who: &str| {
            (0..rs.len())
                .find(|&i| rs.get(i, "id") == Some(&s(who)))
                .and_then(|i| rs.get(i, "band").cloned())
        };
        assert_eq!(band_of("a"), Some(s("other")));
        assert_eq!(band_of("b"), Some(s("senior")));
    }

    #[test]
    fn case_when_no_else_yields_null() {
        let mut fx = Fx::new();
        fx.add("N", "a", vec![("age", i(20))]);
        let v = fx.view();
        let rs = run(
            &v,
            "MATCH (n:N) RETURN CASE WHEN n.age >= 65 THEN 'senior' END AS band",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(rs.get(0, "band"), None);
    }

    #[test]
    fn multi_relationship_type_pattern_matches_either() {
        let mut fx = Fx::new();
        let a = fx.add("N", "a", vec![("id", s("a"))]);
        let b = fx.add("N", "b", vec![("id", s("b"))]);
        let c = fx.add("N", "c", vec![("id", s("c"))]);
        let d = fx.add("N", "d", vec![("id", s("d"))]);
        fx.edge("KNOWS", a, b, vec![]);
        fx.edge("LIKES", a, c, vec![]);
        fx.edge("HATES", a, d, vec![]);
        let v = fx.view();
        // a -[:KNOWS|:LIKES]-> should reach b and c, not d.
        let rs = run(
            &v,
            "MATCH (a:N {id: 'a'})-[r:KNOWS|:LIKES]->(x) RETURN x",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(rs.len(), 2, "KNOWS|LIKES reaches exactly b and c");
    }

    #[test]
    fn collect_grouped_gathers_values_per_group() {
        let mut fx = Fx::new();
        fx.add("P", "a", vec![("city", s("austin")), ("name", s("Ann"))]);
        fx.add("P", "b", vec![("city", s("austin")), ("name", s("Bob"))]);
        fx.add("P", "c", vec![("city", s("boston")), ("name", s("Cy"))]);
        let v = fx.view();
        let rs = run(
            &v,
            "MATCH (n:P) RETURN n.city AS city, collect(n.name) AS names",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(rs.len(), 2, "two city groups");
        // Locate the austin row and check its collected names.
        let austin = (0..rs.len())
            .find(|&i| rs.get(i, "city") == Some(&s("austin")))
            .expect("austin group present");
        assert_eq!(
            rs.get(austin, "names"),
            Some(&Value::List(vec![s("Ann"), s("Bob")]))
        );
    }

    #[test]
    fn collect_ungrouped_gathers_all_into_one_list() {
        let mut fx = Fx::new();
        fx.add("P", "a", vec![("name", s("Ann"))]);
        fx.add("P", "b", vec![("name", s("Bob"))]);
        let v = fx.view();
        let rs = run(
            &v,
            "MATCH (n:P) RETURN collect(n.name) AS names",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(rs.len(), 1);
        assert_eq!(
            rs.get(0, "names"),
            Some(&Value::List(vec![s("Ann"), s("Bob")]))
        );
    }

    #[test]
    fn string_predicate_functions_in_where() {
        let mut fx = Fx::new();
        fx.add("N", "a", vec![("email", s("alice@acme.com"))]);
        fx.add("N", "b", vec![("email", s("bob@other.org"))]);
        let v = fx.view();
        let p = BTreeMap::new();
        assert_eq!(
            run(
                &v,
                "MATCH (n:N) WHERE endsWith(n.email, '.com') RETURN n",
                &p
            )
            .unwrap()
            .len(),
            1
        );
        assert_eq!(
            run(
                &v,
                "MATCH (n:N) WHERE startsWith(n.email, 'bob') RETURN n",
                &p
            )
            .unwrap()
            .len(),
            1
        );
        assert_eq!(
            run(
                &v,
                "MATCH (n:N) WHERE contains(n.email, 'acme') RETURN n",
                &p
            )
            .unwrap()
            .len(),
            1
        );
    }

    #[test]
    fn coercion_functions_in_return() {
        let mut fx = Fx::new();
        fx.add("N", "a", vec![("s", s("42")), ("n", i(7)), ("g", f(3.9))]);
        let v = fx.view();
        let rs = run(
            &v,
            "MATCH (n:N) RETURN toInteger(n.s) AS ti, toFloat(n.n) AS tf, toString(n.g) AS ts",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(rs.get(0, "ti"), Some(&Value::Int(42)));
        assert_eq!(rs.get(0, "tf"), Some(&Value::Float(7.0)));
        assert_eq!(rs.get(0, "ts"), Some(&Value::Str("3.9".into())));
    }

    #[test]
    fn to_integer_unparseable_string_is_null() {
        let mut fx = Fx::new();
        fx.add("N", "a", vec![("s", s("not-a-number"))]);
        let v = fx.view();
        let rs = run(
            &v,
            "MATCH (n:N) RETURN toInteger(n.s) AS ti",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(rs.get(0, "ti"), None);
    }

    #[test]
    fn index_scan_uses_index_and_matches_fallback() {
        use core_storage::property_index::PropertyIndex;
        use std::sync::atomic::Ordering;

        let mut fx = Fx::new();
        let a = fx.add("Person", "a", vec![("city", s("austin"))]);
        let _b = fx.add("Person", "b", vec![("city", s("boston"))]);
        let c = fx.add("Person", "c", vec![("city", s("austin"))]);

        let mut pi = PropertyIndex::new();
        pi.enable("Person", "city");
        pi.set("Person", "city", a, &s("austin"));
        pi.set("Person", "city", 1, &s("boston"));
        pi.set("Person", "city", c, &s("austin"));

        let q = "MATCH (n:Person {city: 'austin'}) RETURN n";

        // Indexed view: the IndexScan fast path must fire and return both nodes.
        let before = super::INDEX_SCAN_FIRES.load(Ordering::Relaxed);
        let indexed = run(&fx.view_indexed(&pi), q, &BTreeMap::new()).unwrap();
        let after = super::INDEX_SCAN_FIRES.load(Ordering::Relaxed);
        assert!(after > before, "IndexScan must take the indexed path");
        assert_eq!(indexed.len(), 2);

        // Unindexed view (prop_index None): same result via scan+filter fallback,
        // and the counter must NOT advance.
        let before2 = super::INDEX_SCAN_FIRES.load(Ordering::Relaxed);
        let fallback = run(&fx.view(), q, &BTreeMap::new()).unwrap();
        let after2 = super::INDEX_SCAN_FIRES.load(Ordering::Relaxed);
        assert_eq!(after2, before2, "fallback must not touch the index counter");
        assert_eq!(
            fallback.len(),
            indexed.len(),
            "fallback matches indexed result"
        );
    }

    #[test]
    fn arithmetic_in_where_comparison() {
        let mut fx = Fx::new();
        fx.add("Person", "alice", vec![("age", Value::Int(5))]); // 5+1=6 > 5 → match
        fx.add("Person", "bob", vec![("age", Value::Int(4))]); // 4+1=5 not > 5 → skip
        let view = fx.view();
        let rs = run(
            &view,
            "MATCH (n:Person) WHERE n.age + 1 > 5 RETURN n",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(rs.len(), 1);
        assert_eq!(rs.get(0, "n"), Some(&s("alice")));
    }

