code_kb_core/

queries.rs

use rusqlite::{Connection, OptionalExtension, Row, ToSql, params};
use rust_stemmers::{Algorithm, Stemmer};
use std::collections::{HashMap, HashSet};
use thiserror::Error;

use crate::db::local_variable_predicate;
use crate::models::{
    BlastRadiusResult, FileFact, ImpactedSymbol, LiteralFact, ReferenceSite, SearchExplain,
    StructuralFact, Symbol, SymbolSearchResult, TestTarget, TypeFact, has_generated_lambda_name,
    is_generated_lambda_name,
};

#[derive(Debug, Error)]
pub enum QueryError {
    #[error("Database query error: {0}")]
    Sqlite(#[from] rusqlite::Error),
    #[error("Symbol '{name}' not found in {workspace}. {hint}")]
    SymbolNotFound {
        name: String,
        workspace: String,
        hint: String,
    },
    #[error(
        "Ambiguous symbol '{0}': found {1} matching candidates. Choose a candidate by its kind and source line, then pass its id as `symbol_id` (CLI: `--symbol-id`) to get_symbol_body, get_symbol_context, find_references, or blast_radius:\n{2}"
    )]
    AmbiguousSymbol(String, usize, String),
    #[error("Invalid direction '{0}': must be 'callers' or 'callees'")]
    InvalidDirection(String),
    #[error("Result limit must be between 0 and {MAX_RESULT_LIMIT}, got {0}")]
    InvalidResultLimit(usize),
}

pub const MAX_RESULT_LIMIT: usize = 200;

pub fn validate_result_limit(limit: usize) -> Result<(), QueryError> {
    if limit > MAX_RESULT_LIMIT {
        return Err(QueryError::InvalidResultLimit(limit));
    }
    Ok(())
}

fn map_symbol(row: &Row) -> rusqlite::Result<Symbol> {
    Ok(Symbol {
        symbol_id: row.get("symbol_id")?,
        file_id: row.get("file_id")?,
        path: row.get::<_, String>("path")?.replace('\\', "/"),
        language: row.get("language")?,
        name: row.get("name")?,
        kind: row.get("kind")?,
        signature: row.get("signature")?,
        doc_comment: row.get("doc_comment")?,
        visibility: row.get("visibility")?,
        parent_symbol_id: row.get("parent_symbol_id")?,
        start_line: row.get::<_, i64>("start_line")? as usize,
        start_column: row.get::<_, i64>("start_column")? as usize,
        end_line: row.get::<_, i64>("end_line")? as usize,
        end_column: row.get::<_, i64>("end_column")? as usize,
        start_byte: row.get::<_, i64>("start_byte")? as usize,
        end_byte: row.get::<_, i64>("end_byte")? as usize,
        body_start_line: row
            .get::<_, Option<i64>>("body_start_line")?
            .map(|v| v as usize),
        body_start_column: row
            .get::<_, Option<i64>>("body_start_column")?
            .map(|v| v as usize),
        body_end_line: row
            .get::<_, Option<i64>>("body_end_line")?
            .map(|v| v as usize),
        body_end_column: row
            .get::<_, Option<i64>>("body_end_column")?
            .map(|v| v as usize),
        body_start_byte: row
            .get::<_, Option<i64>>("body_start_byte")?
            .map(|v| v as usize),
        body_end_byte: row
            .get::<_, Option<i64>>("body_end_byte")?
            .map(|v| v as usize),
        body_hash: row.get("body_hash")?,
        semantic_group: row.get("semantic_group")?,
        is_test: row.get::<_, i64>("is_test")? != 0,
        test_container: row.get::<_, i64>("test_container")? != 0,
    })
}

pub(crate) fn escape_like(value: &str) -> String {
    value
        .replace('\\', "\\\\")
        .replace('%', "\\%")
        .replace('_', "\\_")
}

/// Retrieve indexed files optionally scoped by path filter, pushed down to SQLite.
pub fn load_scoped_files(
    conn: &Connection,
    path_filter: Option<&str>,
) -> Result<Vec<FileFact>, QueryError> {
    let norm = path_filter
        .map(|p| p.replace('\\', "/").trim_matches('/').to_string())
        .filter(|p| !p.is_empty());
    let norm_bs = norm.as_ref().map(|p| p.replace('/', "\\"));
    let prefix = norm.as_ref().map(|path| format!("{}/%", escape_like(path)));
    let prefix_bs = norm_bs
        .as_ref()
        .map(|path| format!("{}\\\\%", escape_like(path)));

    let sql = "SELECT file_id, path, language, content_hash, content_bytes, line_count, indexed_at
               FROM files
               WHERE (:path IS NULL
                  OR path = :path COLLATE NOCASE
                  OR path = :path_bs COLLATE NOCASE
                  OR path LIKE :path_prefix ESCAPE '\\'
                  OR path LIKE :path_prefix_bs ESCAPE '\\')
               ORDER BY (:path IS NOT NULL AND (path = :path OR path = :path_bs)) DESC, path ASC";

    let mut stmt = conn.prepare(sql)?;
    let files = stmt
        .query_map(
            rusqlite::named_params! {
                ":path": norm.as_deref(),
                ":path_bs": norm_bs.as_deref(),
                ":path_prefix": prefix.as_deref(),
                ":path_prefix_bs": prefix_bs.as_deref(),
            },
            |row| {
                Ok(FileFact {
                    file_id: row.get(0)?,
                    path: row.get::<_, String>(1)?.replace('\\', "/"),
                    language: row.get(2)?,
                    content_hash: row.get(3)?,
                    content_bytes: row.get(4)?,
                    line_count: row.get(5)?,
                    indexed_at: row.get(6)?,
                })
            },
        )?
        .collect::<Result<Vec<_>, _>>()?;

    Ok(files)
}

/// The most files one outline lists by name.
pub const OUTLINE_FILE_CAP: usize = 1000;

/// The path filter and depth of one outline, as SQL parameters.
pub struct OutlineScope {
    path: Option<String>,
    path_bs: Option<String>,
    prefix: Option<String>,
    prefix_bs: Option<String>,
    /// The most path separators a file the outline lists by name can have.
    max_slashes: i64,
}

impl OutlineScope {
    pub fn new(path_filter: Option<&str>, depth: usize) -> Self {
        let path = path_filter
            .map(|p| p.replace('\\', "/").trim_matches('/').to_string())
            .filter(|p| !p.is_empty());
        let path_bs = path.as_ref().map(|p| p.replace('/', "\\"));
        let max_slashes = match &path {
            None => depth.saturating_sub(1),
            Some(p) => p.matches('/').count() + depth,
        } as i64;
        Self {
            prefix: path.as_ref().map(|p| format!("{}/%", escape_like(p))),
            prefix_bs: path_bs.as_ref().map(|p| format!("{}\\\\%", escape_like(p))),
            path,
            path_bs,
            max_slashes,
        }
    }

    /// True when the outline lists `path` by name instead of counting it in its folder.
    pub fn lists(&self, path: &str) -> bool {
        path.matches(['/', '\\']).count() as i64 <= self.max_slashes
    }

    fn params(&self) -> [(&'static str, &dyn ToSql); 5] {
        [
            (":path", &self.path),
            (":path_bs", &self.path_bs),
            (":path_prefix", &self.prefix),
            (":path_prefix_bs", &self.prefix_bs),
            (":max_slashes", &self.max_slashes),
        ]
    }
}

/// The `WHERE` terms for the files in an outline scope that have an extractor.
fn outline_file_terms(conn: &Connection) -> String {
    let supported = if has_column(conn, "files", "status") {
        "COALESCE(status, '') != 'unsupported'"
    } else {
        "1 = 1"
    };
    format!(
        "(:path IS NULL
            OR path = :path COLLATE NOCASE
            OR path = :path_bs COLLATE NOCASE
            OR path LIKE :path_prefix ESCAPE '\\'
            OR path LIKE :path_prefix_bs ESCAPE '\\')
         AND {supported}"
    )
}

/// The files an outline lists by name, in path order.
fn outline_files_cte(conn: &Connection) -> String {
    format!(
        "outline_files AS (
            SELECT path FROM files
            WHERE {}
              AND length(path) - length(replace(replace(path, '/', ''), '\\', '')) <= :max_slashes
            ORDER BY path ASC
            LIMIT {OUTLINE_FILE_CAP}
        )",
        outline_file_terms(conn)
    )
}

/// Streams every file path in the outline scope that has an extractor, in path order.
pub fn for_each_outline_path(
    conn: &Connection,
    scope: &OutlineScope,
    mut visit: impl FnMut(String) -> bool,
) -> Result<(), QueryError> {
    let sql = format!(
        "SELECT path FROM files WHERE {} ORDER BY path ASC",
        outline_file_terms(conn)
    );
    let mut stmt = conn.prepare(&sql)?;
    let params = scope.params();
    let mut rows = stmt.query(&params[..4])?;
    while let Some(row) = rows.next()? {
        if !visit(row.get(0)?) {
            break;
        }
    }
    Ok(())
}

/// True when `alias` names a definition at the top of its file: no parent, or only a namespace,
/// module, or package around it.
fn outline_level(alias: &str) -> String {
    format!(
        "({alias}.parent_symbol_id IS NULL OR EXISTS (
            SELECT 1 FROM symbols ns WHERE ns.symbol_id = {alias}.parent_symbol_id
              AND ns.kind IN ('namespace', 'module', 'package')))"
    )
}

const OUTLINE_KINDS: &str =
    "('function', 'method', 'struct', 'enum', 'trait', 'class', 'interface', 'type')";

/// Load up to `limit_per_file` top-level definitions for each file the outline lists by name.
/// Files deeper than `depth` are filtered out in SQLite to keep memory strictly bounded. The
/// unary `+` keeps the overload check on the name index: the parent index would walk every
/// top-level symbol of the repository for each row.
pub fn load_scoped_outline_symbols(
    conn: &Connection,
    path_filter: Option<&str>,
    depth: usize,
    limit_per_file: usize,
) -> Result<HashMap<String, Vec<Symbol>>, QueryError> {
    let scope = OutlineScope::new(path_filter, depth);
    let sql = format!(
        "WITH {files},
        ranked AS (
            SELECT s.symbol_id, s.file_id, s.path, s.language, s.name, s.kind, s.signature, s.doc_comment,
                   s.visibility, s.parent_symbol_id, s.start_line, s.start_column, s.end_line, s.end_column,
                   s.start_byte, s.end_byte, s.body_start_line, s.body_start_column, s.body_end_line,
                   s.body_end_column, s.body_start_byte, s.body_end_byte, s.body_hash, s.semantic_group,
                   s.is_test, s.test_container,
                   ROW_NUMBER() OVER (
                       PARTITION BY s.path ORDER BY (s.is_test = 1 OR s.test_container = 1), s.start_line ASC
                   ) as rn
            FROM symbols s
            JOIN outline_files bf ON s.path = bf.path
            WHERE s.kind IN {OUTLINE_KINDS}
              AND {top}
              AND NOT EXISTS (SELECT 1 FROM symbols overload
                              WHERE +overload.path = s.path AND overload.name = s.name
                                AND overload.kind = s.kind
                                AND +overload.parent_symbol_id IS s.parent_symbol_id
                                AND overload.start_line < s.start_line)
        )
        SELECT symbol_id, file_id, path, language, name, kind, signature, doc_comment,
               visibility, parent_symbol_id, start_line, start_column, end_line, end_column,
               start_byte, end_byte, body_start_line, body_start_column, body_end_line,
               body_end_column, body_start_byte, body_end_byte, body_hash, semantic_group,
               is_test, test_container
        FROM ranked
        WHERE rn <= :limit
        ORDER BY path ASC, start_line ASC",
        files = outline_files_cte(conn),
        top = outline_level("s"),
    );

    let mut stmt = conn.prepare(&sql)?;
    let limit = limit_per_file as i64;
    let mut params = scope.params().to_vec();
    params.push((":limit", &limit));
    let mut rows = stmt.query(&params[..])?;

    let mut symbols_by_file: HashMap<String, Vec<Symbol>> = HashMap::new();
    while let Some(row) = rows.next()? {
        let sym = map_symbol(row)?;
        symbols_by_file
            .entry(sym.path.clone())
            .or_default()
            .push(sym);
    }

    Ok(symbols_by_file)
}

/// How many names one outline row stands for.
#[derive(Debug, Clone, Copy, Default, PartialEq)]
pub struct OutlineCounts {
    /// Distinct top-level definitions that are not tests; overloads count once.
    pub definitions: usize,
    /// Tests at any level, without setup and fixtures.
    pub tests: usize,
    /// Setup, teardown, and fixture members.
    pub fixtures: usize,
}

/// For each file the outline lists by name: its top-level definitions that are not tests, its
/// tests, and its fixtures, so the outline can say how many names it left out.
pub fn load_outline_counts(
    conn: &Connection,
    path_filter: Option<&str>,
    depth: usize,
) -> Result<HashMap<String, OutlineCounts>, QueryError> {
    let scope = OutlineScope::new(path_filter, depth);
    let lifecycle = if has_column(conn, "symbols", "test_lifecycle") {
        "COALESCE(s.test_lifecycle, 0) != 0"
    } else {
        "0"
    };
    let mut stmt = conn.prepare(&format!(
        "WITH {files}
         SELECT replace(s.path, '\\', '/'),
                COUNT(DISTINCT CASE WHEN {top} AND s.is_test = 0 AND s.test_container = 0
                    AND s.kind IN {OUTLINE_KINDS}
                    THEN s.kind || ' ' || s.name END),
                SUM(s.is_test = 1 AND NOT {lifecycle}),
                SUM(s.is_test = 1 AND {lifecycle})
         FROM symbols s
         JOIN outline_files bf ON s.path = bf.path
         GROUP BY replace(s.path, '\\', '/')",
        files = outline_files_cte(conn),
        top = outline_level("s"),
    ))?;
    let rows = stmt.query_map(&scope.params()[..], |row| {
        Ok((
            row.get::<_, String>(0)?,
            OutlineCounts {
                definitions: row.get::<_, i64>(1)? as usize,
                tests: row.get::<_, i64>(2)? as usize,
                fixtures: row.get::<_, i64>(3)? as usize,
            },
        ))
    })?;
    Ok(rows.collect::<Result<_, _>>()?)
}

/// The largest number of files one answer is measured against.
pub const BASELINE_PATH_CAP: usize = 20;

/// Sums the indexed byte size of up to `BASELINE_PATH_CAP` distinct paths.
/// Returns 0 when no path is indexed, so telemetry never fails a tool call.
pub fn file_sizes_for_paths(conn: &Connection, paths: &[String]) -> usize {
    let mut seen = std::collections::HashSet::new();
    let distinct: Vec<&String> = paths
        .iter()
        .filter(|path| seen.insert(path.as_str()))
        .take(BASELINE_PATH_CAP)
        .collect();
    if distinct.is_empty() {
        return 0;
    }

    let placeholders = vec!["?"; distinct.len()].join(", ");
    let sql =
        format!("SELECT COALESCE(SUM(content_bytes), 0) FROM files WHERE path IN ({placeholders})");
    conn.query_row(&sql, rusqlite::params_from_iter(distinct), |row| {
        row.get::<_, i64>(0)
    })
    .map(|bytes| bytes.max(0) as usize)
    .unwrap_or(0)
}

/// Lookup single file metadata by path with slash-boundary matching.
pub fn get_file(conn: &Connection, path: &str) -> Result<Option<FileFact>, QueryError> {
    let normalized = path.replace('\\', "/");
    let backslash = path.replace('/', "\\");

    // Check exact path match first, prioritizing exact case before case-insensitive fallback
    let mut stmt = conn.prepare(
        "SELECT file_id, path, language, content_hash, content_bytes, line_count, indexed_at
         FROM files
         WHERE (path = ?1 COLLATE NOCASE OR path = ?2 COLLATE NOCASE)
         ORDER BY (path = ?1 OR path = ?2) DESC
         LIMIT 1",
    )?;

    let mut rows = stmt.query(params![normalized, backslash])?;
    if let Some(row) = rows.next()? {
        Ok(Some(FileFact {
            file_id: row.get(0)?,
            path: row.get::<_, String>(1)?.replace('\\', "/"),
            language: row.get(2)?,
            content_hash: row.get(3)?,
            content_bytes: row.get(4)?,
            line_count: row.get(5)?,
            indexed_at: row.get(6)?,
        }))
    } else {
        Ok(None)
    }
}

/// Count parse diagnostics recorded for a file, returning 0 when the index has none.
/// Number of indexed symbols in one file, without loading the rows. Like `load_file_symbols`,
/// an exact-case path wins and the case-insensitive match is only the fallback.
pub fn count_file_symbols(conn: &Connection, path: &str) -> usize {
    let count = |sql: &str| {
        conn.query_row(
            sql,
            params![path.replace('\\', "/"), path.replace('/', "\\")],
            |row| row.get::<_, i64>(0),
        )
        .map(|count| count.max(0) as usize)
        .unwrap_or(0)
    };
    match count("SELECT COUNT(*) FROM symbols WHERE path = ?1 OR path = ?2") {
        0 => count(
            "SELECT COUNT(*) FROM symbols
             WHERE path = ?1 COLLATE NOCASE OR path = ?2 COLLATE NOCASE",
        ),
        exact => exact,
    }
}

pub fn count_parse_diagnostics(conn: &Connection, path: &str) -> usize {
    conn.query_row(
        "SELECT COUNT(*) FROM parse_diagnostics
         WHERE path = ?1 COLLATE NOCASE OR path = ?2 COLLATE NOCASE",
        params![path.replace('\\', "/"), path.replace('/', "\\")],
        |row| row.get::<_, i64>(0),
    )
    .map(|count| count as usize)
    .unwrap_or(0)
}

/// Streams the paths of files in the outline scope that no extractor reads.
pub fn for_each_unsupported_path(
    conn: &Connection,
    scope: &OutlineScope,
    mut visit: impl FnMut(&str),
) -> Result<(), QueryError> {
    if !has_column(conn, "files", "status") {
        return Ok(());
    }
    let mut stmt = conn.prepare(
        "SELECT path FROM files
         WHERE status = 'unsupported'
           AND (:path IS NULL
             OR path = :path COLLATE NOCASE
             OR path = :path_bs COLLATE NOCASE
             OR path LIKE :path_prefix ESCAPE '\\'
             OR path LIKE :path_prefix_bs ESCAPE '\\')",
    )?;
    let params = scope.params();
    let mut rows = stmt.query(&params[..4])?;
    while let Some(row) = rows.next()? {
        visit(&row.get::<_, String>(0)?);
    }
    Ok(())
}

/// Load all symbols declared inside a specific file.
pub fn load_file_symbols(conn: &Connection, file_path: &str) -> Result<Vec<Symbol>, QueryError> {
    // Normalizing slashes for path matching
    let normalized = file_path.replace('\\', "/");
    let backslash = file_path.replace('/', "\\");

    // Try exact case matching first to avoid conflating sibling files on case-sensitive filesystems
    let mut stmt = conn.prepare(
        "SELECT symbol_id, file_id, path, language, name, kind, signature, doc_comment,
                visibility, parent_symbol_id, start_line, start_column, end_line, end_column,
                start_byte, end_byte, body_start_line, body_start_column, body_end_line,
                body_end_column, body_start_byte, body_end_byte, body_hash, semantic_group,
                is_test, test_container
         FROM symbols
         WHERE (path = ?1 OR path = ?2)
         ORDER BY start_line ASC, start_column ASC",
    )?;

    let rows = stmt
        .query_map(params![&normalized, &backslash], map_symbol)?
        .collect::<Result<Vec<_>, _>>()?;

    if !rows.is_empty() {
        return Ok(rows);
    }

    // Fall back to case-insensitive match (for Windows or case-variant requests)
    let mut stmt = conn.prepare(
        "SELECT symbol_id, file_id, path, language, name, kind, signature, doc_comment,
                visibility, parent_symbol_id, start_line, start_column, end_line, end_column,
                start_byte, end_byte, body_start_line, body_start_column, body_end_line,
                body_end_column, body_start_byte, body_end_byte, body_hash, semantic_group,
                is_test, test_container
         FROM symbols
         WHERE (path = ?1 COLLATE NOCASE OR path = ?2 COLLATE NOCASE)
         ORDER BY start_line ASC, start_column ASC",
    )?;

    let rows = stmt
        .query_map(params![normalized, backslash], map_symbol)?
        .collect::<Result<Vec<_>, _>>()?;

    Ok(rows)
}

/// Normalizes common symbol kind aliases to their canonical database representation.
pub fn normalize_kind(kind: &str) -> String {
    let lower = kind.trim().to_lowercase();
    match lower.as_str() {
        "fn" | "func" | "function" => "function".to_string(),
        "method" => "method".to_string(),
        "struct" => "struct".to_string(),
        "class" => "class".to_string(),
        "enum" => "enum".to_string(),
        "trait" => "trait".to_string(),
        "interface" => "interface".to_string(),
        "type" | "typedef" => "type".to_string(),
        "mod" | "module" => "module".to_string(),
        "const" | "constant" => "constant".to_string(),
        "var" | "variable" => "variable".to_string(),
        _ => lower,
    }
}

/// Search symbols by name query, kind filter, and test flag.
pub fn search_symbols(
    conn: &Connection,
    query: &str,
    kind_filter: Option<&str>,
    include_tests: bool,
    limit: usize,
) -> Result<Vec<Symbol>, QueryError> {
    search_symbols_scoped(conn, query, kind_filter, None, include_tests, limit)
}

/// Search symbols with optional path scoping filter. Locals and parameters are left out unless
/// the caller passes `kind = "variable"` or names one explicitly as a qualified name such as
/// `open_conn::conn`. With `kind = "variable"` they match by name only, because they are not in
/// the full-text index.
/// The lookup statement. The exact-name bypass of the test filter is written as `+name` so
/// SQLite cannot plan it as a multi-index OR, which scans every non-test row through the
/// test-path predicate (about 8 ms on this repository's index against 1 ms).
/// True when `alias` is an attribute row such as `self.debug = …` whose name a base class already
/// defines with a value or a property. A base class named by a pending `extends` row counts only
/// when no class of that name is as close to the subclass's file, so an unclear base keeps the row. The row is a write to an inherited attribute, not a
/// definition, so lookup, search, and skeletons leave it out. `CASE` keeps SQLite from walking
/// the base classes of every row a search touches: it evaluates an `AND` with `EXISTS` eagerly.
pub(crate) fn inherited_attribute_write(alias: &str) -> String {
    let rival_rank = path_proximity("rival.path", "pe.path");
    let base_rank = path_proximity("c.path", "pe.path");
    format!(
        "(CASE WHEN {alias}.kind = 'property' AND {alias}.parent_symbol_id IS NOT NULL
          AND ({alias}.signature LIKE 'self.%' OR {alias}.signature LIKE 'this.%' OR {alias}.signature LIKE 'cls.%')
          THEN EXISTS (
            WITH RECURSIVE base(symbol_id, depth) AS (
                SELECT {alias}.parent_symbol_id, 0
                UNION
                SELECT r.to_symbol_id, base.depth + 1
                FROM relationships r JOIN base ON r.from_symbol_id = base.symbol_id
                WHERE +r.kind = 'extends' AND base.depth < 8
                UNION
                SELECT c.symbol_id, base.depth + 1
                FROM pending_relationships pe
                JOIN base ON pe.from_symbol_id = base.symbol_id
                JOIN symbols c ON c.name = pe.target_terminal_name
                WHERE +pe.kind = 'extends' AND base.depth < 8
                  AND c.kind IN ('class', 'interface', 'struct', 'trait', 'protocol')
                  AND NOT EXISTS (
                      SELECT 1 FROM symbols rival
                      WHERE rival.name = c.name AND rival.symbol_id != c.symbol_id
                        AND rival.kind IN ('class', 'interface', 'struct', 'trait', 'protocol')
                        AND {rival_rank} >= {base_rank})
            )
            SELECT 1 FROM base JOIN symbols inherited ON inherited.parent_symbol_id = base.symbol_id
            WHERE base.depth > 0 AND inherited.name = {alias}.name
              AND (inherited.kind NOT IN ('variable', 'field') OR inherited.signature LIKE '%=%'))
          ELSE 0 END)"
    )
}

/// ` AND NOT` [`inherited_attribute_write`] for the row `s`, or nothing for an index without
/// relationship tables.
fn inherited_write_exclusion(conn: &Connection) -> String {
    if has_table(conn, "relationships") && has_table(conn, "pending_relationships") {
        format!(" AND NOT {}", inherited_attribute_write("s"))
    } else {
        String::new()
    }
}

/// The ids among `ids` that [`inherited_attribute_write`] leaves out. Callers check only the rows
/// they return, because the check walks base classes and a common word touches thousands of rows.
pub fn inherited_writes_among<'a>(
    conn: &Connection,
    ids: impl Iterator<Item = &'a str>,
) -> Result<HashSet<String>, QueryError> {
    if inherited_write_exclusion(conn).is_empty() {
        return Ok(HashSet::new());
    }
    let ids = serde_json::to_string(&ids.collect::<Vec<_>>()).unwrap_or_default();
    let mut stmt = conn.prepare(&format!(
        "SELECT s.symbol_id FROM symbols s
         WHERE s.symbol_id IN (SELECT value FROM json_each(?1)) AND {}",
        inherited_attribute_write("s")
    ))?;
    let found = stmt
        .query_map(params![ids], |row| row.get(0))?
        .collect::<Result<_, _>>()?;
    Ok(found)
}

fn search_symbols_sql(variables_wanted: bool, include_tests: bool, limit: usize) -> String {
    let mut sql = format!(
        "SELECT symbol_id, file_id, path, language, name, kind, signature, doc_comment,
                visibility, parent_symbol_id, start_line, start_column, end_line, end_column,
                start_byte, end_byte, body_start_line, body_start_column, body_end_line,
                body_end_column, body_start_byte, body_end_byte, body_hash, semantic_group,
                is_test, test_container
         FROM symbols s
         WHERE (name = :query OR name LIKE :pattern ESCAPE '\\')
           AND {KIND_FILTER}
           AND (:path IS NULL OR replace(path, '\\', '/') = :path COLLATE NOCASE OR replace(path, '\\', '/') LIKE :path_like || '/%' ESCAPE '\\' OR replace(path, '\\', '/') LIKE '%/' || :path_like ESCAPE '\\')",
    );
    sql.push_str(&document_link_exclusion());

    if !variables_wanted {
        sql.push_str(&format!(" AND NOT {}", local_variable_predicate("s")));
    }

    if !include_tests {
        sql.push_str(&format!(
            " AND (+name = :query OR (is_test = 0 AND test_container = 0 AND NOT {}))",
            test_path_predicate("s")
        ));
    }

    sql.push_str(
        " ORDER BY (name = :query) DESC, (name LIKE :prefix ESCAPE '\\') DESC, (kind IN ('function', 'struct', 'class', 'trait', 'method', 'enum', 'interface', 'type')) DESC, (COALESCE(signature, '') LIKE 'self.%') ASC, length(name) ASC, path ASC LIMIT ",
    );
    sql.push_str(&limit.to_string());
    sql
}

pub fn search_symbols_scoped(
    conn: &Connection,
    query: &str,
    kind_filter: Option<&str>,
    path_filter: Option<&str>,
    include_tests: bool,
    limit: usize,
) -> Result<Vec<Symbol>, QueryError> {
    validate_result_limit(limit)?;
    if limit == 0 {
        return Ok(Vec::new());
    }
    let norm_kind = kind_filter.map(normalize_kind);
    if let Some(owner) = query
        .strip_suffix("::")
        .or_else(|| query.strip_suffix('.'))
        .and_then(|owner| owner.rsplit(['.', ':']).next())
        .filter(|owner| !owner.is_empty())
    {
        return owner_members(
            conn,
            owner,
            norm_kind.as_deref(),
            path_filter,
            include_tests,
            limit,
        );
    }
    if (query.contains("::") || query.contains('.'))
        && let Some(sym) = get_symbol_by_name(conn, query, path_filter)?
    {
        let kind_matches = norm_kind
            .as_deref()
            .is_none_or(|kind| sym.kind == kind || crate::formatters::display_kind(&sym) == kind);
        return Ok(if kind_matches { vec![sym] } else { Vec::new() });
    }

    let pattern = format!("%{}%", escape_like(query));
    let prefix = format!("{}%", escape_like(query));
    let normalized_path = path_filter.map(|p| {
        p.replace('\\', "/")
            .trim_start_matches("./")
            .trim_matches('/')
            .to_string()
    });
    let escaped_path = normalized_path.as_deref().map(escape_like);

    let path_val = normalized_path.as_deref();
    let path_like = escaped_path.as_deref();
    let kind_val = norm_kind.as_deref();
    let mut fetch = 4 * limit;
    loop {
        let sql = search_symbols_sql(
            norm_kind.as_deref() == Some("variable"),
            include_tests,
            fetch,
        );
        let rows = conn
            .prepare(&sql)?
            .query_map(
                rusqlite::named_params! {
                    ":query": query,
                    ":pattern": pattern,
                    ":prefix": prefix,
                    ":kind": kind_val,
                    ":path": path_val,
                    ":path_like": path_like,
                },
                map_symbol,
            )?
            .collect::<Result<Vec<Symbol>, _>>()?;
        let fetched = rows.len();
        let inherited = inherited_writes_among(conn, rows.iter().map(|s| s.symbol_id.as_str()))?;
        let mut kept: Vec<Symbol> = rows
            .into_iter()
            .filter(|s| !inherited.contains(&s.symbol_id))
            .collect();
        let has_definition = kept.iter().any(|s| s.kind != "import");
        let mut counted = 0;
        let past_limit = kept.iter().position(|s| {
            if !folds_into_import_line(query, s, has_definition) {
                counted += 1;
            }
            counted > limit
        });
        if let Some(cut) = past_limit {
            kept.truncate(cut);
            return Ok(kept);
        }
        if fetched < fetch || fetch >= 4 * MAX_RESULT_LIMIT {
            return Ok(kept);
        }
        fetch *= 2;
    }
}

const MEMBER_LIST_OWNER_KINDS: &str = "'class', 'struct', 'interface', 'trait', 'enum', 'record', 'object', \
                           'protocol', 'union', 'module', 'namespace'";

/// SQL predicate for the `:kind` filter over `symbols s`: the stored kind or the kind that
/// `formatters::display_kind` shows, so a kind copied from the output always filters.
const KIND_FILTER: &str = "(:kind IS NULL OR s.kind = :kind OR :kind = CASE
      WHEN s.language = 'html' AND s.kind = 'class' THEN 'element'
      WHEN s.language = 'sql' AND s.kind = 'class' THEN 'table'
      WHEN s.language = 'markdown' AND s.kind = 'module' THEN 'section'
      WHEN s.language = 'markdown' AND s.kind = 'import' THEN 'link'
      WHEN s.kind = 'property' AND s.language <> 'fsharp'
           AND (s.signature GLOB 'self.*' OR s.signature GLOB 'this.*' OR s.signature GLOB 'cls.*')
        THEN 'attribute'
    END)";