    // --- WHERE equality fold executor equivalence tests (T1) ---

    /// Folded WHERE equality uses index when available; fallback returns same count.
    #[test]
    fn where_fold_indexed_matches_fallback() {
        use core_storage::property_index::PropertyIndex;
        use std::sync::atomic::Ordering;

        let mut fx = Fx::new();
        let a = fx.add("Person", "alice", vec![("city", s("austin"))]);
        let b = fx.add("Person", "bob", vec![("city", s("boston"))]);
        let c = fx.add("Person", "carol", vec![("city", s("austin"))]);

        let mut pi = PropertyIndex::new();
        pi.enable("Person", "city");
        pi.set("Person", "city", a, &s("austin"));
        pi.set("Person", "city", b, &s("boston"));
        pi.set("Person", "city", c, &s("austin"));

        let q = "MATCH (n:Person) WHERE n.city = 'austin' RETURN n";

        // Indexed path must fire and return both austin nodes.
        let before = super::INDEX_SCAN_FIRES.load(Ordering::Relaxed);
        let indexed = run(&fx.view_indexed(&pi), q, &BTreeMap::new()).unwrap();
        let after = super::INDEX_SCAN_FIRES.load(Ordering::Relaxed);
        assert!(
            after > before,
            "WHERE equality fold must take the indexed path"
        );
        assert_eq!(indexed.len(), 2);

        // Fallback (no index) must return byte-identical rows in the same order.
        let fallback = run(&fx.view(), q, &BTreeMap::new()).unwrap();
        assert_eq!(
            rows_of(&fallback),
            rows_of(&indexed),
            "fallback must return identical rows, not just the same count"
        );
    }

    /// Folded WHERE equality that matches no nodes returns empty result.
    #[test]
    fn where_fold_miss_returns_empty() {
        let mut fx = Fx::new();
        fx.add("Person", "alice", vec![("city", s("austin"))]);
        let q = "MATCH (n:Person) WHERE n.city = 'berlin' RETURN n";
        let rs = run(&fx.view(), q, &BTreeMap::new()).unwrap();
        assert_eq!(rs.len(), 0);
    }

    /// WHERE equality with a $param folds to IndexScan and uses the index.
    #[test]
    fn where_fold_param_uses_index() {
        use core_storage::property_index::PropertyIndex;
        use std::sync::atomic::Ordering;

        let mut fx = Fx::new();
        let a = fx.add("Person", "alice", vec![("city", s("austin"))]);
        let b = fx.add("Person", "bob", vec![("city", s("boston"))]);

        let mut pi = PropertyIndex::new();
        pi.enable("Person", "city");
        pi.set("Person", "city", a, &s("austin"));
        pi.set("Person", "city", b, &s("boston"));

        let q = "MATCH (n:Person) WHERE n.city = $c RETURN n";
        let mut params = BTreeMap::new();
        params.insert("c".to_string(), s("austin"));

        let before = super::INDEX_SCAN_FIRES.load(Ordering::Relaxed);
        let rs = run(&fx.view_indexed(&pi), q, &params).unwrap();
        let after = super::INDEX_SCAN_FIRES.load(Ordering::Relaxed);
        assert!(after > before, "$param WHERE equality must use index");
        assert_eq!(rs.len(), 1);
    }

    /// Residual AND predicate is applied after the IndexScan fold.
    #[test]
    fn where_fold_residual_filter_applied() {
        let mut fx = Fx::new();
        fx.add(
            "Person",
            "young-austin",
            vec![("city", s("austin")), ("age", Value::Int(20))],
        );
        fx.add(
            "Person",
            "old-austin",
            vec![("city", s("austin")), ("age", Value::Int(40))],
        );
        fx.add(
            "Person",
            "boston",
            vec![("city", s("boston")), ("age", Value::Int(20))],
        );

        let q = "MATCH (n:Person) WHERE n.city = 'austin' AND n.age > 30 RETURN n";
        let rs = run(&fx.view(), q, &BTreeMap::new()).unwrap();
        assert_eq!(rs.len(), 1, "only old-austin should match city+age filter");
        assert_eq!(rs.get(0, "n"), Some(&s("old-austin")));
    }

    /// IndexScan for an unindexed field falls back to scan+filter; result is correct.
    /// Counter correctness for the fallback path is already covered by the
    /// canonical `index_scan_uses_index_and_matches_fallback` test; this test
    /// focuses on result equivalence when WHERE folds to IndexScan on a field
    /// that has no declared index.
    #[test]
    fn where_fold_unindexed_field_fallback() {
        let mut fx = Fx::new();
        fx.add("Person", "a", vec![("notindexed", s("x"))]);
        fx.add("Person", "b", vec![("notindexed", s("y"))]);
        let rs = run(
            &fx.view(),
            "MATCH (n:Person) WHERE n.notindexed = 'x' RETURN n",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(rs.len(), 1);
        assert_eq!(rs.get(0, "n"), Some(&s("a")));
    }

    // --- IndexIntersect executor equivalence tests (T2) ---

    /// Both fields indexed: IndexIntersect fires and returns only the node matching both.
    #[test]
    fn index_intersect_both_indexed_fires() {
        use core_storage::property_index::PropertyIndex;
        use std::sync::atomic::Ordering;

        let mut fx = Fx::new();
        // alice: city=austin, age=30 — both match
        let a = fx.add(
            "Person",
            "alice",
            vec![("city", s("austin")), ("age", Value::Int(30))],
        );
        // bob: city=austin, age=25 — city matches, age misses
        let b = fx.add(
            "Person",
            "bob",
            vec![("city", s("austin")), ("age", Value::Int(25))],
        );
        // carol: city=boston, age=30 — age matches, city misses
        let c = fx.add(
            "Person",
            "carol",
            vec![("city", s("boston")), ("age", Value::Int(30))],
        );

        let mut pi = PropertyIndex::new();
        pi.enable("Person", "city");
        pi.enable("Person", "age");
        pi.set("Person", "city", a, &s("austin"));
        pi.set("Person", "city", b, &s("austin"));
        pi.set("Person", "city", c, &s("boston"));
        pi.set("Person", "age", a, &Value::Int(30));
        pi.set("Person", "age", b, &Value::Int(25));
        pi.set("Person", "age", c, &Value::Int(30));

        let q = "MATCH (n:Person) WHERE n.city = 'austin' AND n.age = 30 RETURN n";

        let before = super::INDEX_INTERSECT_FIRES.load(Ordering::Relaxed);
        let indexed = run(&fx.view_indexed(&pi), q, &BTreeMap::new()).unwrap();
        let after = super::INDEX_INTERSECT_FIRES.load(Ordering::Relaxed);
        assert!(
            after > before,
            "IndexIntersect must advance counter on indexed path"
        );
        assert_eq!(indexed.len(), 1);
        assert_eq!(indexed.get(0, "n"), Some(&s("alice")));