/// The members of the classes, modules, and namespaces named `owner`, in source order: what
/// `lookup_symbol("App.")` lists.
fn owner_members(
    conn: &Connection,
    owner: &str,
    kind: Option<&str>,
    path_filter: Option<&str>,
    include_tests: bool,
    limit: usize,
) -> Result<Vec<Symbol>, QueryError> {
    let path = path_filter.map(|p| {
        p.replace('\\', "/")
            .trim_start_matches("./")
            .trim_matches('/')
            .to_string()
    });
    let path_like = path.as_deref().map(escape_like);
    let tests = if include_tests {
        String::new()
    } else {
        format!(
            " AND +s.is_test = 0 AND +s.test_container = 0 AND NOT {}",
            test_path_predicate("s")
        )
    };
    let docs = documentation_language_list();
    let sql = format!(
        "SELECT s.symbol_id, s.file_id, s.path, s.language, s.name, s.kind, s.signature, s.doc_comment,
                s.visibility, s.parent_symbol_id, s.start_line, s.start_column, s.end_line, s.end_column,
                s.start_byte, s.end_byte, s.body_start_line, s.body_start_column, s.body_end_line,
                s.body_end_column, s.body_start_byte, s.body_end_byte, s.body_hash, s.semantic_group,
                s.is_test, s.test_container
         FROM symbols o JOIN symbols s ON s.parent_symbol_id = o.symbol_id
         WHERE o.name = :owner
           AND o.kind IN ({MEMBER_LIST_OWNER_KINDS})
           AND (o.language NOT IN ({docs})
                OR NOT EXISTS (SELECT 1 FROM symbols c WHERE c.name = :owner
                                 AND c.kind IN ({MEMBER_LIST_OWNER_KINDS}) AND c.language NOT IN ({docs})))
           AND s.kind NOT IN ('parameter', 'import')
           AND {KIND_FILTER}
           AND (:path IS NULL OR replace(s.path, '\\', '/') = :path COLLATE NOCASE
                OR replace(s.path, '\\', '/') LIKE :path_like || '/%' ESCAPE '\\'
                OR replace(s.path, '\\', '/') LIKE '%/' || :path_like ESCAPE '\\'){tests}
         ORDER BY s.path, s.start_line"
    );
    let rows = conn
        .prepare(&sql)?
        .query_map(
            rusqlite::named_params! {
                ":owner": owner,
                ":kind": kind,
                ":path": path.as_deref(),
                ":path_like": path_like.as_deref(),
            },
            map_symbol,
        )?
        .collect::<Result<Vec<Symbol>, _>>()?;
    let inherited = inherited_writes_among(conn, rows.iter().map(|s| s.symbol_id.as_str()))?;
    Ok(rows
        .into_iter()
        .filter(|s| !inherited.contains(&s.symbol_id))
        .take(limit)
        .collect())
}

/// True when lookup shows `row` inside the one line that counts the imports of `query`, not as a
/// row of its own: an import of that exact name, when some row is a definition. Such rows do not
/// count toward the lookup limit.
pub fn folds_into_import_line(query: &str, row: &Symbol, has_definition: bool) -> bool {
    has_definition
        && row.kind == "import"
        && (row.name == query
            || row.name.ends_with(&format!(".{query}"))
            || row.name.ends_with(&format!("::{query}")))
}

/// Sanitizes a free-form user query into `(and_query, or_query)` formatted for SQLite FTS5.
/// Each alphanumeric/underscore token is quoted and given a prefix wildcard: `"token"*`.
/// Identifiers are split at case boundaries too (`parseHTTPResponse` -> `parse HTTP Response`),
/// so a camelCase query finds a snake_case symbol. Each split identifier keeps its unsplit form
/// as an alternative inside its own required term, so a camelCase symbol, which FTS5 indexes as
/// one token, still matches. A query of two or three words also tries their concatenation.
/// English stop words are dropped unless the whole query is stop words, and tokens under three
/// characters get no prefix wildcard.
pub fn sanitize_fts5_query(query: &str) -> (String, String) {
    let raw_words = query_words(query);
    let split: Vec<(Vec<&str>, Option<&str>)> = raw_words
        .iter()
        .map(|raw| {
            let parts = split_identifier(raw);
            let whole = (parts.len() > 1).then_some(*raw);
            (parts, whole)
        })
        .collect();
    let any_content = split
        .iter()
        .any(|(parts, _)| parts.iter().any(|p| !is_stop_word(p)));

    let mut and_groups: Vec<String> = Vec::new();
    let mut or_terms: Vec<String> = Vec::new();
    for (parts, whole) in split {
        let parts: Vec<String> = parts
            .into_iter()
            .filter(|p| !any_content || !is_stop_word(p))
            .map(fts5_term)
            .collect();
        let whole = whole.map(fts5_term);
        let group = match (parts.is_empty(), whole.as_deref()) {
            (true, None) => continue,
            (true, Some(w)) => w.to_string(),
            (false, None) => parts.join(" "),
            (false, Some(w)) => format!("(({}) OR {w})", parts.join(" ")),
        };
        and_groups.push(group);
        or_terms.extend(parts);
        or_terms.extend(whole);
    }

    if and_groups.is_empty() {
        return (String::new(), String::new());
    }

    let mut and_query = and_groups.join(" AND ");
    if (2..=3).contains(&raw_words.len()) {
        let all: String = raw_words.concat();
        if all.len() <= 64 {
            let all = fts5_term(&all);
            and_query = format!("({and_query}) OR {all}");
            or_terms.push(all);
        }
    }
    (and_query, or_terms.join(" OR "))
}

const STOP_WORDS: &[&str] = &[
    "a", "about", "after", "again", "all", "already", "also", "always", "an", "and", "another",
    "any", "are", "as", "at", "be", "because", "been", "before", "being", "between", "both", "but",
    "by", "can", "could", "did", "do", "does", "each", "either", "else", "ever", "every", "for",
    "from", "had", "has", "have", "here", "how", "if", "in", "instead", "is", "it", "its",
    "itself", "just", "many", "may", "might", "more", "most", "much", "must", "neither", "never",
    "no", "nor", "not", "of", "on", "once", "one", "only", "or", "other", "our", "per", "rather",
    "same", "should", "since", "so", "some", "still", "such", "than", "that", "the", "their",
    "them", "then", "there", "these", "they", "this", "those", "through", "to", "too", "two",
    "until", "very", "via", "was", "we", "were", "what", "when", "where", "whether", "which",
    "while", "who", "whom", "why", "will", "with", "within", "without", "would", "yet", "you",
    "your",
];

fn is_stop_word(word: &str) -> bool {
    STOP_WORDS.contains(&word.to_ascii_lowercase().as_str())
}

/// Splits a query at every character that is not alphanumeric or `_`.
fn query_words(query: &str) -> Vec<&str> {
    query
        .split(|c: char| !c.is_alphanumeric() && c != '_')
        .filter(|s| !s.is_empty())
        .collect()
}

/// Lowercase terms of three or more characters for the trigram name index: every query word
/// and every identifier part of it (`collapse_name` -> `collapse_name`, `collapse`, `name`),
/// deduplicated. Stop words are dropped unless every term is a stop word. Empty when no term
/// qualifies.
fn trigram_name_terms(query: &str) -> Vec<String> {
    let mut terms: Vec<String> = Vec::new();
    for word in query_words(query) {
        for term in std::iter::once(word).chain(split_identifier(word)) {
            let lower = term.to_lowercase();
            if lower.chars().count() >= 3 && !terms.contains(&lower) {
                terms.push(lower);
            }
        }
    }
    let any_content = terms.iter().any(|t| !is_stop_word(t));
    terms.retain(|t| !any_content || !is_stop_word(t));
    terms
}

/// Quotes one token for FTS5 with a prefix wildcard, except for tokens under three characters.
fn fts5_term(token: &str) -> String {
    if token.chars().count() < 3 {
        format!("\"{token}\"")
    } else {
        format!("\"{token}\"*")
    }
}

/// FTS5 query for a symbol name as typed: no case splitting, no stop words.
/// `isReady` -> `"isReady"*`, so related-test lookup stays as strict as the name.
fn name_prefix_query(name: &str) -> String {
    name.split(|c: char| !c.is_alphanumeric() && c != '_')
        .filter(|s| !s.is_empty())
        .map(|s| format!("\"{s}\"*"))
        .collect::<Vec<_>>()
        .join(" ")
}

/// Splits one identifier into words at `_`, digit runs, and case boundaries.
/// `parseHTTPResponse2` -> `["parse", "HTTP", "Response", "2"]`.
fn split_identifier(word: &str) -> Vec<&str> {
    let mut out = Vec::new();
    split_identifier_into(word, &mut out);
    out
}

/// `split_identifier` that appends to a caller-owned vector, so tokenizing a whole doc
/// comment costs one allocation instead of one per word.
fn split_identifier_into<'a>(word: &'a str, out: &mut Vec<&'a str>) {
    let mut chars = word.char_indices().peekable();
    let Some((_, first)) = chars.next() else {
        return;
    };
    let mut prev = char_class(first);
    let mut start = 0;
    while let Some((idx, c)) = chars.next() {
        let cur = char_class(c);
        let next = chars.peek().map_or(OTHER, |(_, n)| char_class(*n));
        if identifier_boundary(prev, cur, next) {
            push_piece(out, &word[start..idx]);
            start = idx;
        }
        prev = cur;
    }
    push_piece(out, &word[start..]);
}

const OTHER: u8 = 0;
const UNDERSCORE: u8 = 1;
const UPPER: u8 = 2;
const LOWER: u8 = 3;
const DIGIT: u8 = 4;

fn char_class(c: char) -> u8 {
    if c == '_' {
        UNDERSCORE
    } else if c.is_uppercase() {
        UPPER
    } else if c.is_lowercase() {
        LOWER
    } else if c.is_ascii_digit() {
        DIGIT
    } else {
        OTHER
    }
}

fn byte_class(b: u8) -> u8 {
    match b {
        b'_' => UNDERSCORE,
        b'A'..=b'Z' => UPPER,
        b'a'..=b'z' => LOWER,
        b'0'..=b'9' => DIGIT,
        _ => OTHER,
    }
}

/// The identifier split rule over character classes: `_` on either side, lower/digit to
/// upper, the last upper of an acronym before a lower (`HTTPResponse`), and digit runs.
fn identifier_boundary(prev: u8, cur: u8, next: u8) -> bool {
    cur == UNDERSCORE
        || prev == UNDERSCORE
        || (cur == UPPER && (prev == LOWER || prev == DIGIT))
        || (cur == UPPER && prev == UPPER && next == LOWER)
        || ((cur == DIGIT) != (prev == DIGIT))
}

fn push_piece<'a>(out: &mut Vec<&'a str>, piece: &'a str) {
    if !piece.is_empty() && piece != "_" {
        out.push(piece);
    }
}

/// Tokens of a signature or doc comment: the text split at non-word characters, each word
/// split like an identifier. ASCII text is walked byte by byte in one pass; other text takes
/// the char path with the same boundary rule.
fn text_tokens_into<'a>(text: &'a str, out: &mut Vec<&'a str>) {
    if !text.is_ascii() {
        for word in text.split(|c: char| !c.is_alphanumeric() && c != '_') {
            split_identifier_into(word, out);
        }
        return;
    }
    let bytes = text.as_bytes();
    let mut start: Option<usize> = None;
    let mut prev = OTHER;
    for (i, &b) in bytes.iter().enumerate() {
        let cur = byte_class(b);
        if cur == OTHER {
            if let Some(s) = start.take() {
                push_piece(out, &text[s..i]);
            }
            continue;
        }
        match start {
            None => start = Some(i),
            Some(s) => {
                let next = bytes.get(i + 1).map_or(OTHER, |n| byte_class(*n));
                if identifier_boundary(prev, cur, next) {
                    push_piece(out, &text[s..i]);
                    start = Some(i);
                }
            }
        }
        prev = cur;
    }
    if let Some(s) = start {
        push_piece(out, &text[s..]);
    }
}

/// One admitted search row with the recall branches that reached it.
/// `result.score` is the word-branch BM25 for word rows and `0.0` otherwise until the
/// rerank replaces it.
#[derive(Clone)]
pub(crate) struct Candidate {
    pub result: SymbolSearchResult,
    pub bm25: Option<f64>,
    pub exact_name: bool,
    pub word_match: bool,
    pub name_match: bool,
    /// Admitted as a member of a class that a query word names.
    pub owner_match: bool,
    pub name_terms: Vec<String>,
    pub documentation: bool,
    /// A function or method declared inside another function or method.
    pub nested: bool,
    /// The class or other type the row is a member of.
    pub owner: Option<String>,
    /// Weak context from a highly ranked code type; never substitutes for a member's own hit.
    owner_context: Option<String>,
}

/// The kinds of symbol whose members search and lookup name as `Owner.member`.
const OWNER_KINDS: &str =
    "('class', 'struct', 'interface', 'trait', 'enum', 'record', 'object', 'protocol', 'union')";

fn candidate_columns(conn: &Connection) -> String {
    format!(
        "s.rowid AS row_id, s.symbol_id, s.file_id, s.path, s.language, s.name, s.kind,
                s.signature, s.doc_comment, s.visibility, s.parent_symbol_id, s.start_line,
                s.start_column, s.end_line, s.end_column, s.start_byte, s.end_byte,
                s.body_start_line, s.body_start_column, s.body_end_line, s.body_end_column,
                s.body_start_byte, s.body_end_byte, s.body_hash, s.semantic_group, s.is_test,
                s.test_container,
                (s.language IN ({doc_langs}) OR NOT ({not_doc})) AS documentation,
                (s.kind IN ('function', 'method')
                 AND EXISTS (SELECT 1 FROM symbols nest WHERE nest.symbol_id = s.parent_symbol_id
                             AND nest.kind IN ('function', 'method', 'constructor')
                             AND s.start_byte > nest.start_byte AND s.end_byte <= nest.end_byte)) AS nested,
                (SELECT o.name FROM symbols o WHERE o.symbol_id = s.parent_symbol_id
                   AND o.kind IN {OWNER_KINDS}) AS owner_name",
        doc_langs = documentation_language_list(),
        not_doc = not_documentation(conn, "s")
    )
}

/// julie records a Markdown or other documentation link as an `import` row; no lookup or search
/// for code wants it.
fn document_link_exclusion() -> String {
    format!(
        " AND NOT (s.kind = 'import' AND s.language IN ({}))",
        documentation_language_list()
    )
}

fn candidate_filters(searching_variables: bool, include_tests: bool) -> String {
    let mut sql = format!(
        " AND {KIND_FILTER}
          AND (:path IS NULL OR replace(s.path, '\\', '/') = :path COLLATE NOCASE OR replace(s.path, '\\', '/') LIKE :path_like || '/%' ESCAPE '\\' OR replace(s.path, '\\', '/') LIKE '%/' || :path_like ESCAPE '\\')",
    );
    sql.push_str(&document_link_exclusion());
    if !searching_variables {
        sql.push_str(&format!(" AND NOT {}", local_variable_predicate("s")));
    }
    if !include_tests {
        // Unary `+` keeps the planner off the test-flag indexes: without ANALYZE statistics it
        // would otherwise prefer them over the name index and walk nearly every row.
        sql.push_str(" AND +s.is_test = 0 AND +s.test_container = 0");
        sql.push_str(&format!(" AND NOT {}", test_path_predicate("s")));
    }
    sql
}

fn map_search_candidate(row: &Row, terms: &[String]) -> rusqlite::Result<(i64, Candidate)> {
    let symbol = map_symbol(row)?;
    let lower_name = symbol.name.to_lowercase();
    let name_terms = terms
        .iter()
        .filter(|t| lower_name.contains(t.as_str()))
        .cloned()
        .collect();
    let candidate = Candidate {
        result: SymbolSearchResult {
            symbol,
            score: 0.0,
            snippet: None,
            explain: None,
        },
        bm25: None,
        exact_name: false,
        word_match: false,
        name_match: false,
        owner_match: false,
        name_terms,
        documentation: row.get::<_, Option<i64>>("documentation")? == Some(1),
        nested: row.get::<_, Option<i64>>("nested")? == Some(1),
        owner: row.get("owner_name")?,
        owner_context: None,
    };
    Ok((row.get("row_id")?, candidate))
}

/// Runs the word, trigram-name, and exact-name branches with the same filters and merges
/// them by `rowid`. The word branch admits the rows that match every query word first and
/// then fills its cap with rows that match any word, so one full match never hides a
/// better partial one. Word BM25 exists only for word rows and never orders across branches.
/// A `variable` kind filter adds locals and parameters by name, because the FTS tables
/// exclude them.
pub(crate) fn collect_search_candidates(
    conn: &Connection,
    query: &str,
    kind_filter: Option<&str>,
    path_filter: Option<&str>,
    include_tests: bool,
    limit: usize,
) -> Result<Vec<Candidate>, QueryError> {
    let (and_q, or_q) = sanitize_fts5_query(query);
    let terms = trigram_name_terms(query);
    let normalized_path = path_filter.map(|p| {
        p.replace('\\', "/")
            .trim_start_matches("./")
            .trim_matches('/')
            .to_string()
    });
    let escaped_path = normalized_path.as_deref().map(escape_like);
    let norm_kind = kind_filter.map(normalize_kind);
    let searching_variables = norm_kind.as_deref() == Some("variable");
    let path_val = normalized_path.as_deref();
    let path_like = escaped_path.as_deref();
    let kind_val = norm_kind.as_deref();
    let columns = candidate_columns(conn);
    let filters = candidate_filters(searching_variables, include_tests);
    let word_cap = (limit * 4).clamp(40, 160);
    let name_cap = (limit * 2).clamp(20, 40);

    let new_candidate = |row: &Row| map_search_candidate(row, &terms);

    let mut candidates: Vec<Candidate> = Vec::new();
    let mut by_rowid: HashMap<i64, usize> = HashMap::new();
    let mut admit = |rowid: i64, incoming: Candidate| match by_rowid.get(&rowid).copied() {
        Some(i) => {
            let existing = &mut candidates[i];
            existing.exact_name |= incoming.exact_name;
            existing.word_match |= incoming.word_match;
            existing.name_match |= incoming.name_match;
            existing.owner_match |= incoming.owner_match;
            if incoming.bm25.is_some() && existing.bm25.is_none() {
                existing.bm25 = incoming.bm25;
                existing.result = incoming.result;
            }
        }
        None => {
            by_rowid.insert(rowid, candidates.len());
            candidates.push(incoming);
        }
    };

    let has_trigram = has_table(conn, "symbol_names_tri");
    let exact_query = query.trim();
    let exact_phrase = format!("\"{}\"", exact_query.replace('"', "\"\""));
    let exact_via_trigram = has_trigram && exact_query.chars().count() >= 3;
    let exact_sql = if exact_via_trigram {
        format!(
            "SELECT {columns} FROM symbol_names_tri
             CROSS JOIN symbols s ON s.rowid = symbol_names_tri.rowid
             WHERE symbol_names_tri MATCH :exact AND length(s.name) = length(:query) {filters}
             ORDER BY s.path ASC, s.start_line ASC LIMIT {MAX_RESULT_LIMIT}"
        )
    } else {
        format!(
            "SELECT {columns} FROM symbols s WHERE s.name = :query {filters}
             ORDER BY s.path ASC, s.start_line ASC LIMIT {MAX_RESULT_LIMIT}"
        )
    };
    let mut exact_params: Vec<(&str, &dyn ToSql)> = vec![
        (":query", &exact_query),
        (":kind", &kind_val),
        (":path", &path_val),
        (":path_like", &path_like),
    ];
    if exact_via_trigram {
        exact_params.push((":exact", &exact_phrase));
    }
    let exact_rows = conn
        .prepare(&exact_sql)?
        .query_map(exact_params.as_slice(), new_candidate)?
        .collect::<Result<Vec<_>, _>>()?;
    for (rowid, mut candidate) in exact_rows {
        candidate.exact_name = true;
        admit(rowid, candidate);
    }

    let owner_sql = if has_trigram {
        format!(
            "SELECT {columns} FROM symbol_names_tri
             CROSS JOIN symbols o ON o.rowid = symbol_names_tri.rowid
             CROSS JOIN symbols s ON s.parent_symbol_id = o.symbol_id
             WHERE symbol_names_tri MATCH :phrase AND length(o.name) = length(:word)
               AND o.kind IN {OWNER_KINDS} {filters}
             ORDER BY s.path ASC, s.start_line ASC LIMIT {name_cap}"
        )
    } else {
        format!(
            "SELECT {columns} FROM symbols o
             CROSS JOIN symbols s ON s.parent_symbol_id = o.symbol_id
             WHERE o.name = :word COLLATE NOCASE AND :phrase IS NOT NULL
               AND o.kind IN {OWNER_KINDS} {filters}
             ORDER BY s.path ASC, s.start_line ASC LIMIT {name_cap}"
        )
    };
    let mut owner_stmt = conn.prepare(&owner_sql)?;
    for word in rerank_words(query).iter().filter(|w| w.len() >= 3) {
        let phrase = format!("\"{word}\"");
        let owner_rows = owner_stmt
            .query_map(
                rusqlite::named_params! {
                    ":word": word,
                    ":phrase": phrase,
                    ":kind": kind_val,
                    ":path": path_val,
                    ":path_like": path_like,
                },
                new_candidate,
            )?
            .collect::<Result<Vec<_>, _>>()?;
        for (rowid, mut candidate) in owner_rows {
            candidate.owner_match = true;
            admit(rowid, candidate);
        }
    }

    if searching_variables {
        let pattern = format!("%{}%", escape_like(exact_query));
        let local_sql = format!(
            "SELECT {columns} FROM symbols s
             WHERE {local} AND (s.name = :query OR s.name LIKE :pattern ESCAPE '\\') {filters}
             ORDER BY (s.name = :query) DESC, length(s.name) ASC, s.path ASC LIMIT {limit}",
            local = local_variable_predicate("s")
        );
        let local_rows = conn
            .prepare(&local_sql)?
            .query_map(
                rusqlite::named_params! {
                    ":query": exact_query,
                    ":pattern": pattern,
                    ":kind": kind_val,
                    ":path": path_val,
                    ":path_like": path_like,
                },
                new_candidate,
            )?
            .collect::<Result<Vec<_>, _>>()?;
        for (rowid, mut candidate) in local_rows {
            candidate.exact_name = candidate.result.symbol.name == exact_query;
            candidate.name_match = true;
            admit(rowid, candidate);
        }
    }

    let word_sql = format!(
        "SELECT {columns},
                bm25(symbols_fts, 10.0, 5.0, 1.0) AS rank_score,
                snippet(symbols_fts, 2, '[', ']', '...', 12) AS doc_snippet,
                snippet(symbols_fts, 1, '[', ']', '...', 12) AS sig_snippet,
                snippet(symbols_fts, 0, '[', ']', '...', 12) AS name_snippet
         FROM symbols_fts
         CROSS JOIN symbols s ON s.rowid = symbols_fts.rowid
         WHERE symbols_fts MATCH :match {filters}
         ORDER BY (s.kind = 'import') ASC, (s.language IN ({doc_langs})) ASC, {not_doc} DESC, (s.name = :query COLLATE NOCASE) DESC, rank_score ASC LIMIT {word_cap}",
        doc_langs = documentation_language_list(),
        not_doc = not_documentation(conn, "s")
    );
    let word_rows = |match_clause: &str| -> Result<Vec<(i64, Candidate)>, QueryError> {
        let map_fn = |row: &Row| -> rusqlite::Result<(i64, Candidate)> {
            let (rowid, mut candidate) = new_candidate(row)?;
            let score: f64 = row.get("rank_score")?;
            let doc_snip: Option<String> = row.get("doc_snippet").ok();
            let sig_snip: Option<String> = row.get("sig_snippet").ok();
            let name_snip: Option<String> = row.get("name_snippet").ok();
            let highlighted = |s: &Option<String>| s.as_ref().is_some_and(|s| s.contains('['));
            candidate.result.snippet = if highlighted(&doc_snip) {
                doc_snip
            } else if highlighted(&sig_snip) {
                sig_snip
            } else if highlighted(&name_snip) {
                name_snip
            } else {
                doc_snip.or(sig_snip).or(name_snip)
            };
            candidate.result.score = score;
            candidate.bm25 = Some(score);
            candidate.word_match = true;
            Ok((rowid, candidate))
        };
        Ok(conn
            .prepare(&word_sql)?
            .query_map(
                rusqlite::named_params! {
                    ":match": match_clause,
                    ":query": query.trim(),
                    ":kind": kind_val,
                    ":path": path_val,
                    ":path_like": path_like,
                },
                map_fn,
            )?
            .collect::<Result<Vec<_>, _>>()?)
    };
    if !and_q.is_empty() {
        let and_rows = word_rows(&and_q)?;
        let mut word_admitted: HashSet<i64> = and_rows.iter().map(|(rowid, _)| *rowid).collect();
        for (rowid, candidate) in and_rows {
            admit(rowid, candidate);
        }
        if and_q != or_q {
            for (rowid, candidate) in word_rows(&or_q)? {
                if word_admitted.len() >= word_cap && !word_admitted.contains(&rowid) {
                    break;
                }
                word_admitted.insert(rowid);
                admit(rowid, candidate);
            }
        }
    }

    if !terms.is_empty() && has_trigram {
        let match_clause = terms
            .iter()
            .map(|t| format!("\"{t}\""))
            .collect::<Vec<_>>()
            .join(" OR ");
        let name_sql = format!(
            "SELECT {columns} FROM symbol_names_tri
             CROSS JOIN symbols s ON s.rowid = symbol_names_tri.rowid
             WHERE symbol_names_tri MATCH :match {filters}
             ORDER BY bm25(symbol_names_tri) ASC, length(s.name) ASC, s.path ASC LIMIT {name_cap}"
        );
        let name_rows = conn
            .prepare(&name_sql)?
            .query_map(
                rusqlite::named_params! {
                    ":match": match_clause,
                    ":kind": kind_val,
                    ":path": path_val,
                    ":path_like": path_like,
                },
                new_candidate,
            )?
            .collect::<Result<Vec<_>, _>>()?;
        for (rowid, mut candidate) in name_rows {
            candidate.name_match = true;
            admit(rowid, candidate);
        }
    }

    let inherited = inherited_writes_among(
        conn,
        candidates
            .iter()
            .map(|c| c.result.symbol.symbol_id.as_str()),
    )?;
    candidates.retain(|c| !inherited.contains(&c.result.symbol.symbol_id));
    Ok(candidates)
}

/// Follow only a few highly ranked code types, one level down. A member must still match
/// the query itself; the owner's context can supply only weaker, otherwise missing terms.
fn expand_ranked_owner_members(
    conn: &Connection,
    candidates: &mut Vec<Candidate>,
    ranked: &[(SymbolSearchResult, SearchExplain)],
    query: &str,
    kind_filter: Option<&str>,
    path_filter: Option<&str>,
    include_tests: bool,
) -> Result<bool, QueryError> {
    // ponytail: First 3 code types in the top 20, at most 40 direct members each.
    // Raise these ceilings only when a larger held-out evaluation validates the recall need.
    let owners: Vec<_> = ranked
        .iter()
        .take(20)
        .filter(|(result, explain)| {
            explain.documentation == 0.0
                && OWNER_KINDS.contains(&format!("'{}'", result.symbol.kind))
        })
        .take(3)
        .map(|(result, _)| &result.symbol)
        .collect();
    if owners.is_empty() {
        return Ok(false);
    }
    let normalized_path = path_filter.map(|p| {
        p.replace('\\', "/")
            .trim_start_matches("./")
            .trim_matches('/')
            .to_string()
    });
    let escaped_path = normalized_path.as_deref().map(escape_like);
    let norm_kind = kind_filter.map(normalize_kind);
    let columns = candidate_columns(conn);
    let filters = candidate_filters(norm_kind.as_deref() == Some("variable"), include_tests);
    let mut stmt = conn.prepare(&format!(
        "SELECT {columns} FROM symbols s WHERE s.parent_symbol_id = :owner {filters}
         ORDER BY s.start_line, s.start_column, s.symbol_id LIMIT 40"
    ))?;
    let terms = trigram_name_terms(query);
    let stemmer = Stemmer::create(Algorithm::English);
    let words = stemmed_query_words(query, &stemmer);
    let mut members = Vec::new();
    for owner in owners {
        let context = format!(
            "{} {}",
            owner.name,
            head_bytes(owner.doc_comment.as_deref().unwrap_or(""), TEXT_HEAD_BYTES)
        );
        let rows = stmt.query_map(
            rusqlite::named_params! {
                ":owner": owner.symbol_id,
                ":kind": norm_kind.as_deref(),
                ":path": normalized_path.as_deref(),
                ":path_like": escaped_path.as_deref(),
            },
            |row| map_search_candidate(row, &terms),
        )?;
        for row in rows {
            let (_, mut member) = row?;
            let symbol = &member.result.symbol;
            let mut tokens = Vec::new();
            let named = name_hits(&symbol.name, &words, &stemmer)
                .iter()
                .any(|hit| *hit > 0);
            let text_match = [symbol.signature.as_deref(), symbol.doc_comment.as_deref()]
                .into_iter()
                .any(|text| {
                    text_hits(
                        text.map(|s| head_bytes(s, TEXT_HEAD_BYTES)),
                        &words,
                        &mut tokens,
                    )
                    .into_iter()
                    .any(|hit| hit)
                });
            if named || text_match {
                member.owner_context = Some(context.clone());
                members.push(member);
            }
        }
    }
    let inherited = inherited_writes_among(
        conn,
        members.iter().map(|c| c.result.symbol.symbol_id.as_str()),
    )?;
    let mut changed = false;
    for member in members {
        let id = &member.result.symbol.symbol_id;
        if inherited.contains(id) {
            continue;
        }
        if let Some(existing) = candidates
            .iter_mut()
            .find(|c| c.result.symbol.symbol_id == *id)
        {
            existing.owner_context = member.owner_context;
        } else {
            candidates.push(member);
        }
        changed = true;
    }
    Ok(changed)
}

/// Conceptual full-text search with optional path scoping filter.
pub fn fts_search_symbols_scoped(
    conn: &Connection,
    query: &str,
    kind_filter: Option<&str>,
    path_filter: Option<&str>,
    include_tests: bool,
    limit: usize,
) -> Result<Vec<SymbolSearchResult>, QueryError> {
    fts_search_symbols_explained(
        conn,
        query,
        kind_filter,
        path_filter,
        include_tests,
        limit,
        false,
    )
}