        // Fallback (no index) returns the same row.
        let fallback = run(&fx.view(), q, &BTreeMap::new()).unwrap();
        assert_eq!(
            rows_of(&fallback),
            rows_of(&indexed),
            "fallback must return identical rows"
        );
    }

    /// One indexed field + one unindexed: indexed path fires; unindexed field is post-filtered.
    #[test]
    fn index_intersect_one_indexed_one_not() {
        use core_storage::property_index::PropertyIndex;
        use std::sync::atomic::Ordering;

        let mut fx = Fx::new();
        let a = fx.add(
            "Person",
            "alice",
            vec![("city", s("austin")), ("role", s("eng"))],
        );
        let b = fx.add(
            "Person",
            "bob",
            vec![("city", s("austin")), ("role", s("mgr"))],
        );
        let _c = fx.add(
            "Person",
            "carol",
            vec![("city", s("boston")), ("role", s("eng"))],
        );

        // Only city is indexed; role is not.
        let mut pi = PropertyIndex::new();
        pi.enable("Person", "city");
        pi.set("Person", "city", a, &s("austin"));
        pi.set("Person", "city", b, &s("austin"));

        let q = "MATCH (n:Person) WHERE n.city = 'austin' AND n.role = 'eng' RETURN n";

        let before = super::INDEX_INTERSECT_FIRES.load(Ordering::Relaxed);
        let indexed = run(&fx.view_indexed(&pi), q, &BTreeMap::new()).unwrap();
        let after = super::INDEX_INTERSECT_FIRES.load(Ordering::Relaxed);
        assert!(
            after > before,
            "IndexIntersect must fire when at least one field is indexed"
        );
        assert_eq!(indexed.len(), 1);
        assert_eq!(indexed.get(0, "n"), Some(&s("alice")));

        let fallback = run(&fx.view(), q, &BTreeMap::new()).unwrap();
        assert_eq!(
            rows_of(&fallback),
            rows_of(&indexed),
            "fallback must return identical rows"
        );
    }

    /// No indexed fields: full scan fallback returns the correct node.
    /// Counter isolation is not asserted here because parallel tests share the
    /// global INDEX_INTERSECT_FIRES atomic; the fires-on-indexed path is already
    /// covered by `index_intersect_both_indexed_fires`.
    #[test]
    fn index_intersect_no_indexed_fallback() {
        let mut fx = Fx::new();
        fx.add("Person", "alice", vec![("x", s("1")), ("y", s("a"))]);
        fx.add("Person", "bob", vec![("x", s("1")), ("y", s("b"))]);
        fx.add("Person", "carol", vec![("x", s("2")), ("y", s("a"))]);

        let q = "MATCH (n:Person) WHERE n.x = '1' AND n.y = 'a' RETURN n";
        let rs = run(&fx.view(), q, &BTreeMap::new()).unwrap();
        assert_eq!(rs.len(), 1);
        assert_eq!(rs.get(0, "n"), Some(&s("alice")));
    }

    /// Two-field intersect with no nodes matching both returns empty.
    #[test]
    fn index_intersect_empty_intersection() {
        use core_storage::property_index::PropertyIndex;

        let mut fx = Fx::new();
        let a = fx.add(
            "Person",
            "alice",
            vec![("city", s("austin")), ("age", Value::Int(30))],
        );
        let b = fx.add(
            "Person",
            "bob",
            vec![("city", s("boston")), ("age", Value::Int(25))],
        );

        let mut pi = PropertyIndex::new();
        pi.enable("Person", "city");
        pi.enable("Person", "age");
        pi.set("Person", "city", a, &s("austin"));
        pi.set("Person", "city", b, &s("boston"));
        pi.set("Person", "age", a, &Value::Int(30));
        pi.set("Person", "age", b, &Value::Int(25));

        // Requesting city=boston AND age=30 — no node has both.
        let q = "MATCH (n:Person) WHERE n.city = 'boston' AND n.age = 30 RETURN n";
        let rs = run(&fx.view_indexed(&pi), q, &BTreeMap::new()).unwrap();
        assert_eq!(rs.len(), 0);
    }

    /// IndexIntersect equality using $params resolves at runtime and fires correctly.
    #[test]
    fn index_intersect_with_params() {
        use core_storage::property_index::PropertyIndex;
        use std::sync::atomic::Ordering;

        let mut fx = Fx::new();
        let a = fx.add(
            "Person",
            "alice",
            vec![("city", s("austin")), ("age", Value::Int(30))],
        );
        let b = fx.add(
            "Person",
            "bob",
            vec![("city", s("boston")), ("age", Value::Int(30))],
        );

        let mut pi = PropertyIndex::new();
        pi.enable("Person", "city");
        pi.enable("Person", "age");
        pi.set("Person", "city", a, &s("austin"));
        pi.set("Person", "city", b, &s("boston"));
        pi.set("Person", "age", a, &Value::Int(30));
        pi.set("Person", "age", b, &Value::Int(30));

        let q = "MATCH (n:Person) WHERE n.city = $c AND n.age = $a RETURN n";
        let mut params = BTreeMap::new();
        params.insert("c".to_string(), s("austin"));
        params.insert("a".to_string(), Value::Int(30));

        let before = super::INDEX_INTERSECT_FIRES.load(Ordering::Relaxed);
        let indexed = run(&fx.view_indexed(&pi), q, &params).unwrap();
        let after = super::INDEX_INTERSECT_FIRES.load(Ordering::Relaxed);
        assert!(after > before, "$param intersect must fire indexed path");
        assert_eq!(indexed.len(), 1);
        assert_eq!(indexed.get(0, "n"), Some(&s("alice")));

        let fallback = run(&fx.view(), q, &params).unwrap();
        assert_eq!(
            rows_of(&fallback),
            rows_of(&indexed),
            "fallback must return identical rows"
        );
    }

    /// Three-field intersect with two indexed fields returns the one node matching all three.
    #[test]
    fn index_intersect_three_fields() {
        use core_storage::property_index::PropertyIndex;

        let mut fx = Fx::new();
        // alice: all three match
        let a = fx.add(
            "Person",
            "alice",
            vec![
                ("city", s("austin")),
                ("age", Value::Int(30)),
                ("role", s("eng")),
            ],
        );
        // bob: city+age match, role misses
        let b = fx.add(
            "Person",
            "bob",
            vec![
                ("city", s("austin")),
                ("age", Value::Int(30)),
                ("role", s("mgr")),
            ],
        );

        let mut pi = PropertyIndex::new();
        pi.enable("Person", "city");
        pi.enable("Person", "age");
        // role intentionally not indexed
        pi.set("Person", "city", a, &s("austin"));
        pi.set("Person", "city", b, &s("austin"));
        pi.set("Person", "age", a, &Value::Int(30));
        pi.set("Person", "age", b, &Value::Int(30));

        let q =
            "MATCH (n:Person) WHERE n.city = 'austin' AND n.age = 30 AND n.role = 'eng' RETURN n";
        let indexed = run(&fx.view_indexed(&pi), q, &BTreeMap::new()).unwrap();
        assert_eq!(indexed.len(), 1);
        assert_eq!(indexed.get(0, "n"), Some(&s("alice")));

        let fallback = run(&fx.view(), q, &BTreeMap::new()).unwrap();
        assert_eq!(
            rows_of(&fallback),
            rows_of(&indexed),
            "fallback must return identical rows"
        );
    }
    // ── Task 17: the Cypher an assistant assumes exists ──────────────────────
    //
    // Every query below came out of a benchmark transcript where the agent
    // wrote it, got null or a parse error, and spent its turn budget
    // recovering. The fixtures are the association store's shape: talents
    // joined to companies by three rule-derived edge types.

    /// Talents and companies joined by three edge types.
    ///
    /// `c1` is reached by all three from both published talents; `c2` by only
    /// two from `t1`; `c3` by all three but from a talent below the
    /// experience floor. An intersection query must return `c1` alone.
    fn assoc_graph() -> Fx {
        let mut fx = Fx::new();
        let t1 = fx.add(
            "Talent",
            "t1",
            vec![
                ("status", s("published")),
                ("years_of_experience", i(12)),
                (
                    "specialties",
                    Value::List(vec![s("hospitality"), s("retail")]),
                ),
                ("location", Value::List(vec![f(40.71), f(-74.01)])),
            ],
        );
        let t2 = fx.add(
            "Talent",
            "t2",
            vec![
                ("status", s("published")),
                ("years_of_experience", i(11)),
                ("location", Value::List(vec![f(41.88), f(-87.63)])),
            ],
        );
        let t3 = fx.add(
            "Talent",
            "t3",
            vec![("status", s("published")), ("years_of_experience", i(3))],
        );
        let c1 = fx.add("Company", "c1", vec![("name", s("Acme Design Works"))]);
        let c2 = fx.add("Company", "c2", vec![("name", s("Beta Studio"))]);
        let c3 = fx.add("Company", "c3", vec![("name", s("Gamma Works"))]);
        for (t, c) in [(t1, c1), (t2, c1), (t3, c3)] {
            fx.edge("INDUSTRY_ALIGNMENT", t, c, vec![]);
            fx.edge("SPECIALTY_MATCH", t, c, vec![]);
            fx.edge("LOCATION_FIT", t, c, vec![]);
        }
        // c2 is joined by two of the three types only.
        fx.edge("INDUSTRY_ALIGNMENT", t1, c2, vec![]);
        fx.edge("SPECIALTY_MATCH", t1, c2, vec![]);
        fx
    }

    /// `n.key` is the node's key after a plain MATCH — it used to be null,
    /// and `RETURN c.key` is what every agent writes.
    #[test]
    fn node_key_reads_as_a_property() {
        let fx = assoc_graph();
        let rs = run(
            &fx.view(),
            "MATCH (n:Company) RETURN n.key",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(
            rows_of(&rs),
            vec![
                vec![Some(s("c1"))],
                vec![Some(s("c2"))],
                vec![Some(s("c3"))]
            ]
        );
    }

    /// The same after a WITH aggregation with a HAVING filter — the shape the
    /// benchmark's multihop cells died on, where the grouping key used to be
    /// flattened to a scalar and both `c.key` and `key(c)` stopped working.
    #[test]
    fn node_key_survives_a_with_aggregation() {
        let fx = assoc_graph();
        let rs = run(
            &fx.view(),
            "MATCH (t:Talent)-[:INDUSTRY_ALIGNMENT]->(c:Company) \
             WITH c, count(*) AS n WHERE n >= 1 RETURN c.key, key(c), n",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(
            rows_of(&rs),
            vec![
                vec![Some(s("c1")), Some(s("c1")), Some(i(2))],
                vec![Some(s("c2")), Some(s("c2")), Some(i(1))],
                vec![Some(s("c3")), Some(s("c3")), Some(i(1))],
            ]
        );
    }

    /// A stored property named `key` wins over the node's identity.
    #[test]
    fn stored_key_property_wins_over_node_key() {
        let mut fx = Fx::new();
        fx.add("N", "a", vec![("key", s("stored"))]);
        let rs = run(&fx.view(), "MATCH (n:N) RETURN n.key", &BTreeMap::new()).unwrap();
        assert_eq!(rows_of(&rs), vec![vec![Some(s("stored"))]]);
    }

    /// `n.id` / `id(n)` fall back to the id-map key when no stored `id` exists.
    #[test]
    fn n_id_falls_back_to_key_when_unstored() {
        let mut fx = Fx::new();
        fx.add("Person", "alice", vec![]);
        let rs = run(
            &fx.view(),
            "MATCH (n:Person) WHERE n.id = 'alice' RETURN n.id, id(n), n.key",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(
            rows_of(&rs),
            vec![vec![Some(s("alice")), Some(s("alice")), Some(s("alice"))]]
        );
    }

    /// A stored `id` property wins over the node's key, including in WHERE.
    ///
    /// Every assertion here also held before `n.id` resolved at all — an
    /// unstored `n.id` was null, so `WHERE n.id = 'k'` was empty for the wrong
    /// reason. What gives the test its teeth is the second node: `fallback`
    /// stores no `id`, so it is reachable only once `n.id` falls back to the
    /// key, and it must *not* be dragged in by the stored-wins hit.
    #[test]
    fn stored_id_property_wins_over_key() {
        let mut fx = Fx::new();
        fx.add("N", "k", vec![("id", s("other"))]);
        fx.add("N", "fallback", vec![]);
        let v = fx.view();
        let by_key = run(
            &v,
            "MATCH (n:N) WHERE n.id = 'fallback' RETURN key(n) AS k",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(
            col(&by_key, "k"),
            vec![Some(s("fallback"))],
            "a node with no stored id is found by its key, and the stored-id \
             node is not swept in with it"
        );
        let miss = run(
            &v,
            "MATCH (n:N) WHERE n.id = 'k' RETURN n.id",
            &BTreeMap::new(),
        )
        .unwrap();
        assert!(
            miss.is_empty(),
            "stored id must win: WHERE n.id = key is empty"
        );
        let hit = run(
            &v,
            "MATCH (n:N) WHERE n.id = 'other' RETURN n.id",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(rows_of(&hit), vec![vec![Some(s("other"))]]);
        let projected = run(&v, "MATCH (n:N) RETURN n.id AS i", &BTreeMap::new()).unwrap();
        let mut got: Vec<String> = col(&projected, "i")
            .into_iter()
            .map(|v| match v {
                Some(Value::Str(s)) => s,
                other => panic!("expected a string, got {other:?}"),
            })
            .collect();
        got.sort();
        assert_eq!(
            got,
            vec!["fallback".to_string(), "other".to_string()],
            "projection is stored-wins per node: the stored id for one, the \
             key fallback for the other"
        );
    }