/// Conceptual full-text search that also attaches the rerank breakdown to every row when
/// `explain` is true. Without it, `explain` stays `None` on every row.
pub fn fts_search_symbols_explained(
    conn: &Connection,
    query: &str,
    kind_filter: Option<&str>,
    path_filter: Option<&str>,
    include_tests: bool,
    limit: usize,
    explain: bool,
) -> Result<Vec<SymbolSearchResult>, QueryError> {
    validate_result_limit(limit)?;
    if limit == 0 {
        return Ok(Vec::new());
    }
    let (and_q, _) = sanitize_fts5_query(query);
    if and_q.is_empty() {
        return Ok(Vec::new());
    }

    let normalized_path = path_filter.map(|p| {
        p.replace('\\', "/")
            .trim_start_matches("./")
            .trim_matches('/')
            .to_string()
    });
    let norm_kind = kind_filter.map(normalize_kind);
    let escaped_path = normalized_path.as_deref().map(escape_like);
    let searching_variables = norm_kind.as_deref() == Some("variable");

    let name_search = |local_clause: &str| -> Result<Vec<SymbolSearchResult>, QueryError> {
        let pattern = format!("%{}%", escape_like(query));
        let mut sql = format!(
            "SELECT symbol_id, file_id, path, language, name, kind, signature, doc_comment,
                    visibility, parent_symbol_id, start_line, start_column, end_line, end_column,
                    start_byte, end_byte, body_start_line, body_start_column, body_end_line,
                    body_end_column, body_start_byte, body_end_byte, body_hash, semantic_group,
                    is_test, test_container
              FROM symbols s
              WHERE (name = :query OR name LIKE :pattern ESCAPE '\\')
                AND {KIND_FILTER}
                AND (:path IS NULL OR replace(path, '\\', '/') = :path COLLATE NOCASE OR replace(path, '\\', '/') LIKE :path_like || '/%' ESCAPE '\\' OR replace(path, '\\', '/') LIKE '%/' || :path_like ESCAPE '\\')",
        );
        sql.push_str(&document_link_exclusion());
        sql.push_str(local_clause);
        if !include_tests {
            sql.push_str(" AND is_test = 0 AND test_container = 0");
            sql.push_str(&format!(" AND NOT {}", test_path_predicate("s")));
        }
        sql.push_str(
            " ORDER BY (name = :query) DESC, (kind IN ('function', 'struct', 'class', 'trait', 'method', 'enum', 'interface', 'type')) DESC, length(name) ASC, path ASC LIMIT ",
        );
        sql.push_str(&limit.to_string());

        let mut stmt = conn.prepare(&sql)?;
        let path_val = normalized_path.as_deref();
        let path_like = escaped_path.as_deref();
        let kind_val = norm_kind.as_deref();
        let rows = stmt
            .query_map(
                rusqlite::named_params! {
                    ":query": query,
                    ":pattern": pattern,
                    ":kind": kind_val,
                    ":path": path_val,
                    ":path_like": path_like,
                },
                map_symbol,
            )?
            .collect::<Result<Vec<_>, _>>()?;

        Ok(rows
            .into_iter()
            .map(|s| SymbolSearchResult {
                symbol: s,
                score: 0.0,
                snippet: None,
                explain: None,
            })
            .collect())
    };

    if !has_table(conn, "symbols_fts") {
        let local_clause = if searching_variables {
            String::new()
        } else {
            format!(" AND NOT {}", local_variable_predicate("s"))
        };
        return name_search(&local_clause);
    }

    let mut candidates =
        collect_search_candidates(conn, query, kind_filter, path_filter, include_tests, limit)?;
    let started = std::time::Instant::now();
    let idf = idf_weights(conn, &rerank_words(query));
    let mut ranked = rerank_with(candidates.clone(), query, include_tests, Some(&idf));
    let mut rerank_us = started.elapsed().as_micros();
    let mut candidate_count = candidates.len();
    if expand_ranked_owner_members(
        conn,
        &mut candidates,
        &ranked,
        query,
        kind_filter,
        path_filter,
        include_tests,
    )? {
        candidate_count = candidates.len();
        let started = std::time::Instant::now();
        ranked = rerank_with(candidates, query, include_tests, Some(&idf));
        rerank_us += started.elapsed().as_micros();
    }
    let unique_query_terms: HashSet<String> = rerank_words(query).into_iter().collect();
    let min_multi_term_results = ((limit * 3).div_ceil(4)).max(15).min(limit);
    let multi_term_results = ranked
        .iter()
        .filter(|(_, breakdown)| covered_query_term_count(&breakdown.terms) >= 2)
        .count();
    if unique_query_terms.len() > 1 && multi_term_results >= min_multi_term_results {
        let mut strong_results_seen = 0;
        ranked.retain(|(_, breakdown)| {
            if covered_query_term_count(&breakdown.terms) >= 2 {
                strong_results_seen += 1;
                true
            } else {
                strong_results_seen < min_multi_term_results || has_distinctive_name_term(breakdown)
            }
        });
    }
    Ok(ranked
        .into_iter()
        .take(limit)
        .map(|(mut result, mut breakdown)| {
            if explain {
                breakdown.candidates = candidate_count;
                breakdown.rerank_us = rerank_us;
                result.explain = Some(breakdown);
            }
            result
        })
        .collect())
}

const W_NAME_WHOLE: f64 = 100.0;
const W_NAME_ALL_WORDS: f64 = 60.0;
const W_KIND_DEFINITION: f64 = 4.0;
const W_KIND_MEMBER: f64 = 0.0;
const W_KIND_IMPORT: f64 = -50.0;
/// A helper declared inside a function ranks below a module or class member of equal score.
const W_NESTED: f64 = -2.0;
const W_PATH_ROLE: f64 = -10.0;
const W_DOCUMENTATION_ROW: f64 = -200.0;
const W_TEST_INTENT: f64 = 5.0;
const W_TERMS: f64 = 52.0;
const MAX_TERM_CREDIT: f64 = 3.0;
const TEXT_CREDIT: f64 = 1.0;
const FIELD_DIVERSITY_BONUS: f64 = 6.0;
const DISTINCTIVE_NAME_IDF_FACTOR: f64 = 1.1;
/// A word that names the class a member belongs to: `Flask init` means `Flask.__init__`.
const OWNER_CREDIT: f64 = 3.0;
const OWNER_CONTEXT_CREDIT: f64 = 0.5;
const TEXT_HEAD_BYTES: usize = 400;
const DOC_HEAD_BYTES: usize = 640;

/// Symbol kinds that define a body: the kinds a touched-symbol or ranking rule prefers over locals.
pub const DEFINITION_KINDS: &[&str] = &[
    "function",
    "method",
    "class",
    "struct",
    "trait",
    "interface",
    "enum",
    "type",
];
const MEMBER_KINDS: &[&str] = &["enum_member", "field", "property", "constant", "variable"];
const DEMOTED_PATH_SEGMENTS: &[&str] = &["scripts", "examples", "benchmarks", "fixtures", "vendor"];
const TEST_INTENT_WORDS: &[&str] = &["test", "tests", "spec", "specs"];

struct QueryWord {
    word: String,
    stem: String,
}

fn stemmed_query_words(query: &str, stemmer: &Stemmer) -> Vec<QueryWord> {
    rerank_words(query)
        .into_iter()
        .map(|word| QueryWord {
            stem: stemmer.stem(&word).into_owned(),
            word,
        })
        .collect()
}

/// Per-word hits of one candidate: which query words its name, signature, and capped doc
/// cover. The name carries a match strength per word, the others a plain hit.
struct Hits {
    name: Vec<u8>,
    /// The query word is the whole name of the class the row belongs to, or its stem. Counted
    /// only when another word matches the row's own name, as in `Flask init`.
    owner: Vec<bool>,
    signature: Vec<bool>,
    doc: Vec<bool>,
    owner_context: Vec<bool>,
}

/// Symbols a query word may be counted in before it is called common: the count walks the
/// word's postings, so the cap bounds the cost of a word held by most of a million symbols.
const DF_CAP: usize = 20_000;

/// Global rarity of each lowercase rerank term: `ln(1 + N / (df + 1))`, where `df` is the
/// number of indexed symbols holding the term (capped at `DF_CAP`) and `N` an upper bound on
/// the index size. The count is an FTS5 `MATCH`, so the term is stemmed by the index's own
/// tokenizer and `news` finds the rows indexed as `new`.
fn idf_weights(conn: &Connection, words: &[String]) -> Vec<f64> {
    let n = conn
        .query_row("SELECT max(rowid) FROM symbols", [], |r| {
            r.get::<_, Option<i64>>(0)
        })
        .ok()
        .flatten()
        .unwrap_or(0) as f64;
    let mut count = conn
        .prepare(&format!(
            "SELECT count(*) FROM (SELECT rowid FROM symbols_fts WHERE symbols_fts MATCH ?1 LIMIT {DF_CAP})"
        ))
        .ok();
    words
        .iter()
        .map(|word| {
            let df = count
                .as_mut()
                .and_then(|stmt| document_frequency(stmt, &format!("\"{word}\"")))
                .unwrap_or(0);
            (1.0 + n / (df as f64 + 1.0)).ln()
        })
        .collect()
}

fn document_frequency(stmt: &mut rusqlite::Statement<'_>, term: &str) -> Option<i64> {
    stmt.query_row([term], |r| r.get(0)).ok()
}

/// The field that credits each query term and the credit it is worth: a name whole token 3,
/// a name stem 2, the owning class `OWNER_CREDIT`, a signature or doc hit `TEXT_CREDIT`, a name
/// substring 1, weak owner context 0.5 for an otherwise unmatched term, nothing 0.
fn term_credits(hits: &Hits, words: &[QueryWord]) -> Vec<(String, String, f64)> {
    words
        .iter()
        .enumerate()
        .map(|(i, w)| {
            let (field, credit) = match hits.name[i] {
                3 => ("name", 3.0),
                2 => ("name", 2.0),
                _ if hits.owner[i] => ("owner", OWNER_CREDIT),
                _ if hits.signature[i] => ("signature", TEXT_CREDIT),
                _ if hits.doc[i] => ("doc", TEXT_CREDIT),
                1 => ("name", 1.0),
                _ if hits.owner_context[i] => ("owner_context", OWNER_CONTEXT_CREDIT),
                _ => ("none", 0.0),
            };
            (w.word.clone(), field.to_string(), credit)
        })
        .collect()
}

/// Adds a prior when one whole name token, the signature, and at least two distinct doc terms
/// independently support the query, which the row covers with at least half its distinct terms.
/// The caller limits it to functions and methods, so request-shaped types cannot earn it.
fn credited_field_bonus(terms: &[(String, String, f64)]) -> f64 {
    let query_terms: HashSet<&str> = terms.iter().map(|(word, _, _)| word.as_str()).collect();
    let covered_terms: HashSet<&str> = terms
        .iter()
        .filter(|(_, _, credit)| *credit >= TEXT_CREDIT)
        .map(|(word, _, _)| word.as_str())
        .collect();
    let adequate_query_coverage =
        !query_terms.is_empty() && covered_terms.len() * 2 >= query_terms.len();
    let mut document_terms = HashSet::new();
    let has_exact_name_term = terms
        .iter()
        .any(|(_, field, credit)| field == "name" && *credit == 3.0);
    let fields = terms.iter().filter(|(_, _, credit)| *credit > 0.0).fold(
        0u8,
        |fields, (word, field, _)| {
            if field == "doc" {
                // The query can repeat a word; count distinct words as independent evidence.
                document_terms.insert(word.as_str());
            }
            fields
                | match field.as_str() {
                    "name" => 1,
                    "owner" => 2,
                    "signature" => 4,
                    "doc" => 8,
                    "owner_context" => 16,
                    _ => 0,
                }
        },
    );
    if has_exact_name_term
        && fields & 4 != 0
        && fields & 8 != 0
        && document_terms.len() >= 2
        && adequate_query_coverage
    {
        FIELD_DIVERSITY_BONUS
    } else {
        0.0
    }
}

fn covered_query_term_count(terms: &[(String, String, f64)]) -> usize {
    terms
        .iter()
        .filter(|(_, _, credit)| *credit > 0.0)
        .map(|(word, _, _)| word.as_str())
        .collect::<HashSet<_>>()
        .len()
}

fn has_distinctive_name_term(explain: &SearchExplain) -> bool {
    let weights = &explain.word_weights;
    let Some(mean_weight) = (!weights.is_empty())
        .then(|| weights.iter().map(|(_, weight)| weight).sum::<f64>() / weights.len() as f64)
    else {
        return false;
    };
    explain
        .terms
        .iter()
        .zip(weights)
        .any(|((term, field, credit), (weighted_term, weight))| {
            term == weighted_term
                && field == "name"
                && *credit >= 2.0
                && *weight >= mean_weight * DISTINCTIVE_NAME_IDF_FACTOR
        })
}

fn term_score(terms: &[(String, String, f64)], weights: &[f64]) -> f64 {
    let total: f64 = weights.iter().sum();
    if total == 0.0 {
        return 0.0;
    }
    let credited: f64 = terms
        .iter()
        .zip(weights)
        .map(|((_, _, credit), weight)| credit * weight)
        .sum();
    W_TERMS * credited / (MAX_TERM_CREDIT * total)
}

/// Lowercase query words for the rerank: every `query_words` token is split like an
/// identifier, and stop words are dropped only when a content word remains.
fn rerank_words(query: &str) -> Vec<String> {
    let words: Vec<String> = query_words(query)
        .into_iter()
        .flat_map(split_identifier)
        .map(str::to_lowercase)
        .collect();
    let any_content = words.iter().any(|w| !is_stop_word(w));
    words
        .into_iter()
        .filter(|w| !any_content || !is_stop_word(w))
        .collect()
}

fn collapse(text: &str) -> String {
    text.chars()
        .filter(|c| c.is_alphanumeric())
        .flat_map(char::to_lowercase)
        .collect()
}

fn head_bytes(text: &str, bytes: usize) -> &str {
    let mut end = bytes.min(text.len());
    while !text.is_char_boundary(end) {
        end -= 1;
    }
    &text[..end]
}

fn token_run_equals(tokens: &[String], word: &str) -> bool {
    (0..tokens.len()).any(|start| {
        let mut joined = String::new();
        for token in &tokens[start..] {
            joined.push_str(token);
            if joined.len() >= word.len() {
                return joined == word;
            }
        }
        false
    })
}

/// Match strength of a name per query word: `3` when a run of its tokens equals the word,
/// `2` when a token stem equals the word stem, `1` when the collapsed name contains the word,
/// `0` when nothing matches.
fn name_hits(name: &str, words: &[QueryWord], stemmer: &Stemmer) -> Vec<u8> {
    let tokens: Vec<String> = split_identifier(name)
        .into_iter()
        .map(str::to_lowercase)
        .collect();
    let stems: Vec<String> = tokens
        .iter()
        .map(|t| stemmer.stem(t).into_owned())
        .collect();
    let collapsed = collapse(name);
    words
        .iter()
        .map(|w| {
            if token_run_equals(&tokens, &w.word) {
                3
            } else if stems.contains(&w.stem) {
                2
            } else if w.word.chars().count() >= 3 && collapsed.contains(&w.word) {
                1
            } else {
                0
            }
        })
        .collect()
}

/// True when `token` lowercased starts with `prefix` (or equals it when `exact`).
/// `prefix` is already lowercase.
fn lowercase_prefix_match(token: &str, prefix: &str, exact: bool) -> bool {
    if token.is_ascii() && prefix.is_ascii() {
        let Some(head) = token.as_bytes().get(..prefix.len()) else {
            return false;
        };
        return head.eq_ignore_ascii_case(prefix.as_bytes())
            && (!exact || token.len() == prefix.len());
    }
    let mut lower = token.chars().flat_map(char::to_lowercase);
    for expected in prefix.chars() {
        if lower.next() != Some(expected) {
            return false;
        }
    }
    !exact || lower.next().is_none()
}

/// Token-level coverage of a signature or doc: a word is covered when some token equals it,
/// or starts with its stem or with the word itself (three or more characters), so `stemming`
/// covers `stemmer` and `stems` but not `system`.
fn text_hits<'a>(
    text: Option<&'a str>,
    words: &[QueryWord],
    tokens: &mut Vec<&'a str>,
) -> Vec<bool> {
    tokens.clear();
    text_tokens_into(text.unwrap_or(""), tokens);
    words
        .iter()
        .map(|w| {
            let stem_prefix = w.stem.chars().count() >= 3;
            let exact_word = w.word.chars().count() < 3;
            tokens.iter().any(|t| {
                lowercase_prefix_match(t, &w.word, exact_word)
                    || (stem_prefix && lowercase_prefix_match(t, &w.stem, false))
            })
        })
        .collect()
}

fn kind_prior(kind: &str) -> f64 {
    let kind = normalize_kind(kind);
    match kind.as_str() {
        "import" => W_KIND_IMPORT,
        k if DEFINITION_KINDS.contains(&k) => W_KIND_DEFINITION,
        k if MEMBER_KINDS.contains(&k) => W_KIND_MEMBER,
        _ => 0.0,
    }
}

fn path_role(path: &str, words: &[QueryWord], stemmer: &Stemmer) -> f64 {
    let Some(segment) = path.split(['/', '\\']).find(|seg| {
        DEMOTED_PATH_SEGMENTS
            .iter()
            .any(|d| d.eq_ignore_ascii_case(seg))
    }) else {
        return 0.0;
    };
    let segment = segment.to_lowercase();
    let segment_stem = stemmer.stem(&segment);
    let named = words.iter().any(|w| {
        w.word == segment || w.word == segment_stem || w.stem == segment || w.stem == segment_stem
    });
    if named { 0.0 } else { W_PATH_ROLE }
}

fn bracket_longest_term(name: &str, terms: &[String]) -> String {
    let lower = name.to_lowercase();
    if lower.len() != name.len() {
        return name.to_string();
    }
    let mut best: Option<(usize, usize)> = None;
    for term in terms {
        if let Some(start) = lower.find(term.as_str()) {
            let end = start + term.len();
            let longer = best.is_none_or(|(s, e)| end - start > e - s);
            if longer && name.is_char_boundary(start) && name.is_char_boundary(end) {
                best = Some((start, end));
            }
        }
    }
    match best {
        Some((start, end)) => {
            format!("{}[{}]{}", &name[..start], &name[start..end], &name[end..])
        }
        None => name.to_string(),
    }
}

fn branch_snippet(candidate: &Candidate) -> Option<String> {
    let name = &candidate.result.symbol.name;
    if candidate.word_match {
        candidate.result.snippet.clone()
    } else if candidate.exact_name {
        Some(name.clone())
    } else {
        Some(bracket_longest_term(name, &candidate.name_terms))
    }
}

/// Scores every admitted candidate and returns them best first. Each query term is credited
/// once from its strongest field and weighted by `idf`, the term's rarity across the index;
/// a weight vector of the wrong length, or none, makes every term count the same. Ties fall
/// to name strength, then word BM25 (rows without one last), then name length, path, name.
fn rerank_with(
    candidates: Vec<Candidate>,
    query: &str,
    include_tests: bool,
    idf: Option<&[f64]>,
) -> Vec<(SymbolSearchResult, SearchExplain)> {
    let stemmer = Stemmer::create(Algorithm::English);
    let words = stemmed_query_words(query, &stemmer);
    let collapsed_query = collapse(query);
    let test_intent = include_tests
        && words
            .iter()
            .any(|w| TEST_INTENT_WORDS.contains(&w.word.as_str()));

    let mut tokens: Vec<&str> = Vec::new();
    let hits: Vec<Hits> = candidates
        .iter()
        .map(|candidate| {
            let symbol = &candidate.result.symbol;
            let name = name_hits(&symbol.name, &words, &stemmer);
            let named = name.iter().any(|strength| *strength > 0);
            Hits {
                owner: match candidate.owner.as_deref().map(collapse) {
                    Some(owner) if named => {
                        let owner_stem = stemmer.stem(&owner);
                        words
                            .iter()
                            .map(|w| owner == w.word || owner_stem == w.stem)
                            .collect()
                    }
                    _ => vec![false; words.len()],
                },
                name,
                signature: text_hits(
                    symbol
                        .signature
                        .as_deref()
                        .map(|signature| head_bytes(signature, TEXT_HEAD_BYTES)),
                    &words,
                    &mut tokens,
                ),
                doc: text_hits(
                    symbol
                        .doc_comment
                        .as_deref()
                        .map(|doc| head_bytes(doc, DOC_HEAD_BYTES)),
                    &words,
                    &mut tokens,
                ),
                owner_context: text_hits(candidate.owner_context.as_deref(), &words, &mut tokens),
            }
        })
        .collect();
    let term_weights: Vec<f64> = match idf {
        Some(weights) if weights.len() == words.len() => weights.to_vec(),
        _ => vec![1.0; words.len()],
    };
    let word_weights: Vec<(String, f64)> = words
        .iter()
        .zip(&term_weights)
        .map(|(w, weight)| (w.word.clone(), *weight))
        .collect();

    let mut scored: Vec<(SymbolSearchResult, SearchExplain)> = candidates
        .into_iter()
        .zip(hits)
        .map(|(candidate, hits)| {
            let symbol = &candidate.result.symbol;
            let name_strength: u32 = hits.name.iter().map(|s| u32::from(*s)).sum();
            let tier = if !collapsed_query.is_empty() && collapse(&symbol.name) == collapsed_query {
                "whole"
            } else if !hits.name.is_empty() && hits.name.iter().all(|s| *s >= 2) {
                "all"
            } else if hits.name.iter().any(|s| *s > 0) {
                "partial"
            } else {
                "none"
            };
            let terms = term_credits(&hits, &words);
            let field_diversity_bonus = if matches!(symbol.kind.as_str(), "function" | "method") {
                credited_field_bonus(&terms)
            } else {
                0.0
            };
            let explain = SearchExplain {
                bm25: candidate.bm25,
                branches: [
                    (candidate.exact_name, "exact"),
                    (candidate.word_match, "word"),
                    (candidate.name_match, "name"),
                    (candidate.owner_match, "owner"),
                    (candidate.owner_context.is_some(), "owner_context"),
                ]
                .into_iter()
                .filter(|(hit, _)| *hit)
                .map(|(_, branch)| branch.to_string())
                .collect(),
                name_tier: tier.to_string(),
                name_strength,
                term_score: term_score(&terms, &term_weights),
                field_diversity_bonus,
                name_bonus: match tier {
                    "whole"
                        if DEFINITION_KINDS.contains(&normalize_kind(&symbol.kind).as_str()) =>
                    {
                        W_NAME_WHOLE
                    }
                    "whole" | "all" => W_NAME_ALL_WORDS,
                    _ => 0.0,
                },
                kind_prior: kind_prior(&symbol.kind),
                path_role: path_role(&symbol.path, &words, &stemmer),
                documentation: if candidate.documentation {
                    W_DOCUMENTATION_ROW
                } else {
                    0.0
                },
                test_intent: if test_intent && (symbol.is_test || symbol.test_container) {
                    W_TEST_INTENT
                } else {
                    0.0
                },
                nested: if candidate.nested { W_NESTED } else { 0.0 },
                terms,
                word_weights: word_weights.clone(),
                candidates: 0,
                rerank_us: 0,
            };
            let score = explain.term_score
                + explain.field_diversity_bonus
                + explain.name_bonus
                + explain.kind_prior
                + explain.path_role
                + explain.documentation
                + explain.test_intent
                + explain.nested;
            let snippet = branch_snippet(&candidate);
            let mut result = candidate.result;
            result.score = score;
            result.snippet = snippet;
            (result, explain)
        })
        .collect();

    scored.sort_by(|(a, ea), (b, eb)| {
        b.score
            .total_cmp(&a.score)
            .then_with(|| eb.name_strength.cmp(&ea.name_strength))
            .then_with(|| {
                a.symbol
                    .name
                    .starts_with('_')
                    .cmp(&b.symbol.name.starts_with('_'))
            })
            .then_with(|| ea.bm25.is_none().cmp(&eb.bm25.is_none()))
            .then_with(|| ea.bm25.unwrap_or(0.0).total_cmp(&eb.bm25.unwrap_or(0.0)))
            .then_with(|| a.symbol.name.len().cmp(&b.symbol.name.len()))
            .then_with(|| a.symbol.path.cmp(&b.symbol.path))
            .then_with(|| a.symbol.name.cmp(&b.symbol.name))
    });
    scored
}

const LOW_SIGNAL_KINDS_SQL: &str =
    "'import','variable','parameter','field','property','module','namespace'";

/// Find tests related to a target symbol by caller relationships, naming pattern, or FTS matching.
pub fn find_related_tests(
    conn: &Connection,
    target_symbol: &Symbol,
    limit: usize,
) -> Result<Vec<Symbol>, QueryError> {
    if limit == 0 || DOCUMENTATION_LANGUAGES.contains(&target_symbol.language.as_str()) {
        return Ok(Vec::new());
    }

    const COLUMNS: &str = "s.symbol_id, s.file_id, s.path, s.language, s.name, s.kind, s.signature, s.doc_comment,
            s.visibility, s.parent_symbol_id, s.start_line, s.start_column, s.end_line, s.end_column,
            s.start_byte, s.end_byte, s.body_start_line, s.body_start_column, s.body_end_line,
            s.body_end_column, s.body_start_byte, s.body_end_byte, s.body_hash, s.semantic_group,
            s.is_test, s.test_container";
    // Rows a scan writes can land in any order, so every query below needs a full ORDER BY.
    const TEST_ORDER: &str = "s.is_test DESC, s.path, s.start_line, s.symbol_id";
    let not_setup = |alias: &str| {
        if has_column(conn, "symbols", "test_lifecycle") {
            format!("COALESCE({alias}.test_lifecycle, 0) = 0")
        } else {
            "1 = 1".to_string()
        }
    };
    let is_test = format!(
        "((s.is_test = 1 OR s.test_container = 1
           OR ({} AND s.kind IN ('function', 'method') AND ({}
               OR EXISTS (SELECT 1 FROM symbols tc
                          WHERE tc.symbol_id = s.parent_symbol_id
                            AND (tc.test_container = 1
                                 OR (tc.kind IN ('function', 'method') AND {}
                                     AND (tc.is_test = 1 OR {})))))))
          AND {})",
        test_path_predicate("s"),
        names_a_test_sql("s"),
        not_setup("tc"),
        names_a_test_sql("tc"),
        not_setup("s")
    );
    let not_documentation = not_documentation(conn, "s");

    let mut tests = Vec::new();
    let mut seen_ids = std::collections::HashSet::from([target_symbol.symbol_id.clone()]);

    let callers_sql = format!(
        "SELECT {COLUMNS}
     FROM symbols s
     JOIN relationships r ON r.from_symbol_id = s.symbol_id
     WHERE r.to_symbol_id = ?1 AND {is_test} AND {not_documentation}
     ORDER BY {TEST_ORDER}
     LIMIT ?2"
    );

    if let Ok(mut stmt) = conn.prepare(&callers_sql)
        && let Ok(rows) = stmt.query_map(params![target_symbol.symbol_id, limit as i64], map_symbol)
    {
        for row in rows.flatten() {
            let row = enclosing_test(conn, row)?;
            if seen_ids.insert(row.symbol_id.clone()) {
                tests.push(row);
                if tests.len() >= limit {
                    return Ok(tests);
                }
            }
        }
    }

    // julie resolves call edges inside one file only; every cross-file caller is a pending edge
    let remaining = limit - tests.len();
    if remaining > 0 && has_pending_namespace_column(conn) {
        let pending_sql = format!(
            "SELECT DISTINCT {COLUMNS}
     FROM pending_relationships p
     JOIN symbols s ON p.from_symbol_id = s.symbol_id
     JOIN symbols s_from ON s_from.symbol_id = s.symbol_id
     JOIN symbols s_target ON s_target.symbol_id = ?1
     LEFT JOIN symbols s_target_parent ON s_target.parent_symbol_id = s_target_parent.symbol_id
     WHERE p.target_terminal_name = s_target.name
       AND {is_test}
       AND {not_documentation}
       AND {pred}
     ORDER BY {TEST_ORDER}
     LIMIT ?2",
            pred = pending_target_predicate(conn, "s_target", "s_target_parent")
        );

        if let Ok(mut stmt) = conn.prepare(&pending_sql)
            && let Ok(rows) = stmt.query_map(
                params![target_symbol.symbol_id, remaining as i64],
                map_symbol,
            )
        {
            for row in rows.flatten() {
                let row = enclosing_test(conn, row)?;
                if seen_ids.insert(row.symbol_id.clone()) {
                    tests.push(row);
                    if tests.len() >= limit {
                        return Ok(tests);
                    }
                }
            }
        }
    }

    // A test that reads or assigns the symbol (`app.wsgi_app = ...`) has no call edge.
    let mut test_by_id = conn.prepare(&format!(
        "SELECT {COLUMNS} FROM symbols s WHERE s.symbol_id = ?1 AND {is_test} AND {not_documentation}"
    ))?;
    for site in find_references_for_symbol(
        conn,
        &target_symbol.name,
        "callers",
        MAX_RESULT_LIMIT,
        &target_symbol.symbol_id,
    )
    .unwrap_or_default()
    {
        if site.kind != "member_access (candidate)"
            && !seen_ids.contains(&site.from_symbol_id)
            && let Some(test) = test_by_id
                .query_row(params![site.from_symbol_id], map_symbol)
                .optional()?
        {
            let test = enclosing_test(conn, test)?;
            if seen_ids.insert(test.symbol_id.clone()) {
                tests.push(test);
                if tests.len() >= limit {
                    return Ok(tests);
                }
            }
        }
    }