    /// `id()` on a non-node is a named error of the same class as `key()`.
    #[test]
    fn id_function_rejects_non_node() {
        let (fx, _, _) = single_edge();
        let v = fx.view();
        let key_err = run(&v, "MATCH (a)-[r:T]->(b) RETURN key(r)", &BTreeMap::new())
            .expect_err("key() on a relationship must error");
        let id_err = run(&v, "MATCH (a)-[r:T]->(b) RETURN id(r)", &BTreeMap::new())
            .expect_err("id() on a relationship must error");
        assert!(
            key_err.contains("not a node"),
            "key() error class: {key_err}"
        );
        assert!(id_err.contains("not a node"), "id() error class: {id_err}");
        assert!(
            id_err.contains("id()"),
            "id() error must name itself: {id_err}"
        );
    }

    /// `WHERE n.id = lit` on an unstored id is a ScanKey, not a label scan.
    #[test]
    fn where_n_id_eq_literal_uses_scan_key() {
        let mut fx = Fx::new();
        fx.add("Person", "alice", vec![]);
        let fires_before = super::SCAN_KEY_FIRES.load(std::sync::atomic::Ordering::Relaxed);
        let rs = run(
            &fx.view(),
            "MATCH (n:Person) WHERE n.id = 'alice' RETURN n",
            &BTreeMap::new(),
        )
        .unwrap();
        let fires_after = super::SCAN_KEY_FIRES.load(std::sync::atomic::Ordering::Relaxed);
        assert!(
            fires_after > fires_before,
            "SCAN_KEY_FIRES must increment; before={fires_before} after={fires_after}"
        );
        assert_eq!(rows_of(&rs), vec![vec![Some(s("alice"))]]);
    }

    /// `WHERE n.key = $k` uses ScanKey.
    #[test]
    fn where_n_key_eq_param_uses_scan_key() {
        let mut fx = Fx::new();
        fx.add("Person", "alice", vec![]);
        let mut params = BTreeMap::new();
        params.insert("k".to_string(), s("alice"));
        let fires_before = super::SCAN_KEY_FIRES.load(std::sync::atomic::Ordering::Relaxed);
        let rs = run(
            &fx.view(),
            "MATCH (n:Person) WHERE n.key = $k RETURN n",
            &params,
        )
        .unwrap();
        let fires_after = super::SCAN_KEY_FIRES.load(std::sync::atomic::Ordering::Relaxed);
        assert!(
            fires_after > fires_before,
            "SCAN_KEY_FIRES must increment; before={fires_before} after={fires_after}"
        );
        assert_eq!(rows_of(&rs), vec![vec![Some(s("alice"))]]);
    }

    /// `key(n)` is always the id-map key, never a stored property of that name.
    /// A node carrying `key = 'K'` as a property must NOT match `key(n) = 'K'`
    /// when its own key is something else — the identity-eq fold must not
    /// apply the `n.key` stored-wins rule to the function form.
    #[test]
    fn where_key_func_ignores_a_stored_key_property() {
        let mut fx = Fx::new();
        fx.add("N", "a", vec![("key", s("K"))]);
        fx.add("N", "K", vec![]);
        let rs = run(
            &fx.view(),
            "MATCH (n:N) WHERE key(n) = 'K' RETURN key(n) AS k",
            &BTreeMap::new(),
        )
        .unwrap();
        let got: Vec<Option<Value>> = col(&rs, "k");
        assert_eq!(
            got,
            vec![Some(s("K"))],
            "key(n) is the id-map key: only the node keyed K matches, not the \
             node whose stored `key` property is K"
        );
    }

    /// `id(n)` is always the id-map key, so a stored `id` property must not
    /// hide a node from `WHERE id(n) = <its own key>`.
    #[test]
    fn where_id_func_ignores_a_stored_id_property() {
        let mut fx = Fx::new();
        fx.add("N", "k", vec![("id", s("other"))]);
        let rs = run(
            &fx.view(),
            "MATCH (n:N) WHERE id(n) = 'k' RETURN id(n) AS k",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(
            col(&rs, "k"),
            vec![Some(s("k"))],
            "id(n) is the id-map key: a stored `id` property must not suppress \
             the match"
        );
    }

    /// The property form is stored-wins, and the fold answers with the union:
    /// the node whose *stored* `id` is the value, plus the node whose *key* is
    /// the value and which stores no `id`. Spec §5.2.
    #[test]
    fn where_n_id_respects_stored_wins() {
        let mut fx = Fx::new();
        fx.add("N", "k", vec![("id", s("other"))]);
        fx.add("N", "other", vec![]);
        let rs = run(
            &fx.view(),
            "MATCH (n:N) WHERE n.id = 'other' RETURN key(n) AS k",
            &BTreeMap::new(),
        )
        .unwrap();
        let mut got: Vec<String> = col(&rs, "k")
            .into_iter()
            .map(|v| match v {
                Some(Value::Str(s)) => s,
                other => panic!("expected a string key, got {other:?}"),
            })
            .collect();
        got.sort();
        assert_eq!(
            got,
            vec!["k".to_string(), "other".to_string()],
            "stored-wins union: the stored-id node and the key-fallback node"
        );
    }

    /// `labels(n)` returns the node's label list; `n.label` is the scalar
    /// spelling. `labels()` used to be an unknown function.
    #[test]
    fn labels_and_label_read_the_node_label() {
        let fx = assoc_graph();
        let rs = run(
            &fx.view(),
            "MATCH (n:Company) RETURN labels(n), n.label LIMIT 1",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(
            rows_of(&rs),
            vec![vec![
                Some(Value::List(vec![s("Company")])),
                Some(s("Company"))
            ]]
        );
    }

    /// An unknown function names itself in the error rather than reading null.
    #[test]
    fn unknown_function_is_a_named_error() {
        let fx = assoc_graph();
        let err = run(
            &fx.view(),
            "MATCH (n:Company) RETURN nodes(n)",
            &BTreeMap::new(),
        )
        .expect_err("unknown function must error");
        assert!(err.contains("unknown function `nodes`"), "{err}");
        assert!(
            err.contains("labels"),
            "error must list what is supported: {err}"
        );
    }

    /// Infix `STARTS WITH` / `ENDS WITH` / `CONTAINS` — all three were parse
    /// errors, and each cost a round trip in the benchmark.
    #[test]
    fn infix_string_predicates_filter() {
        let fx = assoc_graph();
        let p = BTreeMap::new();
        for (q, want) in [
            (
                "MATCH (c:Company) WHERE c.name STARTS WITH 'Acme' RETURN c.key",
                vec!["c1"],
            ),
            (
                "MATCH (c:Company) WHERE c.name ENDS WITH 'Works' RETURN c.key",
                vec!["c1", "c3"],
            ),
            (
                "MATCH (c:Company) WHERE c.name CONTAINS 'Studio' RETURN c.key",
                vec!["c2"],
            ),
            (
                "MATCH (c:Company) WHERE NOT c.name CONTAINS 'Works' RETURN c.key",
                vec!["c2"],
            ),
        ] {
            let rs = run(&fx.view(), q, &p).unwrap_or_else(|e| panic!("{q}: {e}"));
            let got: Vec<String> = (0..rs.len())
                .map(|r| match rs.row(r)[0].clone() {
                    Some(Value::Str(k)) => k,
                    other => panic!("{q}: {other:?}"),
                })
                .collect();
            assert_eq!(got, want, "{q}");
        }
    }

    /// A missing or non-string property makes an infix predicate false, not
    /// an error — same null handling as the `startsWith(a, b)` spelling.
    #[test]
    fn infix_string_predicate_on_missing_property_is_false() {
        let fx = assoc_graph();
        let rs = run(
            &fx.view(),
            "MATCH (t:Talent) WHERE t.name STARTS WITH 'x' RETURN t.key",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(rows_of(&rs), Vec::<Vec<Option<Value>>>::new());
    }

    /// `STARTS` without `WITH` is a named parse error, not a silent
    /// reinterpretation of the word as a variable.
    #[test]
    fn starts_without_with_is_a_named_error() {
        let err = parse(&lex("MATCH (c:Company) WHERE c.name STARTS 'Acme' RETURN c").unwrap())
            .expect_err("must not parse");
        assert!(err.contains("expected WITH after STARTS"), "{err}");
    }

    /// `n.location[0]` — bracket indexing was a parse error, which forced the
    /// benchmark's geo cells to page the whole table and filter by hand.
    #[test]
    fn list_subscript_reads_one_element() {
        let fx = assoc_graph();
        let rs = run(
            &fx.view(),
            "MATCH (t:Talent) WHERE t.key = 't1' \
             RETURN t.location[0] AS lat, t.location[1] AS lon, \
                    t.location[-1] AS last, t.location[7] AS oob, t.status[0] AS notalist",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(
            rows_of(&rs),
            vec![vec![
                Some(f(40.71)),
                Some(f(-74.01)),
                Some(f(-74.01)),
                None,
                None
            ]]
        );
    }

    /// `n.key` filters, orders and matches — not only projects. Each of
    /// these goes down a different path in the executor (index-scan fallback,
    /// bounded pull scan, raw-row ORDER BY, inline pattern property).
    #[test]
    fn node_key_works_in_every_position() {
        let fx = assoc_graph();
        let p = BTreeMap::new();
        for (q, want) in [
            (
                "MATCH (c:Company) WHERE c.key = 'c2' RETURN c.key",
                vec!["c2"],
            ),
            (
                "MATCH (c:Company) WHERE c.key <> 'c1' RETURN c.key LIMIT 1",
                vec!["c2"],
            ),
            (
                "MATCH (c:Company) WITH c ORDER BY c.key DESC RETURN c.key LIMIT 1",
                vec!["c3"],
            ),
            ("MATCH (c:Company {key: 'c3'}) RETURN c.key", vec!["c3"]),
            (
                "MATCH (c:Company) WHERE c.label = 'Company' RETURN c.key LIMIT 1",
                vec!["c1"],
            ),
        ] {
            let rs = run(&fx.view(), q, &p).unwrap_or_else(|e| panic!("{q}: {e}"));
            let got: Vec<String> = (0..rs.len())
                .map(|r| match rs.row(r)[0].clone() {
                    Some(Value::Str(k)) => k,
                    other => panic!("{q}: {other:?}"),
                })
                .collect();
            assert_eq!(got, want, "{q}");
        }
    }

    /// A subscript is usable in WHERE, not only in RETURN.
    #[test]
    fn list_subscript_filters() {
        let fx = assoc_graph();
        let rs = run(
            &fx.view(),
            "MATCH (t:Talent) WHERE t.location[0] > 41.0 RETURN t.key",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(rows_of(&rs), vec![vec![Some(s("t2"))]]);
    }

    /// `count(DISTINCT t)` counts each talent once however many rows it
    /// produced. Without it, an alternation or a multi-pattern match
    /// multiplies the count by the number of matching edge types.
    #[test]
    fn count_distinct_counts_each_binding_once() {
        let fx = assoc_graph();
        let q = "MATCH (t:Talent)-[:INDUSTRY_ALIGNMENT|:SPECIALTY_MATCH]->(c:Company) \
                 WITH c, count(t) AS raw, count(DISTINCT t) AS uniq WHERE raw >= 1 \
                 RETURN c.key, raw, uniq";
        let rs = run(&fx.view(), q, &BTreeMap::new()).unwrap();
        assert_eq!(
            rows_of(&rs),
            vec![
                vec![Some(s("c1")), Some(i(4)), Some(i(2))],
                vec![Some(s("c2")), Some(i(2)), Some(i(1))],
                vec![Some(s("c3")), Some(i(2)), Some(i(1))],
            ]
        );
    }

    /// `count(DISTINCT r)` on a **relationship** variable counts edges, not
    /// nothing.
    ///
    /// A relationship cell has no key, and the DISTINCT gate used to read
    /// that as "no value", which drops the row: the answer was `0` on a graph
    /// full of edges, and silently — no error, no null, a number that looks
    /// like an answer. An edge is identified by its type and its two
    /// endpoints, so on a graph where no pair is joined twice by one type
    /// `count(DISTINCT r)` is `count(r)`.
    #[test]
    fn count_distinct_on_a_relationship_counts_edges() {
        let fx = assoc_graph();
        let rs = run(
            &fx.view(),
            "MATCH (a)-[r]->(b) RETURN count(r) AS raw, count(DISTINCT r) AS uniq",
            &BTreeMap::new(),
        )
        .unwrap();
        let row = &rows_of(&rs)[0];
        assert_eq!(
            row[0], row[1],
            "no pair is joined twice by one type: {row:?}"
        );
        assert_ne!(row[1], Some(i(0)), "a graph full of edges counts them");
    }

    /// And the same alternation that multiplies rows for a node variable
    /// still yields one key per edge, so `DISTINCT` over `r` is the edge
    /// count of the two types rather than the row count.
    #[test]
    fn count_distinct_on_a_relationship_survives_an_alternation() {
        let fx = assoc_graph();
        let rs = run(
            &fx.view(),
            "MATCH (t:Talent)-[r:INDUSTRY_ALIGNMENT|:SPECIALTY_MATCH]->(c:Company) \
             RETURN count(r) AS raw, count(DISTINCT r) AS uniq",
            &BTreeMap::new(),
        )
        .unwrap();
        let row = &rows_of(&rs)[0];
        assert_eq!(row[0], row[1], "every row bound a different edge: {row:?}");
    }

    /// `collect(DISTINCT …)` dedupes the same way.
    #[test]
    fn collect_distinct_dedupes() {
        let fx = assoc_graph();
        let rs = run(
            &fx.view(),
            "MATCH (t:Talent)-[:INDUSTRY_ALIGNMENT|:SPECIALTY_MATCH]->(c:Company) \
             WHERE c.key = 'c1' WITH collect(DISTINCT t.status) AS st RETURN st",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(
            rows_of(&rs),
            vec![vec![Some(Value::List(vec![s("published")]))]]
        );
    }

    /// `DISTINCT *` is rejected at parse time rather than silently ignored.
    #[test]
    fn count_distinct_star_is_rejected() {
        let err =
            parse(&lex("MATCH (n) RETURN count(DISTINCT *)").unwrap()).expect_err("must not parse");
        assert!(
            err.contains("DISTINCT * is not a valid aggregate argument"),
            "{err}"
        );
    }