    // `Flask(...)` runs `Flask.__init__`, so the tests that build the class test its constructor.
    if (target_symbol.kind == "constructor" || target_symbol.name == "__init__")
        && tests.len() < limit
        && let Some(class) = target_symbol
            .parent_symbol_id
            .as_deref()
            .map(|id| get_symbol_by_id(conn, id))
            .transpose()?
            .flatten()
            .filter(|class| class.kind == "class" || class.kind == "struct")
    {
        for test in find_related_tests(conn, &class, limit - tests.len())? {
            if seen_ids.insert(test.symbol_id.clone()) {
                tests.push(test);
            }
        }
        return Ok(tests);
    }

    let name = &target_symbol.name;
    let dunder = name.len() > 4 && name.starts_with("__") && name.ends_with("__");
    let definitions: i64 = conn.query_row(
        "SELECT COUNT(*) FROM symbols WHERE name = ?1 AND kind NOT IN ('import', 'export')",
        params![name],
        |row| row.get(0),
    )?;
    if dunder || definitions > 1 {
        return Ok(tests);
    }

    let remaining = limit - tests.len();
    // SQLite tests these terms in the order written, and the test-path rule costs about 16 times
    // the name match on every row, so the name match comes first.
    let name_sql = format!(
        "SELECT {COLUMNS}
     FROM symbols s
     WHERE (s.name LIKE ?1 ESCAPE '\\' OR s.signature LIKE ?1 ESCAPE '\\')
       AND {is_test}
       AND {not_documentation}
     ORDER BY (s.name LIKE ?1 ESCAPE '\\') DESC, {TEST_ORDER}
     LIMIT ?2"
    );

    if let Ok(mut stmt) = conn.prepare(&name_sql)
        && let Ok(rows) = stmt.query_map(
            params![format!("%{}%", escape_like(name)), (remaining * 2) as i64],
            map_symbol,
        )
    {
        for row in rows.flatten() {
            let row = enclosing_test(conn, row)?;
            if seen_ids.insert(row.symbol_id.clone()) {
                tests.push(row);
                if tests.len() >= limit {
                    return Ok(tests);
                }
            }
        }
    }

    let remaining = limit - tests.len();
    let fts_exists: bool = conn
        .query_row(
            "SELECT 1 FROM sqlite_master WHERE type='table' AND name='symbols_fts'",
            [],
            |_| Ok(true),
        )
        .unwrap_or(false);

    if remaining > 0 && fts_exists {
        let fts_sql = format!(
            "SELECT {COLUMNS}
         FROM symbols_fts
         CROSS JOIN symbols s ON s.rowid = symbols_fts.rowid
         WHERE symbols_fts MATCH ?1 AND {is_test} AND {not_documentation}
         ORDER BY {TEST_ORDER}
         LIMIT ?2"
        );

        let and_q = name_prefix_query(&target_symbol.name);
        if !and_q.is_empty()
            && let Ok(mut stmt) = conn.prepare(&fts_sql)
            && let Ok(rows) = stmt.query_map(params![and_q, (remaining * 2) as i64], map_symbol)
        {
            for row in rows.flatten() {
                let row = enclosing_test(conn, row)?;
                if seen_ids.insert(row.symbol_id.clone()) {
                    tests.push(row);
                    if tests.len() >= limit {
                        break;
                    }
                }
            }
        }
    }

    Ok(tests)
}

/// Renames each row whose parent is a class or another code type to `Owner.name`, so a list of
/// `__init__` rows says which class each one belongs to. For text output only: other tools take
/// the qualified name, and JSON keeps the raw name.
pub fn qualify_members<'a>(
    conn: &Connection,
    symbols: impl IntoIterator<Item = &'a mut Symbol>,
) -> Result<(), QueryError> {
    let mut stmt = conn.prepare(&format!(
        "SELECT name FROM symbols
         WHERE symbol_id = ?1
           AND kind IN ('class', 'struct', 'interface', 'trait', 'enum', 'record', 'object', 'protocol', 'union')
           AND language NOT IN ({})",
        documentation_language_list()
    ))?;
    for symbol in symbols {
        let Some(parent) = symbol.parent_symbol_id.as_deref() else {
            continue;
        };
        if let Some(owner) = stmt
            .query_row(params![parent], |row| row.get::<_, String>(0))
            .optional()?
            && !symbol.name.starts_with(&format!("{owner}."))
        {
            symbol.name = format!("{owner}.{}", symbol.name);
        }
    }
    Ok(())
}

/// Find a specific symbol by name, with an optional path filter for disambiguation.
pub fn get_symbol_by_name(
    conn: &Connection,
    name: &str,
    path_filter: Option<&str>,
) -> Result<Option<Symbol>, QueryError> {
    get_symbol_by_name_internal(conn, name, path_filter, false)
}

/// Find a specific symbol by name, requiring exact path match (used for atomic edits).
pub fn get_symbol_by_name_exact(
    conn: &Connection,
    name: &str,
    exact_path: &str,
) -> Result<Option<Symbol>, QueryError> {
    get_symbol_by_name_internal(conn, name, Some(exact_path), true)
}

/// Retrieve one indexed symbol by its exact current-index identifier.
pub fn get_symbol_by_id(conn: &Connection, symbol_id: &str) -> Result<Option<Symbol>, QueryError> {
    conn.query_row(
        "SELECT symbol_id, file_id, path, language, name, kind, signature, doc_comment,
                visibility, parent_symbol_id, start_line, start_column, end_line, end_column,
                start_byte, end_byte, body_start_line, body_start_column, body_end_line,
                body_end_column, body_start_byte, body_end_byte, body_hash, semantic_group,
                is_test, test_container
         FROM symbols WHERE symbol_id = ?1",
        [symbol_id],
        map_symbol,
    )
    .optional()
    .map_err(QueryError::from)
}

/// Split `Outer::Inner::run` or `Outer.Inner.run` into its ancestor segments, outermost first, and the terminal name.
fn split_qualified_name(name: &str) -> (Vec<&str>, &str) {
    let separator = if name.contains("::") {
        "::"
    } else if name.contains('.') {
        "."
    } else {
        return (Vec::new(), name);
    };
    let mut segments: Vec<&str> = name.split(separator).collect();
    let terminal = segments.pop().unwrap_or(name);
    (segments, terminal)
}

/// Names of the parents of `symbol_id`, innermost first.
fn ancestor_names(conn: &Connection, symbol_id: &str) -> Result<Vec<String>, QueryError> {
    let mut stmt = conn.prepare(
        "SELECT p.symbol_id, p.name FROM symbols s
         JOIN symbols p ON s.parent_symbol_id = p.symbol_id
         WHERE s.symbol_id = ?1",
    )?;
    let mut names = Vec::new();
    let mut current = symbol_id.to_string();
    for _ in 0..32 {
        let mut rows = stmt.query(params![current])?;
        let Some(row) = rows.next()? else { break };
        let parent_id: String = row.get(0)?;
        let parent_name: String = row.get(1)?;
        names.push(parent_name);
        current = parent_id;
    }
    Ok(names)
}

/// A qualified name with several ancestor segments is filtered in Rust after the SQL, so the SQL
/// must not truncate the candidate set the way the 25-row cap does for plain and one-segment names.
const ANCESTOR_WALK_ROW_CAP: i64 = 2000;

fn chain_contains(chain: &[String], wanted: &[&str]) -> bool {
    let mut remaining = chain.iter();
    wanted
        .iter()
        .all(|segment| remaining.any(|name| name == segment))
}

fn name_candidates(
    conn: &Connection,
    name: &str,
    path_filter: Option<&str>,
    exact_path: bool,
) -> Result<Vec<Symbol>, QueryError> {
    let (ancestor_segments, terminal_name) = split_qualified_name(name);
    let parent_name = ancestor_segments.last().copied();

    let sql = "SELECT s.symbol_id, s.file_id, s.path, s.language, s.name, s.kind, s.signature, s.doc_comment,
                s.visibility, s.parent_symbol_id, s.start_line, s.start_column, s.end_line, s.end_column,
                s.start_byte, s.end_byte, s.body_start_line, s.body_start_column, s.body_end_line,
                s.body_end_column, s.body_start_byte, s.body_end_byte, s.body_hash, s.semantic_group,
                s.is_test, s.test_container
         FROM symbols s
         LEFT JOIN symbols p ON s.parent_symbol_id = p.symbol_id
         WHERE (s.name = :name OR (s.name = :term AND (:parent IS NULL OR p.name = :parent)))
           AND (:path IS NULL OR s.path = :path COLLATE NOCASE OR s.path = :path_bs COLLATE NOCASE OR (:exact = 0 AND (s.path LIKE '%/' || :path_like ESCAPE '\\' OR s.path LIKE '%\\\\' || :path_like_bs ESCAPE '\\' OR s.path LIKE :path_like || '/%' ESCAPE '\\' OR s.path LIKE :path_like_bs || '\\\\%' ESCAPE '\\')))
         ORDER BY (s.kind != 'import') DESC,
                  (s.kind IN ('function', 'struct', 'class', 'trait', 'method', 'enum', 'interface', 'type')) DESC,
                  (s.name = :name) DESC,
                  (:path IS NOT NULL AND (s.path = :path COLLATE NOCASE OR s.path = :path_bs COLLATE NOCASE)) DESC,
                  s.is_test ASC
         LIMIT :limit";

    let row_cap: i64 = if ancestor_segments.len() > 1 {
        ANCESTOR_WALK_ROW_CAP
    } else {
        25
    };
    let mut stmt = conn.prepare(sql)?;
    let normalized_path = path_filter.map(|p| p.replace('\\', "/").trim_matches('/').to_string());
    let backslash_path = normalized_path.as_deref().map(|p| p.replace('/', "\\"));
    let path_like = normalized_path.as_deref().map(escape_like);
    let path_like_bs = backslash_path.as_deref().map(escape_like);

    let mut rows = stmt.query(rusqlite::named_params! {
        ":name": name,
        ":term": terminal_name,
        ":parent": parent_name,
        ":path": normalized_path.as_deref(),
        ":path_bs": backslash_path.as_deref(),
        ":path_like": path_like.as_deref(),
        ":path_like_bs": path_like_bs.as_deref(),
        ":exact": if exact_path { 1 } else { 0 },
        ":limit": row_cap,
    })?;

    let mut matches: Vec<Symbol> = Vec::new();
    while let Some(row) = rows.next()? {
        matches.push(map_symbol(row)?);
    }
    drop(rows);

    if ancestor_segments.len() > 1 {
        let required: Vec<&str> = ancestor_segments.iter().rev().copied().collect();
        let mut kept = Vec::with_capacity(matches.len());
        for symbol in matches {
            if symbol.name == name
                || chain_contains(&ancestor_names(conn, &symbol.symbol_id)?, &required)
            {
                kept.push(symbol);
            }
        }
        matches = kept;
    }

    if matches.is_empty() {
        return Ok(Vec::new());
    }

    if matches.len() == 1 {
        return Ok(vec![matches.remove(0)]);
    }

    // Exclude imports if non-import candidates exist
    let candidates: Vec<Symbol> = if matches.iter().any(|s| s.kind != "import") {
        matches.into_iter().filter(|s| s.kind != "import").collect()
    } else {
        matches
    };

    if candidates.len() == 1 {
        return Ok(vec![candidates.into_iter().next().unwrap()]);
    }

    // Check if there's an exact match on full name among candidates
    let exact_name_matches: Vec<_> = candidates
        .iter()
        .filter(|s| s.name == name)
        .cloned()
        .collect();
    if exact_name_matches.len() == 1 {
        return Ok(vec![exact_name_matches.into_iter().next().unwrap()]);
    }

    let definition_candidates = if exact_name_matches.is_empty() {
        &candidates
    } else {
        &exact_name_matches
    };
    let def_matches: Vec<_> = definition_candidates
        .iter()
        .filter(|s| {
            matches!(
                s.kind.as_str(),
                "function"
                    | "struct"
                    | "class"
                    | "trait"
                    | "method"
                    | "enum"
                    | "interface"
                    | "type"
            )
        })
        .cloned()
        .collect();
    if def_matches.len() == 1 {
        return Ok(vec![def_matches.into_iter().next().unwrap()]);
    }

    let active_pool = if !def_matches.is_empty() {
        def_matches
    } else if !exact_name_matches.is_empty() {
        exact_name_matches
    } else {
        candidates
    };

    // If path_filter was given and there's an exact path match
    if let Some(ref p) = normalized_path {
        let exact_path_matches: Vec<_> = active_pool
            .iter()
            .filter(|s| s.path == *p)
            .cloned()
            .collect();
        if exact_path_matches.len() == 1 {
            return Ok(vec![exact_path_matches.into_iter().next().unwrap()]);
        }
    }

    if active_pool.len() == 1 {
        return Ok(vec![active_pool.into_iter().next().unwrap()]);
    }

    let mut outside_tests = active_pool.iter().filter(|s| !is_test_path(&s.path));
    if let (Some(only), None) = (outside_tests.next(), outside_tests.next()) {
        return Ok(vec![only.clone()]);
    }

    let mut outer = outer_definitions(conn, &active_pool)?;
    if outer.len() == 1 {
        return Ok(vec![outer.remove(0)]);
    }

    Ok(active_pool)
}

fn get_symbol_by_name_internal(
    conn: &Connection,
    name: &str,
    path_filter: Option<&str>,
    exact_path: bool,
) -> Result<Option<Symbol>, QueryError> {
    let mut candidates = name_candidates(conn, name, path_filter, exact_path)?;
    match candidates.len() {
        0 | 1 => Ok(candidates.pop()),
        _ => Err(ambiguous_symbol(name, &candidates)),
    }
}

fn ambiguous_symbol(name: &str, candidates: &[Symbol]) -> QueryError {
    let mut candidate_list = String::new();
    for s in candidates {
        candidate_list.push_str(&format!(
            "- {} `{}` in {}:{} (id={})\n",
            crate::formatters::display_kind(s),
            s.name,
            s.path,
            s.start_line,
            s.symbol_id
        ));
    }
    QueryError::AmbiguousSymbol(name.to_string(), candidates.len(), candidate_list)
}

/// The overloads a name selects: one symbol, or every symbol with that name, owner, and file.
/// Any other set of candidates is an `AmbiguousSymbol` error.
pub fn get_symbol_overloads(
    conn: &Connection,
    name: &str,
    path_filter: Option<&str>,
) -> Result<Vec<Symbol>, QueryError> {
    let candidates = name_candidates(conn, name, path_filter, false)?;
    let first = candidates.first();
    let overloads = candidates.iter().all(|s| {
        first.is_some_and(|first| {
            s.name == first.name
                && s.path == first.path
                && s.parent_symbol_id == first.parent_symbol_id
        })
    });
    if overloads {
        Ok(candidates)
    } else {
        Err(ambiguous_symbol(name, &candidates))
    }
}

/// The candidates left after dropping an `export` row that repeats a definition on its line and,
/// when anything else remains, the locals of a function or method.
fn outer_definitions(conn: &Connection, pool: &[Symbol]) -> Result<Vec<Symbol>, QueryError> {
    let definitions: Vec<&Symbol> = pool
        .iter()
        .filter(|s| {
            s.kind != "export"
                || !pool.iter().any(|other| {
                    other.kind != "export"
                        && other.path == s.path
                        && other.start_line == s.start_line
                })
        })
        .collect();
    let mut parent_kind = conn.prepare("SELECT kind FROM symbols WHERE symbol_id = ?1")?;
    let mut outer = Vec::new();
    for symbol in &definitions {
        let kind = match &symbol.parent_symbol_id {
            Some(parent) => parent_kind
                .query_row([parent], |row| row.get::<_, String>(0))
                .optional()?,
            None => None,
        };
        if !matches!(kind.as_deref(), Some("function" | "method" | "constructor")) {
            outer.push((*symbol).clone());
        }
    }
    Ok(if outer.is_empty() {
        definitions.into_iter().cloned().collect()
    } else {
        outer
    })
}

/// The name of the repository this index was built for, for not-found messages.
pub fn workspace_name(conn: &Connection) -> String {
    conn.query_row(
        "SELECT value FROM artifact_metadata WHERE key = 'root_path'",
        [],
        |row| row.get::<_, String>(0),
    )
    .ok()
    .map(|root| root.replace('\\', "/"))
    .and_then(|root| {
        root.trim_end_matches('/')
            .rsplit('/')
            .next()
            .filter(|name| !name.is_empty())
            .map(str::to_string)
    })
    .unwrap_or_else(|| "this workspace".to_string())
}

fn edit_distance(left: &str, right: &str) -> usize {
    let left: Vec<char> = left.chars().collect();
    let right: Vec<char> = right.chars().collect();
    let mut previous: Vec<usize> = (0..=right.len()).collect();
    let mut current = vec![0usize; right.len() + 1];
    for (i, a) in left.iter().enumerate() {
        current[0] = i + 1;
        for (j, b) in right.iter().enumerate() {
            let substitution = previous[j] + usize::from(a != b);
            current[j + 1] = substitution.min(previous[j + 1] + 1).min(current[j] + 1);
        }
        std::mem::swap(&mut previous, &mut current);
    }
    previous[right.len()]
}

/// Names within a small edit distance of `name`, recalled through the name trigram index.
fn near_names_by_trigram(
    conn: &Connection,
    name: &str,
    path_filter: Option<&str>,
) -> Result<Vec<Symbol>, QueryError> {
    if !has_table(conn, "symbol_names_tri") {
        return Ok(Vec::new());
    }
    let lower = name.to_lowercase();
    let chars: Vec<char> = lower.chars().collect();
    let chunks: Vec<String> = chars
        .windows(3)
        .map(|window| window.iter().collect::<String>())
        .filter(|chunk| chunk.chars().all(|c| c.is_alphanumeric() || c == '_'))
        .collect();
    if chunks.is_empty() {
        return Ok(Vec::new());
    }
    let match_clause = chunks
        .iter()
        .map(|chunk| format!("\"{chunk}\""))
        .collect::<Vec<_>>()
        .join(" OR ");

    let normalized_path = path_filter.map(|p| {
        p.replace('\\', "/")
            .trim_start_matches("./")
            .trim_matches('/')
            .to_string()
    });
    let escaped_path = normalized_path.as_deref().map(escape_like);
    let kind_val: Option<&str> = None;
    let columns = candidate_columns(conn);
    let filters = candidate_filters(false, false);
    let sql = format!(
        "SELECT {columns} FROM symbol_names_tri
         CROSS JOIN symbols s ON s.rowid = symbol_names_tri.rowid
         WHERE symbol_names_tri MATCH :match {filters}
         ORDER BY bm25(symbol_names_tri) ASC, length(s.name) ASC, s.path ASC LIMIT 20"
    );
    let rows = conn
        .prepare(&sql)?
        .query_map(
            rusqlite::named_params! {
                ":match": match_clause,
                ":kind": kind_val,
                ":path": normalized_path.as_deref(),
                ":path_like": escaped_path.as_deref(),
            },
            map_symbol,
        )?
        .collect::<Result<Vec<_>, _>>()?;

    let budget = (chars.len() / 4).max(2);
    let mut scored: Vec<(usize, Symbol)> = rows
        .into_iter()
        .map(|symbol| (edit_distance(&lower, &symbol.name.to_lowercase()), symbol))
        .filter(|(distance, _)| *distance <= budget)
        .collect();
    scored.sort_by_key(|(distance, symbol)| (*distance, symbol.name.chars().count()));
    Ok(scored
        .into_iter()
        .map(|(_, symbol)| symbol)
        .take(3)
        .collect())
}

/// Up to three indexed symbols whose names are close to `name`. Never fails.
pub fn suggest_symbol_names(
    conn: &Connection,
    name: &str,
    path_filter: Option<&str>,
) -> Vec<Symbol> {
    let (ancestors, terminal) = split_qualified_name(name);
    let mut found =
        search_symbols_scoped(conn, terminal, None, path_filter, false, 3).unwrap_or_default();
    if found.is_empty() && terminal.chars().count() >= 4 {
        found = near_names_by_trigram(conn, terminal, path_filter).unwrap_or_default();
    }
    if let Some(parent) = ancestors.last().copied() {
        let mut ranked: Vec<(bool, Symbol)> = found
            .into_iter()
            .map(|symbol| {
                let shares_parent = ancestor_names(conn, &symbol.symbol_id)
                    .unwrap_or_default()
                    .first()
                    .is_some_and(|found_parent| found_parent == parent);
                (!shares_parent, symbol)
            })
            .collect();
        ranked.sort_by_key(|(demoted, _)| *demoted);
        found = ranked.into_iter().map(|(_, symbol)| symbol).collect();
    }
    found.truncate(3);
    found
}

/// Up to three indexed file paths close to `rel_path`. Never fails.
pub fn suggest_file_paths(conn: &Connection, rel_path: &str) -> Vec<String> {
    let wanted = rel_path.replace('\\', "/");
    let wanted = wanted.trim_start_matches("./").trim_matches('/');
    let basename = wanted.rsplit('/').next().unwrap_or(wanted).to_lowercase();
    if basename.is_empty() {
        return Vec::new();
    }
    let stem = basename.split('.').next().unwrap_or(&basename).to_string();
    let segments: Vec<&str> = wanted.split('/').collect();
    let tail = if segments.len() >= 2 {
        segments[segments.len() - 2..].join("/").to_lowercase()
    } else {
        basename.clone()
    };

    let path_expr = "replace(files.path, '\\', '/')";
    let file_name = format!(
        "lower(replace({path_expr}, rtrim({path_expr}, replace({path_expr}, '/', '')), ''))"
    );
    let rules = [
        (format!("{file_name} = :value"), basename.clone()),
        (
            format!("{file_name} LIKE :value ESCAPE '\\'"),
            format!("{}%", escape_like(&stem)),
        ),
        (
            format!("lower({path_expr}) LIKE :value ESCAPE '\\'"),
            format!("%{}", escape_like(&tail)),
        ),
    ];

    for (predicate, value) in rules {
        let sql = format!(
            "SELECT {path_expr} FROM files WHERE {predicate}
             ORDER BY length(files.path) ASC, files.path ASC LIMIT 3"
        );
        let found: Vec<String> = conn
            .prepare(&sql)
            .and_then(|mut stmt| {
                stmt.query_map(rusqlite::named_params! { ":value": value }, |row| {
                    row.get::<_, String>(0)
                })?
                .collect()
            })
            .unwrap_or_default();
        if !found.is_empty() {
            return found;
        }
    }

    let max_edits = (basename.chars().count() / 4).clamp(1, 2);
    let mut near: Vec<(usize, String)> = Vec::new();
    let _ = conn
        .prepare(&format!("SELECT {path_expr} FROM files"))
        .and_then(|mut stmt| {
            for path in stmt.query_map([], |row| row.get::<_, String>(0))?.flatten() {
                let name = path.rsplit('/').next().unwrap_or(&path).to_lowercase();
                let edits = edit_distance(&name, &basename);
                if edits <= max_edits {
                    near.push((edits, path));
                }
            }
            Ok(())
        });
    near.sort();
    near.into_iter().take(3).map(|(_, path)| path).collect()
}

/// The workspace name and the recovery hint for a symbol that is not indexed.
pub fn symbol_not_found_parts(
    conn: &Connection,
    name: &str,
    path_filter: Option<&str>,
) -> (String, String) {
    let list = |candidates: &[Symbol]| {
        candidates
            .iter()
            .map(|s| format!("  - {} `{}` ({}:{})", s.kind, s.name, s.path, s.start_line))
            .collect::<Vec<_>>()
            .join("\n")
    };
    let terminal = split_qualified_name(name).1;
    if let Some(path) = path_filter {
        let elsewhere: Vec<Symbol> = suggest_symbol_names(conn, name, None)
            .into_iter()
            .filter(|s| s.name == terminal)
            .collect();
        if !elsewhere.is_empty() {
            let hint = format!(
                "`{path}` does not define it; it is defined in:\n{}",
                list(&elsewhere)
            );
            return (workspace_name(conn), hint);
        }
    }
    let candidates = suggest_symbol_names(conn, name, path_filter);
    let hint = if candidates.is_empty() {
        "No similar name is indexed; check the workspace and spelling.".to_string()
    } else {
        format!("Did you mean one of:\n{}", list(&candidates))
    };
    (workspace_name(conn), hint)
}

/// The workspace name and the recovery hint for a file that is not indexed.
pub fn file_not_found_parts(conn: &Connection, rel_path: &str) -> (String, String) {
    let candidates = suggest_file_paths(conn, rel_path);
    let hint = if candidates.is_empty() {
        "No similar path is indexed; check the workspace and spelling.".to_string()
    } else {
        let list = candidates
            .iter()
            .map(|path| format!("  - {path}"))
            .collect::<Vec<_>>()
            .join("\n");
        format!("Did you mean one of:\n{list}")
    };
    (workspace_name(conn), hint)
}

/// Find callers or callees of a symbol (filters unresolved external stdlib/runtime primitives by default).
pub fn find_references(
    conn: &Connection,
    symbol_name: &str,
    direction: &str,
    limit: usize,
) -> Result<Vec<ReferenceSite>, QueryError> {
    find_references_ext(conn, symbol_name, direction, limit, false)
}

/// Find callers or callees with option to include external runtime/stdlib primitives.
pub fn find_references_ext(
    conn: &Connection,
    symbol_name: &str,
    direction: &str,
    limit: usize,
    include_external: bool,
) -> Result<Vec<ReferenceSite>, QueryError> {
    find_references_scoped(conn, symbol_name, direction, limit, include_external, None)
}

/// Find callers or callees with optional file path disambiguation filter and external symbols toggle.
pub fn find_references_scoped(
    conn: &Connection,
    symbol_name: &str,
    direction: &str,
    limit: usize,
    include_external: bool,
    path_filter: Option<&str>,
) -> Result<Vec<ReferenceSite>, QueryError> {
    validate_result_limit(limit)?;
    if direction != "callers" && direction != "callees" {
        return Err(QueryError::InvalidDirection(direction.to_string()));
    }

    match get_symbol_by_name(conn, symbol_name, path_filter)? {
        Some(target) => find_references_internal(
            conn,
            &target.name,
            direction,
            limit,
            Some(&target.symbol_id),
            include_external,
        ),
        None => {
            let (workspace, hint) = symbol_not_found_parts(conn, symbol_name, path_filter);
            Err(QueryError::SymbolNotFound {
                name: symbol_name.to_string(),
                workspace,
                hint,
            })
        }
    }
}

pub fn find_references_for_symbol(
    conn: &Connection,
    symbol_name: &str,
    direction: &str,
    limit: usize,
    symbol_id: &str,
) -> Result<Vec<ReferenceSite>, QueryError> {
    find_references_for_symbol_ext(conn, symbol_name, direction, limit, symbol_id, false)
}

pub fn find_references_for_symbol_ext(
    conn: &Connection,
    symbol_name: &str,
    direction: &str,
    limit: usize,
    symbol_id: &str,
    include_external: bool,
) -> Result<Vec<ReferenceSite>, QueryError> {
    find_references_internal(
        conn,
        symbol_name,
        direction,
        limit,
        Some(symbol_id),
        include_external,
    )
}

/// SQL expression ranking a candidate path against the call site `p.path`:
/// 2 for the same file, 1 for the same directory, 0 otherwise.
fn call_site_proximity(candidate_path: &str) -> String {
    path_proximity(candidate_path, "p.path")
}

/// SQL expression ranking `candidate_path` against `site_path`: 2 for the same file, 1 for the
/// same directory, 0 otherwise.
pub(crate) fn path_proximity(candidate_path: &str, site_path: &str) -> String {
    let normalized = format!("replace({candidate_path}, '\\', '/')");
    let call_site = format!("replace({site_path}, '\\', '/')");
    format!(
        "CASE WHEN {normalized} = {call_site} THEN 2
              WHEN rtrim({normalized}, replace({normalized}, '/', '')) = rtrim({call_site}, replace({call_site}, '/', '')) THEN 1
              ELSE 0 END"
    )
}

/// True when the import source `value` is a relative file path (`../core/util.js`).
fn is_relative_import_path(value: &str) -> String {
    format!("({value} LIKE '.%' AND instr({value}, '/') > 0)")
}

/// True when `{target}` is the file the relative import source `value` names, resolved from the
/// folder of the call site `p.path`: `../core/util.js` from `src/v4/locales/ru.ts` names
/// `src/v4/core/util.ts`, `.js`, `.tsx`, and the other script files, or that folder's `index`.
fn relative_import_matches(value: &str, target: &str) -> String {
    let stem = format!(
        "CASE WHEN {value} GLOB '*.[jt]s' THEN substr({value}, 1, length({value}) - 3)
              WHEN {value} GLOB '*.[mc][jt]s' OR {value} GLOB '*.[jt]sx'
                   THEN substr({value}, 1, length({value}) - 4)
              ELSE {value} END"
    );
    let caller = "replace(p.path, '\\', '/')";
    let parent_dir = "rtrim(substr(hop.dir, 1, length(hop.dir) - 1), replace(substr(hop.dir, 1, length(hop.dir) - 1), '/', ''))";
    let files = ["ts", "tsx", "js", "jsx", "mjs", "cjs", "mts", "cts", "d.ts"]
        .iter()
        .flat_map(|ext| {
            [
                format!("hop.dir || hop.rest || '.{ext}'"),
                format!("hop.dir || hop.rest || '/index.{ext}'"),
            ]
        })
        .collect::<Vec<_>>()
        .join(", ");
    format!(
        "EXISTS (
            WITH RECURSIVE hop(dir, rest) AS (
                SELECT rtrim({caller}, replace({caller}, '/', '')),
                       CASE WHEN ({stem}) LIKE './%' THEN substr(({stem}), 3) ELSE ({stem}) END
                UNION ALL
                SELECT {parent_dir}, substr(hop.rest, 4) FROM hop WHERE hop.rest LIKE '../%'
            )
            SELECT 1 FROM hop
            WHERE hop.rest NOT LIKE '../%' AND hop.rest != ''
              AND replace({target}.path, '\\', '/') IN ({files})
        )"
    )
}

/// Languages whose files import each other's names, so a name-only reference may cross them.
const JS_FAMILY_LANGUAGES: &str = "'javascript', 'typescript', 'tsx', 'jsx', 'vue'";
/// SQL that holds when a name-only identifier in `identifier` may refer to `target` across
/// languages: the same language, the JavaScript family, Razor and C#, or QML to C++.
fn identifier_language_matches(identifier: &str, target: &str) -> String {
    format!(
        "({target}.language = {identifier}.language
          OR ({identifier}.language IN ({JS_FAMILY_LANGUAGES})
              AND {target}.language IN ({JS_FAMILY_LANGUAGES}))
          OR ({identifier}.language IN ('razor', 'csharp')
              AND {target}.language IN ('razor', 'csharp'))
          OR ({identifier}.language = 'qml' AND {target}.language = 'cpp'))"
    )
}

const NESTED_TYPE_KINDS: &str =
    "'class', 'struct', 'interface', 'enum', 'record', 'trait', 'protocol'";