    /// A variable actually named `distinct` still parses as an argument.
    #[test]
    fn distinct_is_still_usable_as_a_variable_name() {
        let q = parse(&lex("MATCH (distinct) RETURN count(distinct)").unwrap()).unwrap();
        assert_eq!(
            q.returns[0].value,
            RetVal::Agg {
                func: crate::cypher::ast::AggFunc::Count,
                arg: crate::cypher::ast::AggArg::Var("distinct".into()),
            }
        );
    }

    /// Comma-separated patterns in one MATCH bind the same variables across
    /// every pattern — the intersection shape every multihop question needs.
    /// A company joined by only two of the three types must not survive.
    #[test]
    fn comma_patterns_intersect_on_shared_variables() {
        let fx = assoc_graph();
        let q = "MATCH (t:Talent)-[:INDUSTRY_ALIGNMENT]->(c:Company), \
                       (t)-[:SPECIALTY_MATCH]->(c), \
                       (t)-[:LOCATION_FIT]->(c) \
                 WHERE t.status = 'published' AND t.years_of_experience >= 10 \
                 WITH c, count(DISTINCT t) AS n WHERE n >= 2 \
                 RETURN c.key, n ORDER BY n DESC";
        let rs = run(&fx.view(), q, &BTreeMap::new()).unwrap();
        assert_eq!(rows_of(&rs), vec![vec![Some(s("c1")), Some(i(2))]]);
    }

    /// The same query with the threshold dropped lists every company joined
    /// by all three types, and only those: `c2` (two types) is absent, `c3`
    /// (three types, under-experienced talent) is filtered by the WHERE.
    #[test]
    fn comma_patterns_exclude_partial_matches() {
        let fx = assoc_graph();
        let q = "MATCH (t:Talent)-[:INDUSTRY_ALIGNMENT]->(c:Company), \
                       (t)-[:SPECIALTY_MATCH]->(c), \
                       (t)-[:LOCATION_FIT]->(c) \
                 WHERE t.years_of_experience >= 10 \
                 WITH c, count(DISTINCT t) AS n RETURN c.key, n";
        let rs = run(&fx.view(), q, &BTreeMap::new()).unwrap();
        assert_eq!(rows_of(&rs), vec![vec![Some(s("c1")), Some(i(2))]]);
    }

    /// Comma-separated patterns and a run of separate MATCH clauses parse to
    /// the same query.
    #[test]
    fn comma_patterns_equal_separate_match_clauses() {
        let commas =
            parse(&lex("MATCH (a:A)-[:X]->(b:B), (a)-[:Y]->(b) RETURN a.key").unwrap()).unwrap();
        let clauses =
            parse(&lex("MATCH (a:A)-[:X]->(b:B) MATCH (a)-[:Y]->(b) RETURN a.key").unwrap())
                .unwrap();
        assert_eq!(commas.matches, clauses.matches);
    }

    // ── Task 17 fix round 1 ──────────────────────────────────────────────

    /// An aggregate `WITH` with no `WHERE` after it still projects what the
    /// RETURN asked for.
    ///
    /// The streaming group-aggregate path emits the group table itself and
    /// ignores every op after the aggregate but ORDER BY / SKIP / LIMIT, so a
    /// `Project` was silently dropped: this query came back as columns `c`
    /// and `n` carrying the group key and the raw count, with `n * 2` never
    /// evaluated. Adding a `WHERE` after the `WITH` routed it elsewhere and
    /// was correct, which is what made it look like a query-shape problem.
    #[test]
    fn aggregate_with_projects_without_a_having_clause() {
        let fx = assoc_graph();
        let rs = run(
            &fx.view(),
            "MATCH (t:Talent)-[:INDUSTRY_ALIGNMENT]->(c:Company) \
             WITH c, count(t) AS n RETURN c.name, n * 2 AS dbl",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(rs.columns(), ["c.name", "dbl"]);
        assert_eq!(
            rows_of(&rs),
            vec![
                vec![Some(s("Acme Design Works")), Some(i(4))],
                vec![Some(s("Beta Studio")), Some(i(2))],
                vec![Some(s("Gamma Works")), Some(i(2))],
            ]
        );
    }

    /// The same, projecting `key(c)` — the columns are named for what was
    /// asked, not for the grouping variable.
    #[test]
    fn aggregate_with_names_its_projected_columns() {
        let fx = assoc_graph();
        let rs = run(
            &fx.view(),
            "MATCH (t:Talent)-[:INDUSTRY_ALIGNMENT]->(c:Company) \
             WITH c, count(t) AS n RETURN key(c), n",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(rs.columns(), ["key(c)", "n"]);
        assert_eq!(
            rows_of(&rs),
            vec![
                vec![Some(s("c1")), Some(i(2))],
                vec![Some(s("c2")), Some(i(1))],
                vec![Some(s("c3")), Some(i(1))],
            ]
        );
    }

    /// `RETURN c` after an aggregate `WITH` still returns the key string, and
    /// only that column.
    #[test]
    fn aggregate_with_returning_the_bare_node_keeps_the_key() {
        let fx = assoc_graph();
        let rs = run(
            &fx.view(),
            "MATCH (t:Talent)-[:INDUSTRY_ALIGNMENT]->(c:Company) \
             WITH c, count(t) AS n RETURN c",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(rs.columns(), ["c"]);
        assert_eq!(
            rows_of(&rs),
            vec![
                vec![Some(s("c1"))],
                vec![Some(s("c2"))],
                vec![Some(s("c3"))]
            ]
        );
    }

    /// A plain aggregate with no `WITH` keeps the streaming path and its
    /// column names — the fix must not move it.
    #[test]
    fn a_plain_aggregate_is_unchanged() {
        let fx = assoc_graph();
        let rs = run(
            &fx.view(),
            "MATCH (c:Company) RETURN c.name, count(*)",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(rs.columns(), ["c.name", "COUNT(*)"]);
        assert_eq!(rs.len(), 3);
    }

    /// Two unaliased subscripts of the same list are two columns, each named
    /// for what it reads. They used to collide on `<expr>` and fail the
    /// planner's duplicate-column check before the query ever ran.
    #[test]
    fn two_subscripts_of_one_list_are_two_named_columns() {
        let fx = assoc_graph();
        let rs = run(
            &fx.view(),
            "MATCH (t:Talent) RETURN t.location[0], t.location[1]",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(rs.columns(), ["t.location[0]", "t.location[1]"]);
        assert_eq!(
            rows_of(&rs),
            vec![
                vec![Some(f(40.71)), Some(f(-74.01))],
                vec![Some(f(41.88)), Some(f(-87.63))],
                // t3 has no location at all.
                vec![None, None],
            ]
        );
    }