/// Languages where a bare call inside a class reaches the members of that class and its bases,
/// as a call through `this` or `self` does.
const IMPLICIT_RECEIVER_LANGUAGES: &[&str] = &[
    "java", "csharp", "kotlin", "swift", "cpp", "scala", "dart", "ruby", "vbnet",
];

fn receiver_builder_calls(conn: &Connection) -> String {
    // Same-file calls resolve into `relationships`; cross-file calls remain pending. Keep the
    // resolved symbol ID so same-named builders in different lexical scopes do not collapse.
    let resolved_calls = if has_column(conn, "relationships", "from_symbol_id")
        && has_column(conn, "relationships", "to_symbol_id")
        && has_column(conn, "relationships", "kind")
        && has_column(conn, "relationships", "path")
        && has_column(conn, "relationships", "start_line")
    {
        "UNION
         SELECT resolved_builder.name, assign_call.path, resolved_builder.symbol_id
         FROM symbols assigned
         JOIN relationships assign_call
           ON assign_call.from_symbol_id = p.from_symbol_id
          AND assign_call.start_line = assigned.start_line
          AND +assign_call.kind = 'calls'
         JOIN symbols resolved_builder ON resolved_builder.symbol_id = assign_call.to_symbol_id
         WHERE assigned.parent_symbol_id = p.from_symbol_id
           AND +assigned.name = p.target_receiver
           AND +assigned.kind = 'variable'
           AND assigned.start_line <= p.start_line"
    } else {
        ""
    };
    let fixture_calls = if has_column(conn, "symbols", "test_lifecycle") {
        "UNION
         SELECT fixture_call.target_terminal_name, fixture.path, NULL
         FROM symbols fixture
         JOIN pending_relationships fixture_call ON fixture_call.from_symbol_id = fixture.symbol_id
         WHERE fixture.name = p.target_receiver
           AND fixture.kind IN ('function', 'method')
           AND +fixture.test_lifecycle = 1
           AND EXISTS (
               SELECT 1 FROM symbols parameter
               WHERE parameter.parent_symbol_id = p.from_symbol_id
                 AND +parameter.name = p.target_receiver
                 AND (parameter.kind = 'parameter'
                      OR json_valid(parameter.metadata_json)
                         AND json_extract(parameter.metadata_json, '$.role') = 'parameter')
           )
           AND (fixture.path = p.path
                OR fixture.path LIKE '%conftest.py'
                   AND substr(p.path, 1, length(fixture.path) - 11)
                       = substr(fixture.path, 1, length(fixture.path) - 11))"
    } else {
        ""
    };
    // Unary `+` keeps SQLite on the parent index for common receiver names.
    format!(
        "SELECT assign_call.target_terminal_name, assign_call.path, NULL
         FROM symbols assigned
         JOIN pending_relationships assign_call
           ON assign_call.from_symbol_id = p.from_symbol_id
          AND assign_call.start_line = assigned.start_line
          AND +assign_call.kind = 'calls'
         WHERE assigned.parent_symbol_id = p.from_symbol_id
           AND +assigned.name = p.target_receiver
           AND +assigned.kind = 'variable'
           AND assigned.start_line <= p.start_line
         {resolved_calls}
         {fixture_calls}"
    )
}

fn builder_return_type_sql(builder: &str) -> String {
    format!(
        "CASE WHEN json_valid({builder}.metadata_json)
              THEN COALESCE(NULLIF(json_extract({builder}.metadata_json, '$.returnType'), ''),
                            json_extract({builder}.metadata_json, '$.inferredReturnType')) END"
    )
}

fn receiver_definition_scope(target: &str, origin: &str) -> String {
    let source = "json_extract(binding.metadata_json, '$.source')";
    let relative = is_relative_import_path(source);
    let relative_match = relative_import_matches(source, target).replace("p.path", origin);
    let caller = format!("replace({origin}, '\\', '/')");
    let parent_dir = "rtrim(substr(hop.dir, 1, length(hop.dir) - 1), replace(substr(hop.dir, 1, length(hop.dir) - 1), '/', ''))";
    format!(
        "({target}.path = {origin} OR NOT EXISTS (
            SELECT 1 FROM symbols local
            WHERE local.name = {target}.name AND local.path = {origin}
              AND local.kind = {target}.kind
         )) AND NOT EXISTS (
            SELECT 1 FROM symbols binding
            WHERE binding.name = {target}.name AND binding.path = {origin}
              AND binding.kind = 'import' AND json_valid(binding.metadata_json)
              AND NOT ({target}.path = {origin} AND binding.start_line < {target}.start_line)
              AND COALESCE({source}, '') != ''
              AND NOT CASE WHEN {relative} THEN {relative_match}
                  WHEN {source} LIKE '.%' THEN EXISTS (
                      WITH RECURSIVE hop(dir, rest) AS (
                          SELECT rtrim({caller}, replace({caller}, '/', '')), substr({source}, 2)
                          UNION ALL
                          SELECT {parent_dir}, substr(hop.rest, 2) FROM hop WHERE hop.rest LIKE '.%'
                      )
                      SELECT 1 FROM hop WHERE hop.rest NOT LIKE '.%'
                        AND (replace({target}.path, '\\', '/') = hop.dir || replace(hop.rest, '.', '/') || '.py'
                             OR replace({target}.path, '\\', '/') LIKE
                                rtrim(hop.dir || replace(hop.rest, '.', '/'), '/') || '/%')
                  )
                  ELSE instr('/' || replace({target}.path, '\\', '/'), '/' || replace({source}, '.', '/') || '.') > 0
                    OR instr('/' || replace({target}.path, '\\', '/'), '/' || replace({source}, '.', '/') || '/') > 0
                  END
         )"
    )
}

fn member_receiver_match(conn: &Connection) -> String {
    let callable_member = "p.kind = 'member_access'
                    AND CASE WHEN json_valid(p.metadata_json)
                             THEN json_extract(p.metadata_json, '$.role') END IS NOT 'signal_handler'
                    AND EXISTS (
                        SELECT 1 FROM symbols selected
                        JOIN symbols owner ON owner.symbol_id = selected.parent_symbol_id
                        WHERE selected.symbol_id = ?3
                          AND selected.kind IN ('method', 'function', 'constructor')
                          AND owner.kind IN ('class', 'struct', 'interface', 'trait', 'enum', 'protocol')
                    )";
    // A member access without metadata has no receiver evidence: julie leaves `super().x` and
    // HTML and CSS names without it. It stays a candidate for a class method and never
    // reaches another target, where the name alone matches unrelated sites.
    let unproven_member = "p.kind = 'member_access' AND NOT p.has_metadata";
    let documentation_member = format!(
        "{unproven_member} AND p.language IN ({})",
        documentation_language_list()
    );
    if !has_pending_namespace_column(conn) || !has_column(conn, "symbols", "metadata_json") {
        return format!(
            "CASE WHEN {documentation_member} THEN -1 WHEN {callable_member} THEN 0
                  WHEN {unproven_member} THEN -1 ELSE 1 END"
        );
    }
    let matched = pending_target_predicate(conn, "candidate", "candidate_parent");
    let builders = receiver_builder_calls(conn);
    let builder_return_type = builder_return_type_sql("builder");
    let class_scope = receiver_definition_scope("receiver_class", "builder_call.path");
    let builder_scope = receiver_definition_scope("builder", "builder_call.path");
    let returned_scope = receiver_definition_scope("returned_class", "builder.path");
    format!(
        "CASE WHEN {documentation_member} THEN -1
              WHEN NOT ({callable_member}) THEN CASE WHEN {unproven_member} THEN -1 ELSE 1 END
              WHEN {unproven_member} THEN 0
              WHEN (SELECT COUNT(*) FROM symbols binding
                    WHERE binding.parent_symbol_id = p.from_symbol_id
                      AND +binding.name = p.target_receiver
                      AND +binding.kind IN ('variable', 'parameter')
                      AND binding.start_line <= p.start_line) > 1 THEN 0
              WHEN EXISTS (
                  WITH builder_call(name, path, resolved_symbol_id) AS ({builders})
                  SELECT 1 FROM builder_call
                  WHERE (SELECT COUNT(*) FROM builder_call) > 1
                    AND builder_call.resolved_symbol_id IS NULL
                    AND NOT EXISTS (
                        SELECT 1 FROM symbols receiver_class
                        WHERE receiver_class.name = builder_call.name AND receiver_class.kind = 'class'
                          AND (receiver_class.symbol_id = builder_call.resolved_symbol_id
                               OR builder_call.resolved_symbol_id IS NULL AND {class_scope})
                    )
                    AND NOT EXISTS (
                        SELECT 1 FROM symbols builder
                        JOIN symbols returned_class ON returned_class.name =
                            {builder_return_type}
                        WHERE builder.name = builder_call.name AND builder.kind IN ('function', 'method')
                          AND (builder.symbol_id = builder_call.resolved_symbol_id
                               OR CASE WHEN builder_call.resolved_symbol_id IS NULL
                                            AND EXISTS (
                                                SELECT 1 FROM symbols returned_class_probe
                                                WHERE returned_class_probe.name = {builder_return_type}
                                                  AND returned_class_probe.kind = 'class'
                                            )
                                       THEN {builder_scope} ELSE 0 END)
                          AND returned_class.kind = 'class'
                          AND {returned_scope}
                    )
              ) THEN 0
              ELSE (
                  SELECT CASE WHEN COUNT(*) > 1 THEN 0
                              WHEN MAX(candidate.symbol_id = ?3) = 1 THEN 1
                              WHEN COUNT(*) = 1 THEN -1
                              WHEN EXISTS (
                                  WITH builder_call(name, path, resolved_symbol_id) AS ({builders})
                                  SELECT 1 FROM builder_call
                                  JOIN symbols imported ON imported.name = builder_call.name
                                    AND imported.path = builder_call.path AND imported.kind = 'import'
                                  JOIN symbols selected ON selected.symbol_id = ?3
                                  JOIN pending_relationships base ON base.from_symbol_id = selected.parent_symbol_id
                                    AND base.kind = 'extends' AND base.target_terminal_name = builder_call.name
                                  JOIN symbols base_import ON base_import.name = base.target_terminal_name
                                    AND base_import.path = base.path AND base_import.kind = 'import'
                                  WHERE json_valid(imported.metadata_json) AND json_valid(base_import.metadata_json)
                                    AND json_extract(imported.metadata_json, '$.source')
                                        = json_extract(base_import.metadata_json, '$.source')
                              ) THEN -1
                              ELSE 0 END
                  FROM symbols candidate
                  JOIN symbols candidate_parent ON candidate_parent.symbol_id = candidate.parent_symbol_id
                  WHERE candidate.name = p.name
                    AND candidate.kind IN ('method', 'function', 'constructor')
                    AND candidate_parent.kind IN ('class', 'struct', 'interface', 'trait', 'enum', 'protocol')
                    AND {matched}
              ) END"
    )
}

/// SQL that holds when the caller's file `p.path` binds `binding_name` through an import whose
/// source module, or imported name, names a folder or file on the path of `target`.
pub(crate) fn module_import_binding_sql(target: &str, binding_name: &str) -> String {
    let target_path = format!("('/' || replace({target}.path, '\\', '/'))");
    let dotted = "replace(module_path.value, '/', '.')";
    let segment = format!("replace({dotted}, rtrim({dotted}, replace({dotted}, '.', '')), '')");
    let relative = is_relative_import_path("import_source.value");
    let relative_match = relative_import_matches("import_source.value", target);
    format!(
        "EXISTS (
                        SELECT 1 FROM symbols module_import
                        CROSS JOIN json_each(json_array(
                            json_extract(module_import.metadata_json, '$.source')
                        )) import_source
                        CROSS JOIN json_each(json_array(
                            json_extract(module_import.metadata_json, '$.importedName'),
                            import_source.value
                        )) module_path
                        WHERE module_import.kind = 'import'
                          AND module_import.path = p.path
                          AND module_import.name = {binding_name}
                          AND json_valid(module_import.metadata_json)
                          AND CASE WHEN {relative} THEN {relative_match}
                              ELSE {segment} != ''
                                   AND ({target_path} LIKE '%/' || {segment} || '/%'
                                        OR {target_path} LIKE '%/' || {segment} || '.%')
                              END
                    )"
    )
}

/// SQL that holds when the receiver `p.target_receiver` is a parameter or variable of the caller
/// that a call builds as the class owning `target`, directly or through a builder's return type
/// or a fixture, and no nearer class in the chain defines the name. `None` without metadata.
pub(crate) fn built_receiver_sql(conn: &Connection, target: &str) -> Option<String> {
    if !has_column(conn, "symbols", "metadata_json") {
        return None;
    }
    let builder_calls = receiver_builder_calls(conn);
    let builder_return_type = builder_return_type_sql("builder");
    let builder_scope = receiver_definition_scope("builder", "builder_call.path");
    let returned_class_scope = receiver_definition_scope("returned_class", "builder.path");
    let built_scope = receiver_definition_scope("built_class", "built_name.path");
    // CASE prevents scope checks for builders that cannot return an indexed class.
    Some(format!(
        "CASE WHEN EXISTS (
                        SELECT 1 FROM symbols taken
                        WHERE taken.parent_symbol_id = p.from_symbol_id
                          AND taken.name = p.target_receiver
                          AND taken.kind IN ('parameter', 'variable')
                    ) THEN EXISTS (
                        WITH RECURSIVE builder_call(name, path, resolved_symbol_id) AS ({builder_calls}),
                        built_name(name, path, resolved_symbol_id) AS (
                            SELECT name, path, resolved_symbol_id FROM builder_call
                            UNION
                            SELECT returned_class.name, builder.path, returned_class.symbol_id
                            FROM builder_call
                            JOIN symbols builder ON builder.name = builder_call.name
                              AND builder.kind IN ('function', 'method')
                              AND (builder.symbol_id = builder_call.resolved_symbol_id
                                   OR CASE WHEN builder_call.resolved_symbol_id IS NULL
                                                AND EXISTS (
                                                    SELECT 1 FROM symbols returned_class_probe
                                                    WHERE returned_class_probe.name = {builder_return_type}
                                                      AND returned_class_probe.kind = 'class'
                                                )
                                           THEN {builder_scope} ELSE 0 END)
                            JOIN symbols returned_class ON returned_class.name = {builder_return_type}
                              AND returned_class.kind = 'class' AND {returned_class_scope}
                        ),
                        built(symbol_id, depth) AS (
                            SELECT built_class.symbol_id, 0
                            FROM built_name
                            CROSS JOIN symbols built_class ON built_class.name = built_name.name
                            WHERE built_class.kind = 'class' AND {built_scope}
                              AND (built_name.resolved_symbol_id IS NULL
                                   OR built_class.symbol_id = built_name.resolved_symbol_id)
                            UNION
                            SELECT r.to_symbol_id, built.depth + 1
                            FROM relationships r JOIN built ON r.from_symbol_id = built.symbol_id
                            WHERE +r.kind = 'extends' AND built.depth < 8
                            UNION
                            SELECT c.symbol_id, built.depth + 1
                            FROM pending_relationships pe
                            JOIN built ON pe.from_symbol_id = built.symbol_id
                            JOIN symbols c ON c.name = pe.target_terminal_name
                            WHERE +pe.kind = 'extends' AND built.depth < 8
                              AND c.kind IN ('class', 'interface', 'struct', 'trait', 'protocol')
                        )
                        SELECT 1 FROM built WHERE built.symbol_id = {target}.parent_symbol_id
                          AND NOT EXISTS (
                              SELECT 1 FROM built closer
                              JOIN symbols nearer ON nearer.parent_symbol_id = closer.symbol_id
                              WHERE nearer.name = {target}.name
                                AND nearer.kind IN ('method', 'function', 'constructor', 'property')
                                AND closer.depth < built.depth
                          )
                    ) ELSE 0 END"
    ))
}

/// SQL that holds when a variable or parameter named `p.target_receiver` in the caller's file has
/// a type fact naming the parent of `target`, and no other variable of the caller shadows it.
pub(crate) fn receiver_type_sql(conn: &Connection, parent: &str) -> String {
    // Type facts retain only the terminal type name; constrain them to the imported definition.
    let receiver_type_scope = if has_column(conn, "symbols", "metadata_json") {
        receiver_definition_scope(parent, "p.path")
    } else {
        "1 = 1".to_string()
    };
    format!(
        "EXISTS (
                        SELECT 1 FROM symbols receiver
                        JOIN type_facts receiver_type ON receiver_type.symbol_id = receiver.symbol_id
                        WHERE receiver.name = p.target_receiver
                          AND receiver.path = p.path
                          AND receiver_type.resolved_type = {parent}.name
                          AND NOT EXISTS (
                              SELECT 1 FROM symbols shadow
                              WHERE shadow.name = receiver.name
                                AND shadow.path = receiver.path
                                AND shadow.parent_symbol_id = p.from_symbol_id
                                AND shadow.kind IN ('variable', 'parameter')
                                AND shadow.symbol_id != receiver.symbol_id
                          )
                    ) AND {receiver_type_scope}"
    )
}

/// SQL that holds when `target` is a member of the class chain of the caller `s_from`: its own
/// class when `own_class_rule` holds, or a base class named by an `extends` row, where the nearest
/// class that defines the name wins. julie resolves most `this` calls to the caller's own class,
/// but not one inside an ordinary JavaScript function, so a `this`, `self`, or `cls` call also
/// matches a sibling member. A base class that a pending `extends` row names counts unless a class
/// of that name is closer.
pub(crate) fn class_chain_sql(target: &str, own_class_rule: &str) -> String {
    let rival_rank = path_proximity("rival.path", "pe.path");
    let base_rank = path_proximity("c.path", "pe.path");
    let inherited_write = inherited_attribute_write("nearer");
    format!(
        "EXISTS (
                            WITH RECURSIVE base(symbol_id, depth) AS (
                                SELECT s_from.parent_symbol_id, 0
                                UNION
                                SELECT r.to_symbol_id, base.depth + 1
                                FROM relationships r JOIN base ON r.from_symbol_id = base.symbol_id
                                WHERE +r.kind = 'extends' AND base.depth < 8
                                UNION
                                SELECT c.symbol_id, base.depth + 1
                                FROM pending_relationships pe
                                JOIN base ON pe.from_symbol_id = base.symbol_id
                                JOIN symbols c ON c.name = pe.target_terminal_name
                                WHERE +pe.kind = 'extends' AND base.depth < 8
                                  AND c.kind IN ('class', 'interface', 'struct', 'trait', 'protocol')
                                  AND NOT EXISTS (
                                      SELECT 1 FROM symbols rival
                                      WHERE rival.name = c.name
                                        AND rival.kind IN ('class', 'interface', 'struct', 'trait', 'protocol')
                                        AND {rival_rank} > {base_rank})
                            )
                            SELECT 1 FROM base
                            WHERE base.symbol_id = {target}.parent_symbol_id
                              AND (base.depth > 0 OR {own_class_rule})
                              AND NOT EXISTS (
                                  SELECT 1 FROM base closer
                                  JOIN symbols nearer ON nearer.parent_symbol_id = closer.symbol_id
                                  WHERE nearer.name = {target}.name
                                    AND nearer.kind IN ('method', 'function', 'constructor', 'property')
                                    AND closer.depth < base.depth
                                    AND (closer.depth > 0 OR p.target_receiver IS NOT 'super')
                                    AND NOT {inherited_write}
                              )
                        )"
    )
}

/// The quoted, comma-separated [`IMPLICIT_RECEIVER_LANGUAGES`] for an SQL `IN` list.
pub(crate) fn implicit_receiver_language_list() -> String {
    IMPLICIT_RECEIVER_LANGUAGES
        .iter()
        .map(|language| format!("'{language}'"))
        .collect::<Vec<_>>()
        .join(", ")
}

/// SQL that holds when `target` is defined in a scope that encloses the caller's parent, where the
/// nearest scope that defines or imports the name wins. An enclosing class body contributes only
/// its nested types, and only where a bare call reaches members; its methods are left to the
/// class-chain rule, because julie parses macro calls in C++ test files as methods of made-up
/// classes.
pub(crate) fn enclosing_scope_sql(target: &str) -> String {
    let implicit_receiver_languages = implicit_receiver_language_list();
    format!(
        "EXISTS (
            WITH RECURSIVE lexical(symbol_id, depth) AS (
                SELECT s_from.parent_symbol_id, 1
                UNION ALL
                SELECT outer_scope.parent_symbol_id, lexical.depth + 1
                FROM symbols outer_scope JOIN lexical ON outer_scope.symbol_id = lexical.symbol_id
                WHERE outer_scope.parent_symbol_id IS NOT NULL AND lexical.depth < 32
            )
            SELECT 1 FROM lexical JOIN symbols scope ON scope.symbol_id = lexical.symbol_id
            WHERE lexical.symbol_id = {target}.parent_symbol_id
              AND lexical.depth > 1
              AND (scope.kind NOT IN ({MEMBER_LIST_OWNER_KINDS})
                   OR scope.kind IN ('module', 'namespace')
                   OR {target}.kind IN ({NESTED_TYPE_KINDS})
                      AND s_from.language IN ({implicit_receiver_languages}))
              AND NOT EXISTS (
                  SELECT 1 FROM lexical closer
                  JOIN symbols nearer ON nearer.parent_symbol_id = closer.symbol_id
                  WHERE closer.depth < lexical.depth
                    AND nearer.name = {target}.name
                    AND nearer.kind IN ({target}.kind, 'import')
              )
        )"
    )
}

/// SQL predicate that decides whether a pending call edge `p` (with caller `s_from`) points at
/// the candidate definition `target` (whose parent symbol is joined as `parent`).
pub(crate) fn pending_target_predicate(conn: &Connection, target: &str, parent: &str) -> String {
    let ns = "json_each(CASE WHEN json_valid(p.target_namespace_json) THEN p.target_namespace_json ELSE '[]' END)";
    let target_path = format!("('/' || replace({target}.path, '\\', '/'))");
    let like_value = "replace(replace(replace(value, '\\', '\\\\'), '%', '\\%'), '_', '\\_')";
    let closer_rank = call_site_proximity("closer.path");
    let target_rank = call_site_proximity(&format!("{target}.path"));
    let relative = is_relative_import_path("import_source.value");
    let relative_match = relative_import_matches("import_source.value", target);
    // `import { util } from "../helpers/util.js"`: `util.f()` reaches only that file's `util`.
    let receiver_import_elsewhere = if has_column(conn, "symbols", "metadata_json") {
        format!(
            "AND NOT EXISTS (
                            SELECT 1 FROM symbols receiver_import
                            CROSS JOIN json_each(json_array(
                                json_extract(receiver_import.metadata_json, '$.source')
                            )) import_source
                            WHERE receiver_import.kind = 'import'
                              AND receiver_import.path = p.path
                              AND receiver_import.name = p.target_receiver
                              AND json_valid(receiver_import.metadata_json)
                              AND {relative}
                              AND NOT {relative_match}
                        )"
        )
    } else {
        String::new()
    };
    let module_import = module_import_binding_sql(target, "p.target_receiver");
    let import_alias_receiver = if has_column(conn, "symbols", "metadata_json") {
        format!("OR EXISTS (
                        SELECT 1 FROM symbols alias_import
                        WHERE alias_import.kind = 'import'
                          AND alias_import.path = p.path
                          AND json_valid(alias_import.metadata_json)
                          AND (json_extract(alias_import.metadata_json, '$.alias') = p.target_receiver
                               OR json_extract(alias_import.metadata_json, '$.local_name') = p.target_receiver)
                          AND COALESCE(json_extract(alias_import.metadata_json, '$.source'), '') NOT LIKE 'Qt%'
                    )
                    OR {module_import}"
        )
    } else {
        String::new()
    };
    let receiver_type = receiver_type_sql(conn, parent);
    let fixture_receiver = match built_receiver_sql(conn, target) {
        Some(built) => format!("OR {built}"),
        None => String::new(),
    };
    let inherited = class_chain_sql(target, &format!("{target}.kind IN ('variable', 'field')"));
    let receiver_chain = class_chain_sql(target, "p.target_receiver IS NOT 'super'");
    let implicit_receiver_languages = implicit_receiver_language_list();
    let enclosing = enclosing_scope_sql(target);
    let rust_path = rust_path_rule(target, ns);
    format!(
        "(
            NOT (p.kind IS 'extends' AND p.from_symbol_id = {target}.symbol_id)
            AND (
            (
                {target}.parent_symbol_id IS NOT NULL
                AND {parent}.name IS NOT NULL
                AND (
                    EXISTS (SELECT 1 FROM {ns} WHERE value = {parent}.name)
                    OR (EXISTS (SELECT 1 FROM {ns} WHERE value = 'Self')
                        AND s_from.parent_symbol_id = {target}.parent_symbol_id)
                    OR (p.target_receiver IS NOT NULL AND p.target_receiver != '' AND {parent}.name = p.target_receiver
                        {receiver_import_elsewhere})
                    OR ((p.target_receiver IN ('self', 'this', 'cls', 'Self', 'super')
                         OR EXISTS (SELECT 1 FROM {ns} WHERE value = 'Self'))
                        AND {receiver_chain})
                    OR {receiver_type}
                    {fixture_receiver}
                )
                AND NOT EXISTS (
                    SELECT 1 FROM {ns}
                    WHERE value NOT IN ('std', 'core', 'alloc', 'crate', 'super', 'self', 'Self', {parent}.name)
                      AND NOT EXISTS (
                          WITH RECURSIVE ancestor(symbol_id, depth) AS (
                              SELECT {target}.parent_symbol_id, 0
                              UNION ALL
                              SELECT s.parent_symbol_id, ancestor.depth + 1
                              FROM symbols s JOIN ancestor ON s.symbol_id = ancestor.symbol_id
                              WHERE s.parent_symbol_id IS NOT NULL AND ancestor.depth < 32
                          )
                          SELECT 1 FROM ancestor JOIN symbols a ON a.symbol_id = ancestor.symbol_id
                          WHERE a.name = value
                      )
                      AND {target_path} NOT LIKE '%/' || {like_value} || '.%' ESCAPE '\\'
                      AND {target_path} NOT LIKE '%/' || {like_value} || '/%' ESCAPE '\\'
                )
            )
            OR (
                (p.target_namespace_json IS NULL OR p.target_namespace_json = '[]')
                AND (
                    p.target_receiver IS NULL
                    OR p.target_receiver = ''
                    {import_alias_receiver}
                )
                AND ({target}.parent_symbol_id IS NULL
                     OR s_from.parent_symbol_id = {target}.parent_symbol_id
                     OR {target}.parent_symbol_id = p.from_symbol_id
                     OR CASE WHEN p.target_receiver IS NULL OR p.target_receiver = ''
                        THEN {enclosing}
                        ELSE 0 END
                     OR CASE WHEN (p.target_receiver IS NULL OR p.target_receiver = '')
                                  AND s_from.language IN ({implicit_receiver_languages})
                        THEN {parent}.path = s_from.path
                             AND {parent}.name = (
                                 SELECT caller_owner.name FROM symbols caller_owner
                                 WHERE caller_owner.symbol_id = s_from.parent_symbol_id
                                   AND caller_owner.kind = {parent}.kind
                             )
                             OR {inherited}
                        ELSE 0 END)
                AND (p.target_receiver IS NOT NULL AND p.target_receiver != '' OR {target}.path = p.path OR NOT EXISTS (
                    SELECT 1 FROM symbols shadow
                    WHERE shadow.name = {target}.name
                      AND shadow.path = p.path
                      AND shadow.symbol_id != p.from_symbol_id
                      AND shadow.kind NOT IN ('import', 'export', 'module', 'namespace')
                      AND (shadow.parent_symbol_id IS NULL OR shadow.parent_symbol_id = p.from_symbol_id)
                ))
                AND ({target}.parent_symbol_id IS NOT NULL OR NOT EXISTS (
                    SELECT 1 FROM symbols closer
                    WHERE closer.name = {target}.name
                      AND closer.symbol_id != {target}.symbol_id
                      AND closer.parent_symbol_id IS NULL
                      AND closer.kind = {target}.kind
                      AND NOT (p.kind IS 'extends' AND closer.symbol_id = p.from_symbol_id)
                      AND {closer_rank} > {target_rank}
                ))
            )
            OR (
                {target}.parent_symbol_id IS NULL
                AND (
                    EXISTS (
                        SELECT 1 FROM {ns}
                        WHERE value NOT IN ('std', 'core', 'alloc', 'crate', 'super')
                          AND {target_path} LIKE '%/' || {like_value} || '.%' ESCAPE '\\'
                    )
                    OR ({rust_path})
                )
            )
            )
        )"
    )
}

/// SQL predicate for a Rust call written with a path (`crate::f()`, `super::f()`, `my_crate::f()`,
/// `crate::module::f()`) that reaches the top-level function `target`. A package folder is the
/// path before the last `/src/`; a leading `crate`, `self`, or `super` names the caller's own
/// package, a leading crate name names the package folder with `-` read as `_`, and every other
/// segment must be a file stem or folder below that `src/`. A call rooted at `std`, `core`, or
/// `alloc` never matches.
fn rust_path_rule(target: &str, ns: &str) -> String {
    let target_path = format!("('/' || replace({target}.path, '\\', '/'))");
    let caller_path = "('/' || replace(p.path, '\\', '/'))";
    let skipped = "'std', 'core', 'alloc', 'crate', 'self', 'super', 'Self'";
    let like_segment = "replace(replace(replace(seg.value, '\\', '\\\\'), '%', '\\%'), '_', '\\_')";
    let in_module_path = format!(
        "(module_path.value LIKE '%/' || {like_segment} || '.%' ESCAPE '\\'
          OR module_path.value LIKE '%/' || {like_segment} || '/%' ESCAPE '\\')"
    );
    format!(
        "s_from.language = 'rust'
        AND {target}.language = 'rust'
        AND EXISTS (
            SELECT 1
            FROM json_each(json_array(replace({target_path}, '/src/', char(1)))) marked
            CROSS JOIN json_each(json_array(
                CASE WHEN instr(marked.value, char(1)) > 0
                     THEN length({target_path})
                          - length(replace(marked.value, rtrim(marked.value, replace(marked.value, char(1), '')), ''))
                     ELSE 0 END)) src
            CROSS JOIN json_each(json_array(
                CASE WHEN src.value > 0 THEN replace(substr({target_path}, src.value), '-', '_') ELSE '' END)) module_path
            WHERE (
                EXISTS (
                    SELECT 1 FROM {ns} head
                    WHERE head.key = 0
                      AND (head.value IN ('crate', 'self', 'super')
                           AND ({target}.path = p.path
                                OR src.value > 0
                                   AND substr({caller_path}, 1, src.value) = substr({target_path}, 1, src.value))
                           OR head.value NOT IN ({skipped})
                              AND src.value >= length(head.value) + 6
                              AND replace(substr({target_path}, src.value - length(head.value) - 5, length(head.value) + 6), '-', '_')
                                  = '/' || head.value || '/src/')
                )
                AND NOT EXISTS (
                    SELECT 1 FROM {ns} seg
                    WHERE seg.key > 0 AND seg.value NOT IN ({skipped}) AND NOT {in_module_path}
                )
                OR src.value > 0
                   AND EXISTS (SELECT 1 FROM {ns} head WHERE head.key = 0 AND head.value NOT IN ({skipped}))
                   AND NOT EXISTS (
                       SELECT 1 FROM {ns} seg
                       WHERE seg.value NOT IN ({skipped}) AND NOT {in_module_path}
                   )
            )
        )"
    )
}

/// SQL predicate excluding rows julie marked as documentation, or the always-true `1 = 1` when
/// the column is absent, because a bare `1` in ORDER BY means the first result column in SQLite.
pub(crate) const DOCUMENTATION_LANGUAGES: &[&str] = &[
    "markdown", "yaml", "toml", "json", "html", "css", "xml", "ini", "text",
];

pub(crate) fn documentation_language_list() -> String {
    DOCUMENTATION_LANGUAGES
        .iter()
        .map(|l| format!("'{l}'"))
        .collect::<Vec<_>>()
        .join(", ")
}

fn not_documentation(conn: &Connection, alias: &str) -> String {
    let has_content_type: bool = conn
        .query_row(
            "SELECT 1 FROM pragma_table_info('symbols') WHERE name = 'content_type'",
            [],
            |_| Ok(true),
        )
        .unwrap_or(false);
    if has_content_type {
        format!("({alias}.content_type IS NULL OR {alias}.content_type != 'documentation')")
    } else {
        "1 = 1".to_string()
    }
}

pub(crate) fn has_column(conn: &Connection, table: &str, column: &str) -> bool {
    conn.query_row(
        "SELECT 1 FROM pragma_table_info(?1) WHERE name = ?2",
        [table, column],
        |_| Ok(true),
    )
    .unwrap_or(false)
}

pub(crate) fn has_table(conn: &Connection, name: &str) -> bool {
    conn.query_row(
        "SELECT 1 FROM sqlite_master WHERE type = 'table' AND name = ?1",
        [name],
        |_| Ok(true),
    )
    .unwrap_or(false)
}

fn has_pending_namespace_column(conn: &Connection) -> bool {
    let has_ns: bool = conn
        .query_row(
            "SELECT 1 FROM pragma_table_info('pending_relationships') WHERE name = 'target_namespace_json'",
            [],
            |_| Ok(true),
        )
        .unwrap_or(false);
    let has_display: bool = conn
        .query_row(
            "SELECT 1 FROM pragma_table_info('pending_relationships') WHERE name = 'target_display_name'",
            [],
            |_| Ok(true),
        )
        .unwrap_or(false);
    let has_receiver: bool = conn
        .query_row(
            "SELECT 1 FROM pragma_table_info('pending_relationships') WHERE name = 'target_receiver'",
            [],
            |_| Ok(true),
        )
        .unwrap_or(false);
    has_ns && has_display && has_receiver
}

fn find_references_internal(
    conn: &Connection,
    symbol_name: &str,
    direction: &str,
    limit: usize,
    symbol_id: Option<&str>,
    include_external: bool,
) -> Result<Vec<ReferenceSite>, QueryError> {
    let mut sites = find_direct_references(
        conn,
        symbol_name,
        direction,
        limit,
        symbol_id,
        include_external,
    )?;
    let class = match symbol_id.filter(|_| direction == "callers") {
        Some(id) => constructed_class(conn, id)?,
        None => None,
    };
    if let Some((class_id, class_name)) = class.filter(|_| sites.len() < limit) {
        let mut builds = find_direct_references(
            conn,
            &class_name,
            "callers",
            limit - sites.len(),
            Some(&class_id),
            include_external,
        )?;
        builds.retain(|site| site.kind == "calls");
        sites.extend(builds);
        sites.sort_by(|a, b| {
            a.path
                .cmp(&b.path)
                .then(a.start_line.cmp(&b.start_line))
                .then(a.start_column.cmp(&b.start_column))
        });
    }
    if direction == "callers" {
        populate_lambda_enclosing_names(conn, &mut sites)?;
    }
    Ok(sites)
}

pub(crate) fn populate_lambda_enclosing_names(
    conn: &Connection,
    sites: &mut [ReferenceSite],
) -> Result<(), QueryError> {
    if !sites
        .iter()
        .any(|site| has_generated_lambda_name(&site.from_symbol_name))
    {
        return Ok(());
    }
    let mut stmt = conn.prepare(
        "SELECT caller.language, caller.kind, caller.signature, parent.name
         FROM symbols caller
         LEFT JOIN symbols parent ON parent.symbol_id = caller.parent_symbol_id
         WHERE caller.symbol_id = ?1",
    )?;
    for site in sites
        .iter_mut()
        .filter(|site| has_generated_lambda_name(&site.from_symbol_name))
    {
        let details = stmt
            .query_row([&site.from_symbol_id], |row| {
                Ok((
                    row.get::<_, String>(0)?,
                    row.get::<_, String>(1)?,
                    row.get::<_, Option<String>>(2)?,
                    row.get::<_, Option<String>>(3)?,
                ))
            })
            .optional()?;
        if let Some((language, kind, signature, enclosing_symbol_name)) = details
            && is_generated_lambda_name(
                &site.from_symbol_name,
                &language,
                &kind,
                signature.as_deref(),
            )
        {
            site.enclosing_symbol_name = enclosing_symbol_name;
        }
    }
    Ok(())
}

/// The class a constructor builds, so the calls that build the class count as its callers.
pub(crate) fn constructed_class(
    conn: &Connection,
    symbol_id: &str,
) -> Result<Option<(String, String)>, QueryError> {
    Ok(conn
        .query_row(
            "SELECT class.symbol_id, class.name
             FROM symbols ctor JOIN symbols class ON class.symbol_id = ctor.parent_symbol_id
             WHERE ctor.symbol_id = ?1 AND ctor.kind = 'constructor'
               AND class.kind IN ('class', 'struct', 'record')",
            params![symbol_id],
            |row| Ok((row.get(0)?, row.get(1)?)),
        )
        .optional()?)
}

/// A type-usage or member-access identifier that resolves to the selected target.
pub(crate) struct DirectIdentifier {
    pub identifier_id: String,
    pub site: ReferenceSite,
}

/// Type-usage and member-access identifiers named `symbol_name` that reach the selected target,
/// at most `remaining`: supported rows first, then candidates, each in source order.
pub(crate) fn direct_identifier_rows(
    conn: &Connection,
    symbol_name: &str,
    symbol_id: Option<&str>,
    remaining: usize,
) -> Result<Vec<DirectIdentifier>, QueryError> {
    let receiver_match = member_receiver_match(conn);
    // A name-only type usage reaches only the closest same-named type visible from
    // the caller's lexical scopes, then the file's top-level types.
    let type_usage_shadow = format!(
        "i.kind = 'type_usage'
                 AND ?2 IS NOT NULL
                 AND selected_target.kind IN ({NESTED_TYPE_KINDS})
                 AND EXISTS (
                     WITH RECURSIVE lexical(symbol_id, depth) AS (
                         SELECT s_from.symbol_id, 0
                         UNION ALL
                         SELECT scope.parent_symbol_id, lexical.depth + 1
                         FROM symbols scope JOIN lexical ON scope.symbol_id = lexical.symbol_id
                         WHERE scope.parent_symbol_id IS NOT NULL AND lexical.depth < 32
                     ), visible_types(symbol_id, depth) AS (
                         SELECT shadow.symbol_id, lexical.depth
                         FROM lexical
                         JOIN symbols shadow ON shadow.parent_symbol_id = lexical.symbol_id
                         WHERE shadow.name = selected_target.name
                           AND shadow.kind IN ({NESTED_TYPE_KINDS})
                           AND shadow.language = i.language
                         UNION ALL
                         SELECT shadow.symbol_id, 100
                         FROM symbols shadow
                         WHERE shadow.name = selected_target.name
                           AND shadow.kind IN ({NESTED_TYPE_KINDS})
                           AND shadow.language = i.language
                           AND shadow.parent_symbol_id IS NULL
                           AND replace(shadow.path, '\\', '/') = replace(i.path, '\\', '/')
                     )
                     SELECT 1 FROM visible_types nearest
                     WHERE nearest.symbol_id != selected_target.symbol_id
                       AND nearest.depth = (SELECT MIN(depth) FROM visible_types)
                 )"
    );
    let receiver_match = receiver_match.replace("?3", "?1");
    let language_match = identifier_language_matches("i", "selected_target");
    let mut receiver_stmt = conn.prepare(&format!(
                "SELECT {receiver_match}
                 FROM (SELECT name, kind, path, start_line, start_column,
                              containing_symbol_id AS from_symbol_id,
                              CASE WHEN json_valid(metadata_json) THEN metadata_json ELSE '{{}}' END AS metadata_json,
                              COALESCE(json_valid(metadata_json), 0) AS has_metadata,
                              language,
                              CASE WHEN json_valid(metadata_json) THEN json_extract(metadata_json, '$.receiver') END AS target_receiver,
                              '[]' AS target_namespace_json
                       FROM identifiers WHERE identifier_id = ?2) p
                 LEFT JOIN symbols s_from ON p.from_symbol_id = s_from.symbol_id"
            ))?;
    let mut ident_stmt = conn.prepare(&format!(
                "WITH raw AS MATERIALIZED (
                     SELECT identifier_id, name, kind, path, start_line, start_column,
                            containing_symbol_id, language,
                            CASE WHEN json_valid(metadata_json) THEN metadata_json ELSE '{{}}' END AS metadata_json
                     FROM identifiers WHERE name = ?1 AND kind IN ('type_usage', 'member_access')
                 )
                 SELECT i.identifier_id,
                        COALESCE(s_from.name, ''),
                        COALESCE(i.containing_symbol_id, ''),
                        i.name,
                        CASE WHEN json_valid(i.metadata_json)
                                  AND json_extract(i.metadata_json, '$.role') = 'signal_handler'
                             THEN CASE WHEN json_extract(i.metadata_json, '$.receiver') = (
                                      SELECT CASE WHEN t.kind IN ('class', 'struct', 'enum', 'interface', 'trait', 'module', 'namespace')
                                                  THEN t.name ELSE tp.name END
                                      FROM symbols t
                                      LEFT JOIN symbols tp ON tp.symbol_id = t.parent_symbol_id
                                      WHERE t.symbol_id = ?2)
                                  THEN 'handler' ELSE 'handler (candidate)' END
                             ELSE i.kind END,
                        i.path,
                        i.start_line,
                        i.start_column
                 FROM raw i
                 LEFT JOIN symbols s_from ON i.containing_symbol_id = s_from.symbol_id
                 LEFT JOIN symbols selected_target ON selected_target.symbol_id = ?2
                 WHERE i.name = ?1 AND i.kind IN ('type_usage', 'member_access')
                   AND COALESCE(s_from.kind, '') != 'import'
                   AND (?2 IS NULL OR {language_match})
                   AND NOT ({type_usage_shadow})
                   AND NOT EXISTS (
                       SELECT 1 FROM relationships covered
                       JOIN symbols covered_to ON covered.to_symbol_id = covered_to.symbol_id
                       WHERE covered_to.name = i.name
                         AND covered.path = i.path
                         AND covered.start_line = i.start_line
                         AND (?2 IS NULL OR covered.to_symbol_id = ?2)
                   )
                   AND NOT EXISTS (
                       SELECT 1 FROM pending_relationships covered
                       WHERE covered.target_terminal_name = i.name
                         AND covered.path = i.path
                         AND covered.start_line = i.start_line
                   )
                   AND (?2 IS NULL OR NOT EXISTS (
                       SELECT 1 FROM symbols owner
                       JOIN symbols member ON member.parent_symbol_id = owner.symbol_id
                       WHERE owner.name = CASE WHEN json_valid(i.metadata_json) THEN json_extract(i.metadata_json, '$.receiver') END
                         AND member.name = i.name
                         AND owner.name IS NOT (SELECT parent.name FROM symbols target
                                                JOIN symbols parent ON parent.symbol_id = target.parent_symbol_id
                                                WHERE target.symbol_id = ?2)
                   ))
                   AND NOT (i.kind = 'member_access' AND EXISTS (
                       SELECT 1 FROM symbols target
                       LEFT JOIN symbols scope ON scope.symbol_id = target.parent_symbol_id
                       WHERE target.symbol_id = ?2
                         AND (scope.kind IN ('function', 'method', 'constructor')
                              OR (scope.symbol_id IS NULL
                                  AND target.path != i.path
                                  AND target.language IN ('javascript', 'typescript', 'tsx')
                                  AND COALESCE(target.visibility, 'private') = 'private'))
                   ))
                   AND NOT COALESCE((i.kind = 'member_access'
                        AND json_valid(i.metadata_json)
                        AND json_extract(i.metadata_json, '$.role') IS NOT 'signal_handler'
                        AND json_extract(i.metadata_json, '$.receiver') GLOB '[A-Z]*'
                        AND NOT EXISTS (SELECT 1 FROM symbols known
                                        WHERE known.name = json_extract(i.metadata_json, '$.receiver')
                                          AND known.kind NOT IN ('variable', 'parameter', 'method'))), 0)
                 ORDER BY i.path, i.start_line, i.start_column, i.identifier_id"
            ))?;
    let mut supported = Vec::new();
    let mut candidates = Vec::new();
    let mut supported_keys = HashSet::new();
    let mut candidate_keys = HashSet::new();
    let site_key = |site: &ReferenceSite| {
        (
            site.from_symbol_id.clone(),
            site.to_symbol_name.clone(),
            site.kind.clone(),
            site.path.clone(),
            site.start_line,
        )
    };
    let mut ident_rows = ident_stmt.query(params![symbol_name, symbol_id])?;
    while let Some(row) = ident_rows.next()? {
        let identifier_id: String = row.get(0)?;
        let mut site = ReferenceSite {
            from_symbol_name: row.get(1)?,
            from_symbol_id: row.get(2)?,
            to_symbol_name: row.get(3)?,
            kind: row.get(4)?,
            path: row.get::<_, String>(5)?.replace('\\', "/"),
            start_line: row.get::<_, Option<i64>>(6)?.map(|v| v as usize),
            start_column: row.get::<_, Option<i64>>(7)?.map(|v| v as usize),
            occurrences: None,
            target: None,
            enclosing_symbol_name: None,
        };
        let receiver_match: i64 =
            receiver_stmt.query_row(params![symbol_id, identifier_id], |row| row.get(0))?;
        if receiver_match < 0 {
            continue;
        }
        if receiver_match == 0 && site.kind == "member_access" {
            site.kind = "member_access (candidate)".to_string();
        }
        if site.kind == "member_access (candidate)" {
            if candidates.len() < remaining && candidate_keys.insert(site_key(&site)) {
                candidates.push(DirectIdentifier {
                    identifier_id,
                    site,
                });
            }
        } else if supported_keys.insert(site_key(&site)) {
            supported.push(DirectIdentifier {
                identifier_id,
                site,
            });
            if supported.len() >= remaining {
                break;
            }
        }
    }

    let site_order = |a: &DirectIdentifier, b: &DirectIdentifier| {
        a.site
            .path
            .cmp(&b.site.path)
            .then(a.site.start_line.cmp(&b.site.start_line))
            .then(a.site.start_column.cmp(&b.site.start_column))
    };
    supported.sort_by(site_order);
    supported.truncate(remaining);
    candidates.sort_by(site_order);
    candidates.truncate(remaining - supported.len());
    let mut rows = supported;
    rows.extend(candidates);
    Ok(rows)
}

/// Member accesses whose receiver names the type `symbol_id`, one row per file with an
/// `occurrences` count, at most `remaining`.
pub(crate) fn grouped_member_access_rows(
    conn: &Connection,
    symbol_id: &str,
    remaining: usize,
) -> Result<Vec<ReferenceSite>, QueryError> {
    let language_match = identifier_language_matches("i", "target");
    let mut receiver_stmt = conn.prepare(&format!(
                    "SELECT COALESCE(s.name, ''),
                            COALESCE(i.containing_symbol_id, ''),
                            i.name,
                            i.kind,
                            i.path,
                            MIN(i.start_line),
                            i.start_column,
                            COUNT(*)
                     FROM identifiers i
                     LEFT JOIN symbols s ON i.containing_symbol_id = s.symbol_id
                     JOIN symbols target ON target.symbol_id = ?2
                     LEFT JOIN symbols target_parent ON target_parent.symbol_id = target.parent_symbol_id
                     WHERE i.kind = 'member_access'
                       AND i.name != target.name
                       AND COALESCE(s.kind, '') != 'import'
                       AND target.kind IN ('class', 'struct', 'enum', 'interface', 'trait', 'module', 'namespace')
                       AND json_valid(i.metadata_json)
                       AND json_extract(i.metadata_json, '$.receiver') = target.name
                       AND (json_extract(i.metadata_json, '$.receiver_qualifier') IS NULL
                            OR json_extract(i.metadata_json, '$.receiver_qualifier') = target_parent.name)
                       AND {language_match}
                       AND NOT EXISTS (
                           SELECT 1 FROM symbols nearer
                           WHERE nearer.name = target.name
                             AND nearer.path = i.path
                             AND nearer.symbol_id != target.symbol_id
                             AND nearer.kind IN ({NESTED_TYPE_KINDS}, 'module', 'namespace')
                       )
                     GROUP BY i.path
                     ORDER BY i.path, i.start_line
                     LIMIT ?1"
                ))?;
    let rows = receiver_stmt.query_map(params![remaining as i64, symbol_id], |row| {
        Ok(ReferenceSite {
            from_symbol_name: row.get(0)?,
            from_symbol_id: row.get(1)?,
            to_symbol_name: row.get(2)?,
            kind: row.get(3)?,
            path: row.get::<_, String>(4)?.replace('\\', "/"),
            start_line: row.get::<_, Option<i64>>(5)?.map(|v| v as usize),
            start_column: row.get::<_, Option<i64>>(6)?.map(|v| v as usize),
            occurrences: Some(row.get::<_, i64>(7)? as usize),
            target: None,
            enclosing_symbol_name: None,
        })
    })?;
    Ok(rows.collect::<Result<_, _>>()?)
}

fn find_direct_references(
    conn: &Connection,
    symbol_name: &str,
    direction: &str,
    limit: usize,
    symbol_id: Option<&str>,
    include_external: bool,
) -> Result<Vec<ReferenceSite>, QueryError> {
    let mut results = Vec::new();

    if direction == "callers" {
        // Find callers: references pointing to target symbol
        let mut stmt = conn.prepare(
            "SELECT s_from.name AS from_name,
                    r.from_symbol_id,
                    s_to.name AS to_name,
                    r.kind,
                    r.path,
                    r.start_line,
                    r.start_column
             FROM relationships r
             JOIN symbols s_from ON r.from_symbol_id = s_from.symbol_id
             JOIN symbols s_to ON r.to_symbol_id = s_to.symbol_id
             WHERE s_to.name = ?1 AND (?3 IS NULL OR r.to_symbol_id = ?3)
             LIMIT ?2",
        )?;

        let rows = stmt.query_map(params![symbol_name, limit as i64, symbol_id], |row| {
            Ok(ReferenceSite {
                from_symbol_name: row.get(0)?,
                from_symbol_id: row.get(1)?,
                to_symbol_name: row.get(2)?,
                kind: row.get(3)?,
                path: row.get::<_, String>(4)?.replace('\\', "/"),
                start_line: row.get::<_, Option<i64>>(5)?.map(|v| v as usize),
                start_column: row.get::<_, Option<i64>>(6)?.map(|v| v as usize),
                occurrences: None,
                target: None,
                enclosing_symbol_name: None,
            })
        })?;

        for r in rows {
            results.push(r?);
        }

        // Also query pending_relationships for callers if results < limit
        if results.len() < limit {
            let remaining = limit - results.len();
            if has_pending_namespace_column(conn) {
                if let Some(sid) = symbol_id {
                    let mut pending_stmt = conn.prepare(
                        &format!("SELECT s_from.name AS from_name,
                                p.from_symbol_id,
                                p.target_terminal_name AS to_name,
                                p.kind,
                                p.path,
                                p.start_line,
                                p.start_column
                         FROM pending_relationships p
                         JOIN symbols s_from ON p.from_symbol_id = s_from.symbol_id
                         JOIN symbols s_target ON s_target.symbol_id = ?3
                         LEFT JOIN symbols s_target_parent ON s_target.parent_symbol_id = s_target_parent.symbol_id
                          WHERE p.target_terminal_name = ?1
                            AND {pred}
                          LIMIT ?2", pred = pending_target_predicate(conn, "s_target", "s_target_parent")),
                    )?;

                    let p_rows = pending_stmt.query_map(
                        params![symbol_name, remaining as i64, sid],
                        |row| {
                            Ok(ReferenceSite {
                                from_symbol_name: row.get(0)?,
                                from_symbol_id: row.get(1)?,
                                to_symbol_name: row.get(2)?,
                                kind: row.get(3)?,
                                path: row.get::<_, String>(4)?.replace('\\', "/"),
                                start_line: Some(row.get::<_, i64>(5)? as usize),
                                start_column: row.get::<_, Option<i64>>(6)?.map(|v| v as usize),
                                occurrences: None,
                                target: None,
                                enclosing_symbol_name: None,
                            })
                        },
                    )?;
                    for r in p_rows {
                        results.push(r?);
                    }
                } else {
                    let mut pending_stmt = conn.prepare(
                        "SELECT s_from.name AS from_name,
                                p.from_symbol_id,
                                p.target_terminal_name AS to_name,
                                p.kind,
                                p.path,
                                p.start_line,
                                p.start_column
                         FROM pending_relationships p
                         JOIN symbols s_from ON p.from_symbol_id = s_from.symbol_id
                         WHERE p.target_terminal_name = ?1
                           AND (
                               (p.target_namespace_json IS NULL OR p.target_namespace_json = '[]')
                               OR EXISTS (
                                   SELECT 1 FROM symbols s_any
                                   JOIN symbols s_any_parent ON s_any.parent_symbol_id = s_any_parent.symbol_id
                                   WHERE s_any.name = p.target_terminal_name
                                     AND EXISTS (SELECT 1 FROM json_each(CASE WHEN json_valid(p.target_namespace_json) THEN p.target_namespace_json ELSE '[]' END) WHERE value = s_any_parent.name)
                               )
                           )
                         LIMIT ?2",
                    )?;

                    let p_rows =
                        pending_stmt.query_map(params![symbol_name, remaining as i64], |row| {
                            Ok(ReferenceSite {
                                from_symbol_name: row.get(0)?,
                                from_symbol_id: row.get(1)?,
                                to_symbol_name: row.get(2)?,
                                kind: row.get(3)?,
                                path: row.get::<_, String>(4)?.replace('\\', "/"),
                                start_line: Some(row.get::<_, i64>(5)? as usize),
                                start_column: row.get::<_, Option<i64>>(6)?.map(|v| v as usize),
                                occurrences: None,
                                target: None,
                                enclosing_symbol_name: None,
                            })
                        })?;
                    for r in p_rows {
                        results.push(r?);
                    }
                }
            } else {
                let is_nested = if let Some(sid) = symbol_id {
                    conn.query_row(
                        "SELECT 1 FROM symbols WHERE symbol_id = ?1 AND parent_symbol_id IS NOT NULL",
                        params![sid],
                        |_| Ok(true),
                    )
                    .unwrap_or(false)
                } else {
                    false
                };

                if !is_nested {
                    let mut pending_stmt = conn.prepare(
                        "SELECT s_from.name AS from_name,
                                p.from_symbol_id,
                                p.target_terminal_name AS to_name,
                                p.kind,
                                p.path,
                                p.start_line,
                                p.start_column
                         FROM pending_relationships p
                         JOIN symbols s_from ON p.from_symbol_id = s_from.symbol_id
                         WHERE p.target_terminal_name = ?1
                         LIMIT ?2",
                    )?;

                    let p_rows =
                        pending_stmt.query_map(params![symbol_name, remaining as i64], |row| {
                            Ok(ReferenceSite {
                                from_symbol_name: row.get(0)?,
                                from_symbol_id: row.get(1)?,
                                to_symbol_name: row.get(2)?,
                                kind: row.get(3)?,
                                path: row.get::<_, String>(4)?.replace('\\', "/"),
                                start_line: Some(row.get::<_, i64>(5)? as usize),
                                start_column: row.get::<_, Option<i64>>(6)?.map(|v| v as usize),
                                occurrences: None,
                                target: None,
                                enclosing_symbol_name: None,
                            })
                        })?;

                    for r in p_rows {
                        results.push(r?);
                    }
                }
            }
        }

        if results.len() < limit && has_table(conn, "identifiers") {
            let remaining = limit - results.len();
            let identifiers = direct_identifier_rows(conn, symbol_name, symbol_id, remaining)?;
            results.extend(identifiers.into_iter().map(|row| row.site));
            if let Some(sid) = symbol_id.filter(|_| results.len() < limit) {
                let remaining = limit - results.len();
                results.extend(grouped_member_access_rows(conn, sid, remaining)?);
            }
        }
    } else {
        // Find callees: symbols called by target symbol
        let mut stmt = conn.prepare(
            "SELECT s_from.name AS from_name,
                    r.from_symbol_id,
                    s_to.name AS to_name,
                    r.kind,
                    r.path,
                    r.start_line,
                    r.start_column
             FROM relationships r
             JOIN symbols s_from ON r.from_symbol_id = s_from.symbol_id
             JOIN symbols s_to ON r.to_symbol_id = s_to.symbol_id
             WHERE s_from.name = ?1 AND (?3 IS NULL OR r.from_symbol_id = ?3)
             LIMIT ?2",
        )?;

        let rows = stmt.query_map(params![symbol_name, limit as i64, symbol_id], |row| {
            Ok(ReferenceSite {
                from_symbol_name: row.get(0)?,
                from_symbol_id: row.get(1)?,
                to_symbol_name: row.get(2)?,
                kind: row.get(3)?,
                path: row.get::<_, String>(4)?.replace('\\', "/"),
                start_line: row.get::<_, Option<i64>>(5)?.map(|v| v as usize),
                start_column: row.get::<_, Option<i64>>(6)?.map(|v| v as usize),
                occurrences: None,
                target: None,
                enclosing_symbol_name: None,
            })
        })?;

        for r in rows {
            results.push(r?);
        }

        // Also query pending_relationships for callees
        if results.len() < limit {
            let remaining = limit - results.len();
            let p_rows: Vec<ReferenceSite> = if has_pending_namespace_column(conn) {
                let sql = if include_external {
                    String::from("SELECT DISTINCT s_from.name AS from_name,
                            p.from_symbol_id,
                            COALESCE(NULLIF(p.target_display_name, ''), p.target_terminal_name) AS to_name,
                            p.kind,
                            p.path,
                            p.start_line,
                            p.start_column
                     FROM pending_relationships p
                     JOIN symbols s_from ON p.from_symbol_id = s_from.symbol_id
                     WHERE s_from.name = ?1 AND (?3 IS NULL OR p.from_symbol_id = ?3)
                     LIMIT ?2")
                } else {
                    format!("SELECT DISTINCT s_from.name AS from_name,
                            p.from_symbol_id,
                            p.target_terminal_name AS to_name,
                            p.kind,
                            p.path,
                            p.start_line,
                            p.start_column
                     FROM pending_relationships p
                     JOIN symbols s_from ON p.from_symbol_id = s_from.symbol_id
                     WHERE s_from.name = ?1 AND (?3 IS NULL OR p.from_symbol_id = ?3)
                       AND EXISTS (
                           SELECT 1 FROM symbols s_to
                           LEFT JOIN symbols s_to_parent ON s_to.parent_symbol_id = s_to_parent.symbol_id
                           WHERE s_to.name = p.target_terminal_name
                             AND {kind}
                             AND {pred}
                       )
                     LIMIT ?2", kind = callee_target_kind("s_to_parent"), pred = pending_target_predicate(conn, "s_to", "s_to_parent"))
                };
                let mut pending_stmt = conn.prepare(&sql)?;
                let rows = pending_stmt.query_map(
                    params![symbol_name, remaining as i64, symbol_id],
                    |row| {
                        Ok(ReferenceSite {
                            from_symbol_name: row.get(0)?,
                            from_symbol_id: row.get(1)?,
                            to_symbol_name: row.get(2)?,
                            kind: row.get(3)?,
                            path: row.get::<_, String>(4)?.replace('\\', "/"),
                            start_line: Some(row.get::<_, i64>(5)? as usize),
                            start_column: row.get::<_, Option<i64>>(6)?.map(|v| v as usize),
                            occurrences: None,
                            target: None,
                            enclosing_symbol_name: None,
                        })
                    },
                )?;
                let mut out = Vec::new();
                for r in rows {
                    out.push(r?);
                }
                out
            } else {
                let sql = if include_external {
                    "SELECT DISTINCT s_from.name AS from_name,
                            p.from_symbol_id,
                            p.target_terminal_name AS to_name,
                            p.kind,
                            p.path,
                            p.start_line,
                            p.start_column
                     FROM pending_relationships p
                     JOIN symbols s_from ON p.from_symbol_id = s_from.symbol_id
                     WHERE s_from.name = ?1 AND (?3 IS NULL OR p.from_symbol_id = ?3)
                     LIMIT ?2"
                } else {
                    "SELECT DISTINCT s_from.name AS from_name,
                            p.from_symbol_id,
                            p.target_terminal_name AS to_name,
                            p.kind,
                            p.path,
                            p.start_line,
                            p.start_column
                     FROM pending_relationships p
                     JOIN symbols s_from ON p.from_symbol_id = s_from.symbol_id
                     WHERE s_from.name = ?1 AND (?3 IS NULL OR p.from_symbol_id = ?3)
                       AND EXISTS (SELECT 1 FROM symbols s_to WHERE s_to.name = p.target_terminal_name)
                     LIMIT ?2"
                };
                let mut pending_stmt = conn.prepare(sql)?;

                let rows = pending_stmt.query_map(
                    params![symbol_name, remaining as i64, symbol_id],
                    |row| {
                        Ok(ReferenceSite {
                            from_symbol_name: row.get(0)?,
                            from_symbol_id: row.get(1)?,
                            to_symbol_name: row.get(2)?,
                            kind: row.get(3)?,
                            path: row.get::<_, String>(4)?.replace('\\', "/"),
                            start_line: Some(row.get::<_, i64>(5)? as usize),
                            start_column: row.get::<_, Option<i64>>(6)?.map(|v| v as usize),
                            occurrences: None,
                            target: None,
                            enclosing_symbol_name: None,
                        })
                    },
                )?;
                let mut out = Vec::new();
                for r in rows {
                    out.push(r?);
                }
                out
            };

            for r in p_rows {
                results.push(r);
            }
        }
        results.sort_by(|a, b| {
            a.path
                .cmp(&b.path)
                .then(a.start_line.cmp(&b.start_line))
                .then(a.start_column.cmp(&b.start_column))
        });
        let mut sites = merge_same_site(results);
        describe_callee_targets(conn, &mut sites)?;
        return Ok(sites);
    }

    Ok(merge_same_site(results))
}

/// A pending call reaches a definition, or a class attribute that holds a callable:
/// `self.should_ignore_error(error)` with `should_ignore_error: None = None` in the class.
fn callee_target_kind(parent: &str) -> String {
    format!(
        "(s_to.kind NOT IN ('import', 'variable', 'parameter', 'field', 'property', 'module', 'namespace')
         OR p.kind = 'calls' AND s_to.kind IN ('variable', 'field')
            AND {parent}.kind IN ('class', 'struct', 'interface', 'trait', 'record', 'object', 'protocol'))"
    )
}