    /// A subscript carries through a `WITH` stage — the case that breaks if
    /// the name the executor projects and the one `collect_vars` interns ever
    /// disagree, both aliased and left to name itself.
    #[test]
    fn a_subscript_carries_through_a_with_stage() {
        let fx = assoc_graph();
        let rs = run(
            &fx.view(),
            "MATCH (t:Talent) WHERE t.location[0] > 41.0 \
             WITH t, t.location[0] AS lat, t.location[1] \
             RETURN key(t), lat, t.location[1]",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(rs.columns(), ["key(t)", "lat", "t.location[1]"]);
        assert_eq!(
            rows_of(&rs),
            vec![vec![Some(s("t2")), Some(f(41.88)), Some(f(-87.63))]]
        );
    }

    /// A subscript inside a function call is named the same way, and a
    /// non-literal index falls back to the operand's own label.
    #[test]
    fn subscript_column_names_cover_nested_and_computed_forms() {
        use crate::cypher::ast::operand_label;
        let prop = || Operand::Prop {
            var: "t".into(),
            field: "location".into(),
        };
        let at = |idx: Operand| Operand::Index {
            base: Box::new(prop()),
            index: Box::new(idx),
        };
        assert_eq!(operand_label(&at(Operand::Lit(i(0)))), "t.location[0]");
        assert_eq!(operand_label(&at(Operand::Lit(i(-1)))), "t.location[-1]");
        assert_eq!(
            operand_label(&at(Operand::Param("k".into()))),
            "t.location[$k]"
        );
        assert_eq!(
            operand_label(&at(Operand::Var("j".into()))),
            "t.location[j]"
        );
        assert_eq!(
            operand_label(&Operand::Index {
                base: Box::new(at(Operand::Lit(i(0)))),
                index: Box::new(Operand::Lit(i(1))),
            }),
            "t.location[0][1]"
        );
    }

    // ── Task 17 fix round 2 ─────────────────────────────────────────────
    //
    // `WITH … WHERE …` filters *after* the projection, so an alias the WITH
    // introduces is in scope for it. The executor always did that; the
    // planner's pre-flight check scoped the WHERE to the pre-WITH variables
    // and rejected every such query before it ran.

    /// An alias a non-aggregate `WITH` introduces is visible to its `WHERE`.
    #[test]
    fn a_with_alias_is_visible_to_its_where() {
        let fx = assoc_graph();
        let rs = run(
            &fx.view(),
            "MATCH (t:Talent) WITH t, t.years_of_experience AS x WHERE x > 11 \
             RETURN key(t), x",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(rs.columns(), ["key(t)", "x"]);
        assert_eq!(rows_of(&rs), vec![vec![Some(s("t1")), Some(i(12))]]);
    }

    /// …and to its `ORDER BY`, in the same clause.
    #[test]
    fn a_with_alias_is_visible_to_where_and_order_by_together() {
        let fx = assoc_graph();
        let rs = run(
            &fx.view(),
            "MATCH (t:Talent) WITH t, t.years_of_experience AS x WHERE x > 3 \
             ORDER BY x DESC RETURN key(t), x",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(rs.columns(), ["key(t)", "x"]);
        assert_eq!(
            rows_of(&rs),
            vec![
                vec![Some(s("t1")), Some(i(12))],
                vec![Some(s("t2")), Some(i(11))],
            ]
        );
    }

    /// A `WITH` that carries only the variable, with no alias at all, filters
    /// on the node's own properties. This always worked; it is pinned so the
    /// scoping change cannot quietly take it away.
    #[test]
    fn a_with_carrying_only_a_variable_still_filters_on_it() {
        let fx = assoc_graph();
        let rs = run(
            &fx.view(),
            "MATCH (t:Talent) WHERE t.status = 'published' \
             WITH t WHERE t.years_of_experience > 11 RETURN key(t)",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(rs.columns(), ["key(t)"]);
        assert_eq!(rows_of(&rs), vec![vec![Some(s("t1"))]]);
    }

    /// A `WITH` alias with no `WHERE` projects under its new name. Also
    /// always worked; pinned alongside the two above.
    #[test]
    fn a_with_alias_projects_under_its_new_name() {
        let fx = assoc_graph();
        let rs = run(
            &fx.view(),
            "MATCH (c:Company) WITH c, c.name AS nm RETURN nm",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(rs.columns(), ["nm"]);
        assert_eq!(
            rows_of(&rs),
            vec![
                vec![Some(s("Acme Design Works"))],
                vec![Some(s("Beta Studio"))],
                vec![Some(s("Gamma Works"))],
            ]
        );
    }

    /// Every alias shape a WITH can introduce is in scope for its WHERE: a
    /// property, an arithmetic expression, a subscript, a renamed node, and
    /// an alias that drops the node it came from.
    #[test]
    fn every_with_alias_shape_is_in_scope_for_the_where() {
        let fx = assoc_graph();
        let p = BTreeMap::new();
        for (q, want) in [
            (
                "MATCH (t:Talent) WITH t, t.years_of_experience + 1 AS x WHERE x > 12 \
                 RETURN key(t)",
                vec!["t1"],
            ),
            (
                "MATCH (t:Talent) WITH t, t.location[0] AS lat WHERE lat > 41.0 RETURN key(t)",
                vec!["t2"],
            ),
            (
                "MATCH (t:Talent) WITH t AS u WHERE u.years_of_experience > 11 RETURN key(u)",
                vec!["t1"],
            ),
            (
                "MATCH (t:Talent) WITH t.key AS k WHERE k = 't2' RETURN k",
                vec!["t2"],
            ),
        ] {
            let rs = run(&fx.view(), q, &p).unwrap_or_else(|e| panic!("{q}: {e}"));
            let got: Vec<String> = (0..rs.len())
                .map(|r| match rs.row(r)[0].clone() {
                    Some(Value::Str(k)) => k,
                    other => panic!("{q}: {other:?}"),
                })
                .collect();
            assert_eq!(got, want, "{q}");
        }
    }

    /// A name that is neither in scope before the WITH nor introduced by it
    /// is still a named error — the scope widened, it did not disappear.
    #[test]
    fn an_unknown_name_in_a_with_where_is_still_an_error() {
        let err = plan(
            &parse(
                &lex(
                    "MATCH (t:Talent) WITH t, t.years_of_experience AS x WHERE nope > 1 \
                     RETURN key(t)",
                )
                .unwrap(),
            )
            .unwrap(),
        )
        .expect_err("must not plan");
        assert_eq!(err, "unbound variable `nope` in WHERE");
    }

    /// Comma-separated patterns with no shared variable are a cartesian
    /// product, the openCypher meaning.
    #[test]
    fn comma_patterns_without_shared_vars_are_a_product() {
        let mut fx = Fx::new();
        fx.add("A", "a1", vec![]);
        fx.add("A", "a2", vec![]);
        fx.add("B", "b1", vec![]);
        let rs = run(
            &fx.view(),
            "MATCH (a:A), (b:B) RETURN a.key, b.key",
            &BTreeMap::new(),
        )
        .unwrap();
        assert_eq!(
            rows_of(&rs),
            vec![
                vec![Some(s("a1")), Some(s("b1"))],
                vec![Some(s("a2")), Some(s("b1"))],
            ]
        );
    }
}