/// Names the definitions each callee site can reach: the resolved target, or the definitions a
/// pending call matches. Several matches are listed, so a call to a common name stays clear.
fn describe_callee_targets(
    conn: &Connection,
    sites: &mut [ReferenceSite],
) -> Result<(), QueryError> {
    let owner = "CASE WHEN par.kind IN ('class', 'struct', 'interface', 'trait', 'enum', 'record', 'object', 'protocol', 'union')
                 THEN par.name || '.' ELSE '' END";
    let pending = if has_pending_namespace_column(conn) {
        format!(
            "UNION
             SELECT {owner} || s_to.name, replace(s_to.path, '\\', '/'), s_to.start_line
             FROM pending_relationships p
             JOIN symbols s_from ON p.from_symbol_id = s_from.symbol_id
             JOIN symbols s_to ON s_to.name = p.target_terminal_name
             LEFT JOIN symbols par ON s_to.parent_symbol_id = par.symbol_id
             LEFT JOIN symbols s_to_parent ON s_to.parent_symbol_id = s_to_parent.symbol_id
             WHERE p.from_symbol_id = ?1 AND p.start_line = ?2
               AND (p.target_terminal_name = ?3 OR p.target_display_name = ?3)
               AND {kind}
               AND {pred}",
            kind = callee_target_kind("s_to_parent"),
            pred = pending_target_predicate(conn, "s_to", "s_to_parent")
        )
    } else {
        String::new()
    };
    let mut stmt = conn.prepare(&format!(
        "SELECT {owner} || s_to.name, replace(s_to.path, '\\', '/'), s_to.start_line
         FROM relationships r
         JOIN symbols s_to ON r.to_symbol_id = s_to.symbol_id
         LEFT JOIN symbols par ON s_to.parent_symbol_id = par.symbol_id
         WHERE r.from_symbol_id = ?1 AND r.start_line = ?2 AND s_to.name = ?3
         {pending}
         ORDER BY 2, 3
         LIMIT 4"
    ))?;
    for site in sites.iter_mut() {
        let Some(line) = site.start_line else {
            continue;
        };
        let targets = stmt
            .query_map(
                params![site.from_symbol_id, line as i64, site.to_symbol_name],
                |row| {
                    Ok(format!(
                        "`{}` ({}:{})",
                        row.get::<_, String>(0)?,
                        row.get::<_, String>(1)?,
                        row.get::<_, i64>(2)?
                    ))
                },
            )?
            .collect::<Result<Vec<_>, _>>()?;
        site.target = match targets.len() {
            0 => None,
            1 => targets.into_iter().next(),
            2 | 3 => Some(format!("one of {}", targets.join(", "))),
            _ => Some(format!(
                "more than 3 definitions named `{}`",
                site.to_symbol_name
            )),
        };
    }
    Ok(())
}

/// The import statements of the top-level definition `name` (under `path_filter` when given): an
/// import row with that name whose source module names a folder or file on the definition's path.
/// `find_references` lists calls and type uses, so these go in a summary line after them.
pub fn import_sites(
    conn: &Connection,
    name: &str,
    path_filter: Option<&str>,
) -> Result<Vec<(String, usize)>, QueryError> {
    import_sites_from(conn, name, path_filter, false)
}

/// The import statements of the top-level name `name` that the index no longer defines in the file
/// `path`: the same rows as [`import_sites`], for a definition that an edit removed or renamed.
pub fn import_sites_of_absent(
    conn: &Connection,
    name: &str,
    path: &str,
) -> Result<Vec<(String, usize)>, QueryError> {
    let mut sites = import_sites_from(conn, name, Some(path), true)?;
    sites.extend(rust_use_lines(conn, name, path)?);
    sites.sort();
    sites.dedup();
    Ok(sites)
}

/// The Rust `use` lines of `name` whose last module segment is the stem of the file `path` or the
/// folder of its `mod.rs`. A Rust import row carries no source, so the pending `imports` rows give
/// the module.
fn rust_use_lines(
    conn: &Connection,
    name: &str,
    path: &str,
) -> Result<Vec<(String, usize)>, QueryError> {
    let mut stmt = conn.prepare(
        "SELECT DISTINCT replace(pr.path, '\\', '/'), pr.start_line
         FROM pending_relationships pr
         JOIN files f ON f.file_id = pr.file_id AND f.language = 'rust'
         WHERE pr.kind = 'imports' AND pr.target_terminal_name = ?1
           AND pr.start_line IS NOT NULL AND json_valid(pr.target_namespace_json)
           AND json_extract(pr.target_namespace_json, '$[#-1]') NOT IN ('crate', 'self', 'super')
           AND (?2 GLOB '*/' || json_extract(pr.target_namespace_json, '$[#-1]') || '.rs'
                OR ?2 GLOB '*/' || json_extract(pr.target_namespace_json, '$[#-1]') || '/mod.rs')",
    )?;
    let target_path = format!("/{}", path.replace('\\', "/"));
    let rows = stmt.query_map(params![name, target_path], |row| {
        Ok((row.get::<_, String>(0)?, row.get::<_, i64>(1)? as usize))
    })?;
    Ok(rows.collect::<Result<_, _>>()?)
}

fn import_sites_from(
    conn: &Connection,
    name: &str,
    path_filter: Option<&str>,
    absent: bool,
) -> Result<Vec<(String, usize)>, QueryError> {
    if !has_column(conn, "symbols", "metadata_json") {
        return Ok(Vec::new());
    }
    let targets = if absent {
        "(SELECT ?1 AS name, ?2 AS path, NULL AS parent_symbol_id, 'function' AS kind)"
    } else {
        "symbols"
    };
    let dotted = "replace(module_path.value, '/', '.')";
    let segment = format!("replace({dotted}, rtrim({dotted}, replace({dotted}, '.', '')), '')");
    let target_path = "('/' || replace(target.path, '\\', '/'))";
    // `from . import Flask`: a source of dots only names the importing file's package.
    let dir = |path: &str| {
        format!("rtrim(replace({path}, '\\', '/'), replace(replace({path}, '\\', '/'), '/', ''))")
    };
    let (import_dir, target_dir) = (dir("i.path"), dir("target.path"));
    let mut stmt = conn.prepare(&format!(
        "SELECT DISTINCT replace(i.path, '\\', '/'), i.start_line
         FROM {targets} target
         JOIN symbols i ON i.name = target.name AND i.kind = 'import'
         CROSS JOIN json_each(json_array(json_extract(i.metadata_json, '$.source'))) module_path
         WHERE target.name = ?1
           AND target.parent_symbol_id IS NULL
           AND target.kind NOT IN ({LOW_SIGNAL_KINDS_SQL})
           AND (?2 IS NULL OR replace(target.path, '\\', '/') = ?2
                OR replace(target.path, '\\', '/') LIKE ?3 ESCAPE '\\')
           AND json_valid(i.metadata_json)
           AND i.language NOT IN ({docs})
           AND (
               {segment} != ''
               AND ({target_path} GLOB '*/' || {segment} || '/*'
                    OR {target_path} GLOB '*/' || {segment} || '.*')
               OR module_path.value != '' AND replace(module_path.value, '.', '') = ''
               AND substr({import_dir}, 1, length({target_dir})) = {target_dir}
           )
         ORDER BY 1, 2",
        docs = documentation_language_list()
    ))?;
    let norm = path_filter.map(|p| p.replace('\\', "/").trim_matches('/').to_string());
    let prefix = norm.as_ref().map(|p| format!("{}/%", escape_like(p)));
    let rows = stmt.query_map(params![name, norm, prefix], |row| {
        Ok((row.get::<_, String>(0)?, row.get::<_, i64>(1)? as usize))
    })?;
    Ok(rows.collect::<Result<_, _>>()?)
}

/// One row for each caller, target, kind, and line: `app.wsgi_app = Wrap(app.wsgi_app)` is one
/// site, not two identical rows.
fn merge_same_site(sites: Vec<ReferenceSite>) -> Vec<ReferenceSite> {
    let mut merged: Vec<ReferenceSite> = Vec::with_capacity(sites.len());
    let mut index = HashMap::new();
    for site in sites {
        let key = (
            site.from_symbol_id.clone(),
            site.to_symbol_name.clone(),
            site.kind.clone(),
            site.path.clone(),
            site.start_line,
        );
        match index.get(&key) {
            Some(&at) => {
                let kept: &mut ReferenceSite = &mut merged[at];
                kept.start_column = kept.start_column.min(site.start_column);
            }
            None => {
                index.insert(key, merged.len());
                merged.push(site);
            }
        }
    }
    merged
}

/// Resolve callee signatures directly in a single joined query, avoiding N+1 queries
/// and preserving ambiguous methods across types. Prioritizes functions/methods over enum variants.
pub fn find_callee_signatures(
    conn: &Connection,
    symbol_name: &str,
    symbol_id: &str,
    limit: usize,
    include_external: bool,
) -> Result<Vec<String>, QueryError> {
    let mut stmt = conn.prepare(
        "SELECT DISTINCT s_to.name, s_to.signature, s_to.path, s_to.start_line, s_to.kind
         FROM relationships r
         JOIN symbols s_from ON r.from_symbol_id = s_from.symbol_id
         JOIN symbols s_to ON r.to_symbol_id = s_to.symbol_id
         WHERE s_from.name = ?1 AND r.from_symbol_id = ?2
         LIMIT ?3",
    )?;

    let rows = stmt.query_map(params![symbol_name, symbol_id, (limit * 2) as i64], |row| {
        Ok((
            row.get::<_, String>(0)?,
            row.get::<_, Option<String>>(1)?,
            row.get::<_, String>(2)?.replace('\\', "/"),
            row.get::<_, Option<i64>>(3)?.unwrap_or(1) as usize,
            row.get::<_, String>(4)?,
        ))
    })?;

    let mut signatures = Vec::new();
    let mut variants = Vec::new();

    for r in rows.flatten() {
        let (name, sig_opt, path, line, kind) = r;
        let sig = sig_opt.unwrap_or(name);
        let entry = format!("{sig} ({path}:{line})");
        if kind == "variant" {
            if !variants.contains(&entry) {
                variants.push(entry);
            }
        } else if !signatures.contains(&entry) {
            signatures.push(entry);
        }
    }

    if signatures.len() < limit {
        let remaining = (limit - signatures.len()) * 2;
        let p_rows: Vec<(String, Option<String>, String, usize, String)> =
            if has_pending_namespace_column(conn) {
                let mut p_stmt = conn.prepare(
                &format!("SELECT DISTINCT s_to.name, s_to.signature, s_to.path, s_to.start_line, s_to.kind
                 FROM pending_relationships p
                 JOIN symbols s_from ON p.from_symbol_id = s_from.symbol_id
                 JOIN symbols s_to ON s_to.name = p.target_terminal_name
                 LEFT JOIN symbols s_parent ON s_to.parent_symbol_id = s_parent.symbol_id
                 WHERE s_from.name = ?1 AND p.from_symbol_id = ?2
                   AND {kind}
                    AND {pred}
                 LIMIT ?3", kind = callee_target_kind("s_parent"), pred = pending_target_predicate(conn, "s_to", "s_parent")),
            )?;

                let rows =
                    p_stmt.query_map(params![symbol_name, symbol_id, remaining as i64], |row| {
                        Ok((
                            row.get::<_, String>(0)?,
                            row.get::<_, Option<String>>(1)?,
                            row.get::<_, String>(2)?.replace('\\', "/"),
                            row.get::<_, Option<i64>>(3)?.unwrap_or(1) as usize,
                            row.get::<_, String>(4)?,
                        ))
                    })?;
                rows.flatten().collect()
            } else {
                let mut p_stmt = conn.prepare(
                "SELECT DISTINCT s_to.name, s_to.signature, s_to.path, s_to.start_line, s_to.kind
                 FROM pending_relationships p
                 JOIN symbols s_from ON p.from_symbol_id = s_from.symbol_id
                 JOIN symbols s_to ON s_to.name = p.target_terminal_name
                 WHERE s_from.name = ?1 AND p.from_symbol_id = ?2
                   AND s_to.kind NOT IN ('import', 'variable', 'parameter', 'field', 'property', 'module', 'namespace')
                 LIMIT ?3",
            )?;

                let rows =
                    p_stmt.query_map(params![symbol_name, symbol_id, remaining as i64], |row| {
                        Ok((
                            row.get::<_, String>(0)?,
                            row.get::<_, Option<String>>(1)?,
                            row.get::<_, String>(2)?.replace('\\', "/"),
                            row.get::<_, Option<i64>>(3)?.unwrap_or(1) as usize,
                            row.get::<_, String>(4)?,
                        ))
                    })?;
                rows.flatten().collect()
            };

        for r in p_rows {
            let (name, sig_opt, path, line, kind) = r;
            let sig = sig_opt.unwrap_or(name);
            let entry = format!("{sig} ({path}:{line})");
            if kind == "variant" {
                if !variants.contains(&entry) {
                    variants.push(entry);
                }
            } else if !signatures.contains(&entry) {
                signatures.push(entry);
            }
        }
    }

    if include_external && signatures.len() < limit {
        let remaining = (limit - signatures.len()) * 2;
        let ext_rows: Vec<(String, String, usize)> = if has_pending_namespace_column(conn) {
            let mut ext_stmt = conn.prepare(
                &format!("SELECT DISTINCT COALESCE(NULLIF(p.target_display_name, ''), p.target_terminal_name), p.path, p.start_line
                 FROM pending_relationships p
                 JOIN symbols s_from ON p.from_symbol_id = s_from.symbol_id
                 WHERE s_from.name = ?1 AND p.from_symbol_id = ?2
                   AND NOT EXISTS (
                       SELECT 1 FROM symbols s_to
                       LEFT JOIN symbols s_parent ON s_to.parent_symbol_id = s_parent.symbol_id
                       WHERE s_to.name = p.target_terminal_name
                         AND {kind}
                         AND {pred}
                   )
                 LIMIT ?3", kind = callee_target_kind("s_parent"), pred = pending_target_predicate(conn, "s_to", "s_parent")),
            )?;

            let rows =
                ext_stmt.query_map(params![symbol_name, symbol_id, remaining as i64], |row| {
                    Ok((
                        row.get::<_, String>(0)?,
                        row.get::<_, String>(1)?.replace('\\', "/"),
                        row.get::<_, Option<i64>>(2)?.unwrap_or(1) as usize,
                    ))
                })?;
            rows.flatten().collect()
        } else {
            let mut ext_stmt = conn.prepare(
                "SELECT DISTINCT p.target_terminal_name, p.path, p.start_line
                 FROM pending_relationships p
                 JOIN symbols s_from ON p.from_symbol_id = s_from.symbol_id
                 WHERE s_from.name = ?1 AND p.from_symbol_id = ?2
                   AND NOT EXISTS (
                       SELECT 1 FROM symbols s_to
                       WHERE s_to.name = p.target_terminal_name
                         AND s_to.kind NOT IN ('import', 'variable', 'parameter', 'field', 'property', 'module', 'namespace')
                   )
                 LIMIT ?3",
            )?;

            let rows =
                ext_stmt.query_map(params![symbol_name, symbol_id, remaining as i64], |row| {
                    Ok((
                        row.get::<_, String>(0)?,
                        row.get::<_, String>(1)?.replace('\\', "/"),
                        row.get::<_, Option<i64>>(2)?.unwrap_or(1) as usize,
                    ))
                })?;
            rows.flatten().collect()
        };

        for r in ext_rows {
            let (name, path, line) = r;
            let entry = format!("{name} ({path}:{line})");
            if !signatures.contains(&entry) {
                signatures.push(entry);
            }
        }
    }

    for v in variants {
        if signatures.len() >= limit {
            break;
        }
        if !signatures.contains(&v) {
            signatures.push(v);
        }
    }

    signatures.truncate(limit);
    Ok(signatures)
}

const SQL_FAMILIES: &[&str] = &["sql."];

pub(crate) const ROUTE_FAMILIES: &[&str] = &[
    ".route.",
    ".attribute_route.",
    ".scope_route.",
    ".file_route.",
    ".route_handler.",
    ".route_reference.",
    ".route_group.",
    ".route_prefix.",
    ".resource_route.",
    ".router_mount.",
    ".include_router.",
    ".server_route.",
    ".route_definition.",
];

const CONFIG_FAMILIES: &[&str] = &["toml.", "yaml.", "json.", "env."];

const MODEL_FAMILIES: &[&str] = &[
    "sql.table_definition.v1",
    "sql.column_definition.v1",
    "sql.constraint.v1",
    "sql.foreign_key.v1",
    "sql.index_definition.v1",
    "json.schema.v1",
];

const SIGNAL_FAMILIES: &[&str] = &[".signal_declaration."];

const IMPORT_FAMILIES: &[&str] = &[".import_statement.", ".import."];

const BINDING_FAMILIES: &[&str] = &[".binding."];

const COMPONENT_FAMILIES: &[&str] = &[".object_instantiation.", ".object_type."];

const MODULE_FAMILIES: &[&str] = &[".module."];

const PRAGMA_FAMILIES: &[&str] = &[".pragma.", "javascript.qml_directive.v1"];

const PROPERTY_FAMILIES: &[&str] = &[".property_declaration.", ".qt_property."];

/// Category aliases mapped to the pattern-id families they name.
///
/// A rule that starts with a dot matches any pattern id that contains it, a rule that ends with a
/// dot matches any pattern id that starts with it, and any other rule matches one exact id.
pub const CATEGORY_ALIASES: &[(&str, &[&str])] = &[
    ("sql", SQL_FAMILIES),
    ("query", SQL_FAMILIES),
    ("queries", SQL_FAMILIES),
    ("route", ROUTE_FAMILIES),
    ("routes", ROUTE_FAMILIES),
    ("config", CONFIG_FAMILIES),
    ("model", MODEL_FAMILIES),
    ("models", MODEL_FAMILIES),
    ("signal", SIGNAL_FAMILIES),
    ("signals", SIGNAL_FAMILIES),
    ("import", IMPORT_FAMILIES),
    ("imports", IMPORT_FAMILIES),
    ("binding", BINDING_FAMILIES),
    ("bindings", BINDING_FAMILIES),
    ("component", COMPONENT_FAMILIES),
    ("components", COMPONENT_FAMILIES),
    ("module", MODULE_FAMILIES),
    ("modules", MODULE_FAMILIES),
    ("pragma", PRAGMA_FAMILIES),
    ("property", PROPERTY_FAMILIES),
    ("properties", PROPERTY_FAMILIES),
];

fn category_families(category: &str) -> Option<&'static [&'static str]> {
    let wanted = category.trim().to_ascii_lowercase();
    CATEGORY_ALIASES
        .iter()
        .find(|(alias, _)| *alias == wanted)
        .map(|(_, families)| *families)
}

pub(crate) fn matches_family(pattern_id: &str, rule: &str) -> bool {
    if rule.starts_with('.') {
        pattern_id.contains(rule)
    } else if rule.ends_with('.') {
        pattern_id.starts_with(rule)
    } else {
        pattern_id == rule
    }
}

fn family_clause(column: &str, families: &[&str]) -> String {
    let alternatives: Vec<String> = families
        .iter()
        .map(|rule| {
            let escaped = escape_like(rule);
            if rule.starts_with('.') {
                format!("{column} LIKE '%{escaped}%' ESCAPE '\\'")
            } else if rule.ends_with('.') {
                format!("{column} LIKE '{escaped}%' ESCAPE '\\'")
            } else {
                format!("{column} = '{rule}'")
            }
        })
        .collect();
    format!("({})", alternatives.join(" OR "))
}

/// True when the category text names one of the aliases in `CATEGORY_ALIASES`.
pub fn is_category_alias(category: &str) -> bool {
    category_families(category).is_some()
}

/// Counts the patterns and facts each category alias reaches, given a category listing.
///
/// Only one alias per family list is reported, and an alias with no facts is left out.
pub fn alias_fact_counts(categories: &[(String, usize)]) -> Vec<(&'static str, usize, usize)> {
    let mut reported: Vec<&[&str]> = Vec::new();
    let mut counts = Vec::new();
    for (alias, families) in CATEGORY_ALIASES {
        if reported.contains(families) {
            continue;
        }
        reported.push(families);
        let mut patterns = 0;
        let mut facts = 0;
        for (pattern_id, count) in categories {
            if families.iter().any(|rule| matches_family(pattern_id, rule)) {
                patterns += 1;
                facts += count;
            }
        }
        if facts > 0 {
            counts.push((*alias, patterns, facts));
        }
    }
    counts
}

/// Find structural facts by category (e.g. route, query, model, config), optionally scoped by path.
pub fn find_structural_facts_scoped(
    conn: &Connection,
    category: &str,
    path_filter: Option<&str>,
    limit: usize,
) -> Result<Vec<StructuralFact>, QueryError> {
    validate_result_limit(limit)?;
    let norm_path = path_filter
        .map(|p| {
            p.replace('\\', "/")
                .trim_start_matches("./")
                .trim_matches('/')
                .to_string()
        })
        .filter(|p| !p.is_empty());
    let dir_prefix = norm_path
        .as_deref()
        .map(|p| format!("{}/%", escape_like(p)));
    let cat_pattern = format!("%{}%", escape_like(category));

    let cat_clause = match category_families(category) {
        Some(families) => family_clause("sf.pattern_id", families),
        None => {
            "(sf.pattern_id LIKE :cat ESCAPE '\\' OR sf.capture_name LIKE :cat ESCAPE '\\' OR sf.node_kind LIKE :cat ESCAPE '\\')"
                .to_string()
        }
    };

    let sql = format!(
        "SELECT sf.structural_fact_id, sf.path, sf.language, sf.pattern_id,
                sf.capture_name, sf.node_kind, s.name AS containing_symbol_name,
                sf.start_line, sf.end_line, sf.confidence,
                COALESCE(
                    CASE WHEN json_extract(sf.metadata_json, '$.key_path') LIKE '$.%'
                         THEN substr(json_extract(sf.metadata_json, '$.key_path'), 3)
                         ELSE json_extract(sf.metadata_json, '$.key_path') END,
                    json_extract(sf.metadata_json, '$.key'),
                    CASE WHEN json_extract(sf.metadata_json, '$.normalized_route_template') IS NOT NULL
                         THEN COALESCE(json_extract(sf.metadata_json, '$.verb') || ' ', '')
                              || json_extract(sf.metadata_json, '$.normalized_route_template') END,
                    CASE WHEN sf.pattern_id = 'cpp.qt_property.v1'
                         THEN json_extract(sf.metadata_json, '$.name') END
                ) AS display_key,
                sf.metadata_json
         FROM structural_facts sf
         LEFT JOIN symbols s ON sf.containing_symbol_id = s.symbol_id
         WHERE (:cat IS NOT NULL AND {cat_clause})
           AND (:path IS NULL OR replace(sf.path, '\\', '/') = :path COLLATE NOCASE OR replace(sf.path, '\\', '/') LIKE :dir_prefix ESCAPE '\\')
         ORDER BY sf.path ASC, sf.start_line ASC,
                  COALESCE(
                      json_extract(sf.metadata_json, '$.effective_route_template'),
                      json_extract(sf.metadata_json, '$.route_template'),
                      json_extract(sf.metadata_json, '$.normalized_route_template'),
                      ''
                  ) ASC,
                  COALESCE(json_extract(sf.metadata_json, '$.verb'), '') ASC,
                  sf.structural_fact_id ASC
         LIMIT :limit"
    );

    let mut stmt = conn.prepare(&sql)?;
    let rows = stmt.query_map(
        rusqlite::named_params! {
            ":cat": cat_pattern,
            ":path": norm_path.as_deref(),
            ":dir_prefix": dir_prefix.as_deref(),
            ":limit": limit as i64,
        },
        |row| {
            Ok(StructuralFact {
                structural_fact_id: row.get(0)?,
                path: row.get::<_, String>(1)?.replace('\\', "/"),
                language: row.get(2)?,
                pattern_id: row.get(3)?,
                capture_name: row.get(4)?,
                node_kind: row.get(5)?,
                key: row.get(10)?,
                metadata: row
                    .get::<_, Option<String>>(11)?
                    .map(|json| serde_json::from_str(&json))
                    .transpose()
                    .map_err(|err| {
                        rusqlite::Error::FromSqlConversionFailure(
                            11,
                            rusqlite::types::Type::Text,
                            Box::new(err),
                        )
                    })?,
                containing_symbol_name: row.get(6)?,
                start_line: row.get::<_, i64>(7)? as usize,
                end_line: row.get::<_, i64>(8)? as usize,
                confidence: row.get(9)?,
            })
        },
    )?;

    let mut results = Vec::new();
    for r in rows {
        let mut fact = r?;
        let display = fact.metadata.as_ref().and_then(|metadata| {
            route_display(metadata).or_else(|| mount_display(&fact.capture_name, metadata))
        });
        if let Some(display) = display {
            fact.key = Some(display);
        }
        results.push(fact);
    }
    Ok(results)
}

/// A route as its source spells it, with its verb: `GET /<int:id>/update`, not the normalized
/// `/:id/update`. The prefixed source template is used when the extractor joined a prefix;
/// otherwise the raw template, unless the normalized one holds a prefix the raw one lacks.
fn route_display(metadata: &serde_json::Value) -> Option<String> {
    let text = |key: &str| metadata.get(key).and_then(|value| value.as_str());
    let normalized = text("normalized_route_template")?;
    let template = match (text("effective_route_template"), text("route_template")) {
        (Some(effective), _) => effective,
        (None, Some(raw)) if route_shape(raw) == route_shape(normalized) => raw,
        _ => normalized,
    };
    Some(match text("verb") {
        Some(verb) => format!("{verb} {template}"),
        None => template.to_string(),
    })
}

/// A router or blueprint mount with what it mounts and where: ``blueprint_registration `auth.bp` ``,
/// ``router_mount `router` at /api``. The path includes any enclosing scope the extractor joined.
fn mount_display(capture: &str, metadata: &serde_json::Value) -> Option<String> {
    let text = |key: &str| metadata.get(key).and_then(|value| value.as_str());
    if text("query_family") != Some("framework") {
        return None;
    }
    let target = text("mount_target")?;
    Some(
        match text("normalized_mount_path").or_else(|| text("mount_path")) {
            Some(path) => format!("{capture} `{target}` at {path}"),
            None => format!("{capture} `{target}`"),
        },
    )
}

/// A route template with every parameter (`<int:id>`, `{id}`, `:id`) replaced by `*`.
fn route_shape(template: &str) -> String {
    let mut shape = String::new();
    let mut chars = template.chars().peekable();
    while let Some(c) = chars.next() {
        match c {
            '<' | '{' => {
                let close = if c == '<' { '>' } else { '}' };
                for inner in chars.by_ref() {
                    if inner == close {
                        break;
                    }
                }
                shape.push('*');
            }
            ':' => {
                while chars.peek().is_some_and(|next| *next != '/') {
                    chars.next();
                }
                shape.push('*');
            }
            _ => shape.push(c),
        }
    }
    shape
}

/// Find structural facts by category (e.g. route, query, model, config).
pub fn find_structural_facts(
    conn: &Connection,
    category: &str,
    limit: usize,
) -> Result<Vec<StructuralFact>, QueryError> {
    find_structural_facts_scoped(conn, category, None, limit)
}

/// Find literals (endpoints, SQL queries, configs) matching category, optionally scoped by path.
pub fn find_literals_scoped(
    conn: &Connection,
    category: &str,
    path_filter: Option<&str>,
    limit: usize,
) -> Result<Vec<LiteralFact>, QueryError> {
    validate_result_limit(limit)?;
    let norm_path = path_filter
        .map(|p| {
            p.replace('\\', "/")
                .trim_start_matches("./")
                .trim_matches('/')
                .to_string()
        })
        .filter(|p| !p.is_empty());
    let dir_prefix = norm_path
        .as_deref()
        .map(|p| format!("{}/%", escape_like(p)));
    let cat_pattern = format!("%{}%", escape_like(category));

    let cat_lower = category.trim().to_ascii_lowercase();
    let cat_clause = match cat_lower.as_str() {
        "config" => {
            "(l.kind LIKE '%config%' OR l.kind LIKE '%toml%' OR l.kind LIKE '%json%' OR l.kind LIKE '%yaml%')"
        }
        "route" | "routes" => "l.kind LIKE '%route%'",
        "query" | "queries" | "sql" => "(l.kind LIKE '%sql%' OR l.kind LIKE '%query%')",
        "model" | "models" => "l.kind LIKE '%model%'",
        _ => "(l.kind LIKE :cat ESCAPE '\\' OR l.literal_text LIKE :cat ESCAPE '\\')",
    };

    let sql = format!(
        "SELECT l.literal_id, l.path, l.literal_text, l.kind, l.carrier,
                l.start_line, s.name AS containing_symbol_name
         FROM literals l
         LEFT JOIN symbols s ON l.containing_symbol_id = s.symbol_id
         WHERE (:cat IS NOT NULL AND {cat_clause})
           AND (:path IS NULL OR replace(l.path, '\\', '/') = :path COLLATE NOCASE OR replace(l.path, '\\', '/') LIKE :dir_prefix ESCAPE '\\')
         ORDER BY l.path ASC, l.start_line ASC
         LIMIT :limit"
    );

    let mut stmt = conn.prepare(&sql)?;
    let rows = stmt.query_map(
        rusqlite::named_params! {
            ":cat": cat_pattern,
            ":path": norm_path.as_deref(),
            ":dir_prefix": dir_prefix.as_deref(),
            ":limit": limit as i64,
        },
        |row| {
            Ok(LiteralFact {
                literal_id: row.get(0)?,
                path: row.get::<_, String>(1)?.replace('\\', "/"),
                literal_text: row.get(2)?,
                kind: row.get(3)?,
                carrier: row.get(4)?,
                start_line: row.get::<_, i64>(5)? as usize,
                containing_symbol_name: row.get(6)?,
            })
        },
    )?;

    let mut results = Vec::new();
    for r in rows {
        results.push(r?);
    }
    Ok(results)
}

/// Find literals (endpoints, SQL queries, configs) matching category.
pub fn find_literals(
    conn: &Connection,
    category: &str,
    limit: usize,
) -> Result<Vec<LiteralFact>, QueryError> {
    find_literals_scoped(conn, category, None, limit)
}

/// List available structural fact and literal categories with counts, optionally scoped by path.
pub fn list_structural_fact_categories_scoped(
    conn: &Connection,
    path_filter: Option<&str>,
) -> Result<Vec<(String, usize)>, QueryError> {
    let norm_path = path_filter
        .map(|p| {
            p.replace('\\', "/")
                .trim_start_matches("./")
                .trim_matches('/')
                .to_string()
        })
        .filter(|p| !p.is_empty());
    let dir_prefix = norm_path
        .as_deref()
        .map(|p| format!("{}/%", escape_like(p)));

    let mut categories = Vec::new();

    let sql = "SELECT pattern_id, COUNT(*) AS cnt FROM structural_facts
               WHERE (:path IS NULL OR replace(path, '\\', '/') = :path COLLATE NOCASE OR replace(path, '\\', '/') LIKE :dir_prefix ESCAPE '\\')
               GROUP BY pattern_id ORDER BY cnt DESC";
    let mut stmt = conn.prepare(sql)?;
    let rows = stmt.query_map(
        rusqlite::named_params! {
            ":path": norm_path.as_deref(),
            ":dir_prefix": dir_prefix.as_deref(),
        },
        |row| Ok((row.get::<_, String>(0)?, row.get::<_, i64>(1)? as usize)),
    )?;
    for r in rows {
        categories.push(r?);
    }

    let lit_sql = "SELECT kind, COUNT(*) AS cnt FROM literals
                   WHERE (:path IS NULL OR replace(path, '\\', '/') = :path COLLATE NOCASE OR replace(path, '\\', '/') LIKE :dir_prefix ESCAPE '\\')
                   GROUP BY kind ORDER BY cnt DESC";
    let mut lit_stmt = conn.prepare(lit_sql)?;
    let lit_rows = lit_stmt.query_map(
        rusqlite::named_params! {
            ":path": norm_path.as_deref(),
            ":dir_prefix": dir_prefix.as_deref(),
        },
        |row| Ok((row.get::<_, String>(0)?, row.get::<_, i64>(1)? as usize)),
    )?;
    for r in lit_rows {
        categories.push(r?);
    }

    Ok(categories)
}

/// List all available structural fact and literal categories with counts.
pub fn list_structural_fact_categories(
    conn: &Connection,
) -> Result<Vec<(String, usize)>, QueryError> {
    list_structural_fact_categories_scoped(conn, None)
}

/// Find type facts for a symbol.
pub fn find_type_facts(conn: &Connection, symbol_id: &str) -> Result<Vec<TypeFact>, QueryError> {
    let has_table: bool = conn
        .query_row(
            "SELECT 1 FROM sqlite_master WHERE type='table' AND name='type_facts'",
            [],
            |_| Ok(true),
        )
        .unwrap_or(false);
    if !has_table {
        return Ok(Vec::new());
    }

    let mut stmt = conn.prepare(
        "SELECT type_fact_id, symbol_id, language, resolved_type, generic_params_json
         FROM type_facts
         WHERE symbol_id = ?1",
    )?;

    let rows = stmt.query_map(params![symbol_id], |row| {
        Ok(TypeFact {
            type_fact_id: row.get(0)?,
            symbol_id: row.get(1)?,
            language: row.get(2)?,
            resolved_type: row.get(3)?,
            generic_params: row.get(4)?,
        })
    })?;

    let mut results = Vec::new();
    for r in rows {
        results.push(r?);
    }
    Ok(results)
}

/// The workspace type definitions named by the type facts of `target` and of its parameters and
/// locals, one per name, as `signature (path:line)`. A name with no type definition in the
/// index, such as `Result` or `String`, is left out. Same-file definitions win, then shorter paths.
pub fn find_related_types(
    conn: &Connection,
    target: &Symbol,
    limit: usize,
) -> Result<Vec<String>, QueryError> {
    if !has_table(conn, "type_facts") {
        return Ok(Vec::new());
    }
    let mut stmt = conn.prepare(
        "SELECT d.name, d.kind, d.signature, d.path, d.start_line
         FROM symbols d
         WHERE d.kind IN ('struct', 'enum', 'trait', 'class', 'interface', 'type', 'union')
           AND d.name IN (
               SELECT t.resolved_type FROM type_facts t
               JOIN symbols s ON s.symbol_id = t.symbol_id
               WHERE s.symbol_id = ?1 OR s.parent_symbol_id = ?1
           )
         ORDER BY d.path = ?2 DESC, length(d.path), d.path, d.start_line",
    )?;
    let rows = stmt.query_map(params![target.symbol_id, target.path], |row| {
        Ok((
            row.get::<_, String>(0)?,
            row.get::<_, String>(1)?,
            row.get::<_, Option<String>>(2)?,
            row.get::<_, String>(3)?,
            row.get::<_, i64>(4)?,
        ))
    })?;

    let mut seen = std::collections::HashSet::new();
    let mut types = Vec::new();
    for row in rows {
        let (name, kind, signature, path, line) = row?;
        if types.len() >= limit {
            break;
        }
        if !seen.insert(name.clone()) {
            continue;
        }
        let signature = signature.unwrap_or_else(|| format!("{kind} {name}"));
        types.push(format!(
            "{} ({path}:{line})",
            type_declaration_line(&signature)
        ));
    }
    Ok(types)
}

/// A type signature without leading attributes, on one line of at most 120 characters.
fn type_declaration_line(signature: &str) -> String {
    let mut declaration = signature;
    while declaration.starts_with("#[")
        && let Some((_, rest)) = declaration.split_once("] ")
    {
        declaration = rest;
    }
    let line = declaration.split_whitespace().collect::<Vec<_>>().join(" ");
    if line.chars().count() <= 120 {
        return line;
    }
    line.chars().take(119).collect::<String>() + "…"
}

/// True when a repository-relative path looks like a test file. Directory rules and file-name
/// rules are kept apart: a `test`, `tests`, `autotests`, or `__tests__` directory anywhere
/// including the repository root, or a file name that starts with Qt's `tst_`, starts with
/// `test_` in Python or Ruby, contains `_test.`,
/// `.test.`, or `.spec.`, is exactly `test.rs` or `tests.rs`, or ends with the C# `Tests.cs`
/// (case-sensitive, so `Contests.cs` is a production file).
/// A name that reads like a test (`test_run`, `TestRoutes`, `runSpec`), for a symbol in a test file
/// that julie did not flag and that is not in a test class: an app factory such as `create_app` in
/// `tests/test_apps` is not a test, but a helper method of a test class leads to that class.
/// The SQL mirror of [`names_a_test`] for the row `alias`; `LIKE` ignores ASCII case.
fn names_a_test_sql(alias: &str) -> String {
    ["test%", "%test", "%tests", "%spec", "%specs"]
        .map(|pattern| format!("{alias}.name LIKE '{pattern}'"))
        .join(" OR ")
}

/// The test a related-test row stands for: the row, or the test that encloses a helper defined
/// inside it (`@app.url_defaults def add_language_code` in `test_url_processors`).
fn enclosing_test(conn: &Connection, row: Symbol) -> Result<Symbol, QueryError> {
    if row.is_test || row.test_container || names_a_test(&row.name) {
        return Ok(row);
    }
    let parent = row
        .parent_symbol_id
        .as_deref()
        .map(|id| get_symbol_by_id(conn, id))
        .transpose()?
        .flatten();
    Ok(match parent {
        Some(parent) if matches!(parent.kind.as_str(), "function" | "method") => parent,
        _ => row,
    })
}

pub(crate) fn names_a_test(name: &str) -> bool {
    let lowered = name.to_ascii_lowercase();
    lowered.starts_with("test")
        || ["test", "tests", "spec", "specs"]
            .iter()
            .any(|suffix| lowered.ends_with(suffix))
}

pub fn is_test_path(path: &str) -> bool {
    let p = path.replace('\\', "/");
    let cut = p.rfind('/').map_or(0, |i| i + 1);
    let directories = format!("/{}/", p[..cut].to_lowercase());
    let file_name = &p[cut..];
    let lower_name = file_name.to_lowercase();
    directories.contains("/test/")
        || directories.contains("/tests/")
        || directories.contains("/autotests/")
        || directories.contains("/__tests__/")
        || lower_name.starts_with("tst_")
        || (lower_name.starts_with("test_")
            && (lower_name.ends_with(".py") || lower_name.ends_with(".rb")))
        || lower_name.contains("_test.")
        || lower_name.contains(".test.")
        || lower_name.contains(".spec.")
        || lower_name == "test.rs"
        || lower_name == "tests.rs"
        || file_name.ends_with("Tests.cs")
}

/// SQL boolean over `alias.path` that mirrors [`is_test_path`] rule for rule.
///
/// `test_path_rule_and_its_sql_mirror_agree_on_every_path` runs both forms over one path list so
/// the two cannot drift apart.
///
/// Every rule needs `test`, `spec`, or `tst_` in the path, so a cheap substring test guards the
/// rules and lets most rows skip the path split. Without the guard the split costs about seven
/// times more over a half-million rows.
pub(crate) fn test_path_predicate(alias: &str) -> String {
    let guard = format!(
        "(lower({alias}.path) LIKE '%test%' OR lower({alias}.path) LIKE '%spec%' OR lower({alias}.path) LIKE '%tst\\_%' ESCAPE '\\')"
    );
    let p = format!("replace({alias}.path, '\\', '/')");
    let directories = format!("'/' || lower(rtrim({p}, replace({p}, '/', ''))) || '/'");
    let file_name = format!("replace({p}, rtrim({p}, replace({p}, '/', '')), '')");
    let lower_name = format!("lower({file_name})");
    let like = |subject: &String, pattern: &str| format!("{subject} LIKE '{pattern}' ESCAPE '\\'");
    let clauses = [
        like(&directories, "%/test/%"),
        like(&directories, "%/tests/%"),
        like(&directories, "%/autotests/%"),
        like(&directories, "%/\\_\\_tests\\_\\_/%"),
        like(&lower_name, "tst\\_%"),
        like(&lower_name, "test\\_%.py"),
        like(&lower_name, "test\\_%.rb"),
        like(&lower_name, "%\\_test.%"),
        like(&lower_name, "%.test.%"),
        like(&lower_name, "%.spec.%"),
        format!("{lower_name} = 'test.rs'"),
        format!("{lower_name} = 'tests.rs'"),
        format!("{file_name} GLOB '*Tests.cs'"),
    ]
    .join(" OR ");
    format!("({guard} AND ({clauses}))")
}

/// Compute blast radius and likely tests for given seed symbols or seed file paths.
/// Recursively walks reverse reachability (transitive callers) up to `max_depth` in SQLite.
pub fn compute_blast_radius_scoped(
    conn: &Connection,
    seed_symbols: &[&str],
    symbol_path_filter: Option<&str>,
    seed_paths: &[&str],
    max_depth: usize,
    limit: usize,
) -> Result<BlastRadiusResult, QueryError> {
    compute_blast_radius_scoped_with_ids(
        conn,
        seed_symbols,
        &[],
        symbol_path_filter,
        seed_paths,
        max_depth,
        limit,
    )
}

const MAX_BLAST_DEPTH: usize = 5;

/// Compute blast radius with exact current-index IDs kept as traversal seeds.
pub fn compute_blast_radius_scoped_with_ids(
    conn: &Connection,
    seed_symbols: &[&str],
    seed_ids: &[&str],
    symbol_path_filter: Option<&str>,
    seed_paths: &[&str],
    max_depth: usize,
    limit: usize,
) -> Result<BlastRadiusResult, QueryError> {
    validate_result_limit(limit)?;
    let max_depth = max_depth.min(MAX_BLAST_DEPTH);
    let mut resolved_seed_symbols = seed_symbols
        .iter()
        .map(|name| {
            get_symbol_by_name(conn, name, symbol_path_filter)?.ok_or_else(|| {
                let (workspace, hint) = symbol_not_found_parts(conn, name, symbol_path_filter);
                QueryError::SymbolNotFound {
                    name: (*name).to_string(),
                    workspace,
                    hint,
                }
            })
        })
        .collect::<Result<Vec<_>, _>>()?;
    for id in seed_ids {
        let symbol = get_symbol_by_id(conn, id)?.ok_or_else(|| {
            let workspace = workspace_name(conn);
            QueryError::SymbolNotFound {
                name: format!("symbol id {id}"),
                workspace,
                hint: "Run lookup_symbol again and select a current id.".to_string(),
            }
        })?;
        resolved_seed_symbols.push(symbol);
    }
    let mut seeds = Vec::new();
    let seed_type = if (!seed_symbols.is_empty() || !seed_ids.is_empty()) && !seed_paths.is_empty()
    {
        for s in seed_symbols {
            seeds.push(s.to_string());
        }
        for id in seed_ids {
            seeds.push(id.to_string());
        }
        for p in seed_paths {
            seeds.push(p.to_string());
        }
        "mixed".to_string()
    } else if !seed_symbols.is_empty() || !seed_ids.is_empty() {
        for s in seed_symbols {
            seeds.push(s.to_string());
        }
        for id in seed_ids {
            seeds.push(id.to_string());
        }
        "symbol".to_string()
    } else if !seed_paths.is_empty() {
        for p in seed_paths {
            seeds.push(p.to_string());
        }
        "file".to_string()
    } else {
        return Ok(BlastRadiusResult {
            seed_type: "none".to_string(),
            seeds: Vec::new(),
            likely_tests: Vec::new(),
            impacted_symbols: Vec::new(),
            likely_tests_truncated: false,
            impacted_symbols_truncated: false,
            traversal_ceiling_reached: false,
            limit_at_maximum: false,
            likely_tests_found: 0,
            impacted_symbols_found: 0,
            stopped_at_depth: None,
        });
    };

    let mut where_clauses = Vec::new();
    let mut params_vec: Vec<rusqlite::types::Value> = Vec::new();

    if !resolved_seed_symbols.is_empty() {
        let placeholders: Vec<String> = (1..=resolved_seed_symbols.len())
            .map(|i| format!("?{}", i))
            .collect();
        where_clauses.push(format!("symbol_id IN ({})", placeholders.join(", ")));
        for symbol in &resolved_seed_symbols {
            params_vec.push(rusqlite::types::Value::Text(symbol.symbol_id.clone()));
        }
    }

    if !seed_paths.is_empty() {
        let mut path_conds = Vec::new();
        for p in seed_paths.iter() {
            let raw = p
                .replace('\\', "/")
                .trim_start_matches("./")
                .trim_matches('/')
                .to_string();
            let exact_idx = params_vec.len() + 1;
            params_vec.push(rusqlite::types::Value::Text(raw.clone()));
            let dir_pattern = format!("{}/%", escape_like(&raw));
            let like_idx = params_vec.len() + 1;
            params_vec.push(rusqlite::types::Value::Text(dir_pattern));
            path_conds.push(format!(
                "replace(path, '\\', '/') = ?{exact_idx} COLLATE NOCASE OR replace(path, '\\', '/') LIKE ?{like_idx} ESCAPE '\\'"
            ));
        }
        where_clauses.push(format!("({})", path_conds.join(" OR ")));
    }

    let seed_condition = where_clauses.join(" OR ");
    let max_depth_idx = params_vec.len() + 1;
    // One level past `max_depth` tells a walk that a deeper call would extend from one that ran out.
    let walk_depth = if max_depth < MAX_BLAST_DEPTH {
        max_depth + 1
    } else {
        max_depth
    };
    params_vec.push(rusqlite::types::Value::Integer(walk_depth as i64));

    let mut traversal_ceiling_reached = false;
    let mut callers_beyond_depth = false;

    let has_relationships: bool = conn
        .query_row(
            "SELECT 1 FROM sqlite_master WHERE type='table' AND name='relationships'",
            [],
            |_| Ok(true),
        )
        .unwrap_or(false);

    let has_pending: bool = conn
        .query_row(
            "SELECT 1 FROM sqlite_master WHERE type='table' AND name='pending_relationships'",
            [],
            |_| Ok(true),
        )
        .unwrap_or(false);

    let mut likely_tests = Vec::new();
    let mut impacted_symbols = Vec::new();
    let mut seen_test_keys = HashSet::new();

    let mut recursive_branches = Vec::new();

    if has_relationships {
        recursive_branches.push(format!(
            "SELECT r.from_symbol_id, iw.depth + 1, iw.via
             FROM relationships r
             JOIN impact_walk iw ON r.to_symbol_id = iw.symbol_id
             JOIN symbols s_from ON r.from_symbol_id = s_from.symbol_id
             WHERE iw.depth < ?{max_depth_idx}
               AND s_from.kind NOT IN ({LOW_SIGNAL_KINDS_SQL})"
        ));
    }

    if has_pending {
        let (parent_join, ns_condition) = if conn
            .query_row(
                "SELECT 1 FROM pragma_table_info('pending_relationships') WHERE name='target_namespace_json'",
                [],
                |_| Ok(true),
            )
            .unwrap_or(false)
        {
            (
                "LEFT JOIN symbols s_target_parent ON s_target.parent_symbol_id = s_target_parent.symbol_id
            LEFT JOIN symbols s_from ON p.from_symbol_id = s_from.symbol_id",
                format!("AND {pred}", pred = pending_target_predicate(conn, "s_target", "s_target_parent")),
            )
        } else {
            ("", String::new())
        };

        recursive_branches.push(format!(
            "SELECT p.from_symbol_id, iw.depth + 1, iw.via
             FROM pending_relationships p
             JOIN symbols s_target ON p.target_terminal_name = s_target.name
             JOIN impact_walk iw ON s_target.symbol_id = iw.symbol_id
             {parent_join}
             WHERE iw.depth < ?{max_depth_idx}
               AND s_target.kind NOT IN ({LOW_SIGNAL_KINDS_SQL})
               {ns_condition}"
        ));
    }

    // A constructor runs wherever its class is built: `Flask()` reaches `Flask.__init__`. Rows
    // reached only this way (`via` 1) sort last: nearly every test builds the app.
    if has_relationships {
        recursive_branches.push(format!(
            "SELECT r.from_symbol_id, iw.depth + 1, 1
             FROM impact_walk iw
             CROSS JOIN symbols ctor ON ctor.symbol_id = iw.symbol_id
             JOIN relationships r ON r.to_symbol_id = ctor.parent_symbol_id
             WHERE iw.depth < ?{max_depth_idx}
               AND ctor.kind = 'constructor'
               AND +r.kind = 'calls'"
        ));
    }
    if has_pending && has_pending_namespace_column(conn) {
        recursive_branches.push(format!(
            "SELECT p.from_symbol_id, iw.depth + 1, 1
             FROM impact_walk iw
             CROSS JOIN symbols ctor ON ctor.symbol_id = iw.symbol_id
             CROSS JOIN symbols s_target ON s_target.symbol_id = ctor.parent_symbol_id
             JOIN pending_relationships p ON p.target_terminal_name = s_target.name
             LEFT JOIN symbols s_target_parent ON s_target.parent_symbol_id = s_target_parent.symbol_id
             LEFT JOIN symbols s_from ON p.from_symbol_id = s_from.symbol_id
             WHERE iw.depth < ?{max_depth_idx}
               AND ctor.kind = 'constructor'
               AND s_target.kind IN ('class', 'struct', 'record')
               AND +p.kind = 'calls'
               AND {pred}",
            pred = pending_target_predicate(conn, "s_target", "s_target_parent")
        ));
    }

    let lifecycle = if has_column(conn, "symbols", "test_lifecycle") {
        "COALESCE(s.test_lifecycle, 0) != 0"
    } else {
        "0"
    };
    let mut fixtures = Vec::new();
    let mut setups = Vec::new();
    let mut constructor_only_test_keys = HashSet::new();
    let mut non_constructor_test_keys = HashSet::new();
    if !recursive_branches.is_empty() {
        let recursive_sql = recursive_branches.join("\n UNION \n");
        let not_documentation = not_documentation(conn, "s");
        let sql = format!(
            "WITH RECURSIVE impact_walk(symbol_id, depth, via) AS (
                SELECT symbol_id, 0, 0
                FROM symbols
                WHERE ({seed_condition})
                  AND kind NOT IN ({LOW_SIGNAL_KINDS_SQL})

                UNION

                {recursive_sql}
            )
            SELECT s.name, s.kind, s.path, s.start_line, s.is_test, s.test_container,
                   MIN(iw.depth) as min_depth, MIN(iw.via) as min_via,
                   {lifecycle} AS is_fixture, s.parent_symbol_id,
                   EXISTS (SELECT 1 FROM symbols owner
                           WHERE owner.symbol_id = s.parent_symbol_id
                             AND COALESCE(owner.test_container, 0) != 0) AS in_test_class,
                   s.language, s.signature
            FROM impact_walk iw
            CROSS JOIN symbols s ON iw.symbol_id = s.symbol_id
            WHERE s.kind NOT IN ({LOW_SIGNAL_KINDS_SQL})
              AND {not_documentation}
            GROUP BY s.symbol_id, s.name, s.kind, s.path, s.start_line, s.is_test, s.test_container
            HAVING MIN(iw.depth) > 0
            ORDER BY MIN(iw.depth) > {max_depth} ASC, MIN(iw.via) ASC, min_depth ASC, s.path ASC, s.name ASC
            LIMIT 201"
        );

        let mut stmt = conn.prepare(&sql)?;
        let param_refs: Vec<&dyn rusqlite::ToSql> = params_vec
            .iter()
            .map(|v| v as &dyn rusqlite::ToSql)
            .collect();

        let rows = stmt.query_map(param_refs.as_slice(), |row| {
            Ok((
                row.get::<_, String>(0)?,
                row.get::<_, String>(1)?,
                row.get::<_, String>(2)?,
                row.get::<_, i64>(3)? as usize,
                row.get::<_, bool>(4)?,
                row.get::<_, bool>(5)?,
                row.get::<_, i64>(6)? as usize,
                row.get::<_, i64>(7)? as usize,
                row.get::<_, bool>(8)?,
                row.get::<_, Option<String>>(9)?,
                row.get::<_, bool>(10)?,
                row.get::<_, String>(11)?,
                row.get::<_, Option<String>>(12)?,
            ))
        })?;

        for (index, r) in rows.enumerate() {
            let (
                name,
                kind,
                raw_path,
                line,
                is_test,
                test_container,
                depth,
                min_via,
                is_fixture,
                parent,
                in_test_class,
                language,
                signature,
            ) = r?;
            if depth > max_depth {
                callers_beyond_depth = true;
                break;
            }
            if index == 200 {
                traversal_ceiling_reached = true;
                break;
            }
            let path = raw_path.replace('\\', "/");
            let is_test_target = is_test
                || test_container
                || is_fixture
                || (is_test_path(&path) && (in_test_class || names_a_test(&name)));
            let constructor_only = min_via > 0;
            if is_test_target {
                let key = format!("{}:{}", path, line);
                let test_key = (path.clone(), line);
                if constructor_only {
                    constructor_only_test_keys.insert(test_key);
                } else {
                    non_constructor_test_keys.insert(test_key);
                }
                if seen_test_keys.insert(key) {
                    let lowered = name.to_ascii_lowercase();
                    let pytest_fixture = path.ends_with(".py")
                        && !["setup", "teardown", "asyncsetup", "asyncteardown"]
                            .iter()
                            .any(|prefix| lowered.starts_with(prefix));
                    let caller = if depth == 1 {
                        "direct caller".to_string()
                    } else {
                        format!("indirect caller [depth {depth}]")
                    };
                    let reason = if is_fixture && pytest_fixture {
                        fixtures.push((name.clone(), path.clone(), line, constructor_only));
                        format!("fixture ({caller})")
                    } else if is_fixture {
                        if let Some(class_id) = parent {
                            setups.push((
                                name.clone(),
                                class_id,
                                path.clone(),
                                line,
                                constructor_only,
                            ));
                        }
                        format!("setup ({caller})")
                    } else {
                        caller
                    };
                    likely_tests.push(TestTarget {
                        name,
                        path,
                        line,
                        reason,
                    });
                }
            } else {
                let enclosing_symbol_name =
                    if is_generated_lambda_name(&name, &language, &kind, signature.as_deref()) {
                        match parent.as_deref() {
                            Some(parent_id) => {
                                get_symbol_by_id(conn, parent_id)?.map(|parent| parent.name)
                            }
                            None => None,
                        }
                    } else {
                        None
                    };
                impacted_symbols.push(ImpactedSymbol {
                    name,
                    kind,
                    path,
                    line,
                    depth,
                    enclosing_symbol_name,
                });
            }
        }
    }

    // A fixture or setup member is not a test to run; the tests it serves stand in for it.
    let mut replaced = HashSet::new();
    for (fixture, fixture_path, line, constructor_only) in &fixtures {
        let shown = fixture_name(conn, fixture, fixture_path);
        let users = fixture_users(conn, fixture, fixture_path)?;
        if !users.is_empty() {
            replaced.insert((fixture_path.clone(), *line));
        }
        for test in users {
            let test_key = (test.path.clone(), test.line);
            if *constructor_only {
                constructor_only_test_keys.insert(test_key);
            } else {
                non_constructor_test_keys.insert(test_key);
            }
            if seen_test_keys.insert(format!("{}:{}", test.path, test.line)) {
                likely_tests.push(TestTarget {
                    reason: format!("uses fixture `{shown}`"),
                    ..test
                });
            }
        }
    }

    for (setup, class_id, setup_path, line, constructor_only) in &setups {
        let tests = tests_in_class(conn, class_id)?;
        if !tests.is_empty() {
            replaced.insert((setup_path.clone(), *line));
        }
        for test in tests {
            let test_key = (test.path.clone(), test.line);
            if *constructor_only {
                constructor_only_test_keys.insert(test_key);
            } else {
                non_constructor_test_keys.insert(test_key);
            }
            if seen_test_keys.insert(format!("{}:{}", test.path, test.line)) {
                likely_tests.push(TestTarget {
                    reason: format!("setup `{setup}` runs before it"),
                    ..test
                });
            }
        }
    }
    likely_tests.retain(|test| !replaced.contains(&(test.path.clone(), test.line)));
    constructor_only_test_keys.retain(|key| !non_constructor_test_keys.contains(key));

    qualify_test_methods(conn, &mut likely_tests)?;
    // A whole-file target often reaches many unrelated tests only because they construct an app;
    // keep that coverage visible while collapsing each affected test file to one run target.
    if seed_type == "file" {
        let mut constructor_tests_by_file =
            std::collections::BTreeMap::<String, Vec<TestTarget>>::new();
        let mut other_tests = Vec::with_capacity(likely_tests.len());
        for test in likely_tests.drain(..) {
            if constructor_only_test_keys.contains(&(test.path.clone(), test.line)) {
                constructor_tests_by_file
                    .entry(test.path.clone())
                    .or_default()
                    .push(test);
            } else {
                other_tests.push(test);
            }
        }
        for (path, mut tests) in constructor_tests_by_file {
            if tests.len() > 1 {
                let count = tests.len();
                other_tests.push(TestTarget {
                    name: path.clone(),
                    path,
                    line: 1,
                    reason: format!("constructor-only callers ({count} targets)"),
                });
            } else if let Some(test) = tests.pop() {
                other_tests.push(test);
            }
        }
        likely_tests = other_tests;
    }
    let likely_tests_found = likely_tests.len();
    let impacted_symbols_found = impacted_symbols.len();
    let likely_tests_truncated = likely_tests.len() > limit;
    let impacted_symbols_truncated = impacted_symbols.len() > limit;
    if likely_tests.len() > limit {
        likely_tests.truncate(limit);
    }
    if impacted_symbols.len() > limit {
        impacted_symbols.truncate(limit);
    }

    Ok(BlastRadiusResult {
        seed_type,
        seeds,
        likely_tests,
        impacted_symbols,
        likely_tests_truncated,
        impacted_symbols_truncated,
        traversal_ceiling_reached,
        limit_at_maximum: limit >= MAX_RESULT_LIMIT,
        likely_tests_found,
        impacted_symbols_found,
        stopped_at_depth: callers_beyond_depth.then_some(max_depth),
    })
}

/// Prefixes a test method with its class, as `TestRoutes::test_simple`, so the name selects the
/// test the way pytest and most runners address it.
fn qualify_test_methods(conn: &Connection, tests: &mut [TestTarget]) -> Result<(), QueryError> {
    let mut stmt = conn.prepare(
        "SELECT class.name FROM symbols t
         JOIN symbols class ON class.symbol_id = t.parent_symbol_id
         WHERE replace(t.path, '\\', '/') = ?1 AND t.start_line = ?2 AND t.name = ?3
           AND class.kind = 'class'
         LIMIT 1",
    )?;
    for test in tests.iter_mut().filter(|test| test.name != test.path) {
        if let Some(class) = stmt
            .query_row(params![test.path, test.line as i64, test.name], |row| {
                row.get::<_, String>(0)
            })
            .optional()?
        {
            test.name = format!("{class}::{}", test.name);
        }
    }
    Ok(())
}

/// The name tests use for the fixture function `function` in `path`: the `name=` argument of its
/// `@pytest.fixture(...)` decorator when it has one, else the function name.
fn fixture_name(conn: &Connection, function: &str, path: &str) -> String {
    let signature: Option<String> = conn
        .query_row(
            "SELECT signature FROM symbols
             WHERE name = ?1 AND replace(path, '\\', '/') = ?2 AND signature LIKE '%fixture(%'
             LIMIT 1",
            params![function, path],
            |row| row.get(0),
        )
        .optional()
        .ok()
        .flatten();
    signature
        .as_deref()
        .and_then(declared_fixture_name)
        .unwrap_or_else(|| function.to_string())
}

/// The `name=` argument of a `fixture(...)` decorator in `signature`.
fn declared_fixture_name(signature: &str) -> Option<String> {
    let args = &signature[signature.find("fixture(")? + "fixture(".len()..];
    let args = &args[..args.find(')').unwrap_or(args.len())];
    let value = args[args.find("name=")? + "name=".len()..].trim_start();
    let quote = value.chars().next().filter(|c| *c == '"' || *c == '\'')?;
    let end = value[1..].find(quote)?;
    Some(value[1..1 + end].to_string())
}

/// The test functions that take the fixture `name` as a parameter: in the fixture's own file,
/// or anywhere under the directory of the `conftest.py` that defines it. A fixture that takes it
/// is not a test, so its own users are listed in its place.
fn fixture_users(
    conn: &Connection,
    name: &str,
    fixture_path: &str,
) -> Result<Vec<TestTarget>, QueryError> {
    let lifecycle = if has_column(conn, "symbols", "test_lifecycle") {
        "COALESCE(t.test_lifecycle, 0) != 0"
    } else {
        "0"
    };
    let mut stmt = conn.prepare(&format!(
        "SELECT DISTINCT t.name, t.path, t.start_line, {lifecycle}
         FROM symbols parameter
         JOIN symbols t ON t.symbol_id = parameter.parent_symbol_id
         WHERE parameter.name = ?1
           AND (parameter.kind = 'parameter'
                OR (parameter.kind = 'variable' AND parameter.start_byte < t.body_start_byte))
           AND t.is_test = 1
           AND replace(t.path, '\\', '/') LIKE ?2 ESCAPE '\\'
         ORDER BY t.path, t.start_line
         LIMIT 2000"
    ))?;
    // ponytail: the callers rank these rows by name, so the cap must hold a whole test suite's
    // users of one fixture; raise it if a suite passes 2,000 users of a single fixture.
    let mut users = Vec::new();
    let mut listed = HashSet::new();
    let mut visited = HashSet::new();
    let mut pending = vec![(name.to_string(), fixture_path.to_string())];
    while let Some((function, fixture_path)) = pending.pop() {
        if !visited.insert((function.clone(), fixture_path.clone())) {
            continue;
        }
        let name = fixture_name(conn, &function, &fixture_path);
        let scope = match fixture_path.rsplit_once('/') {
            Some((dir, "conftest.py")) => format!("{}/%", escape_like(dir)),
            None if fixture_path == "conftest.py" => "%".to_string(),
            _ => escape_like(&fixture_path),
        };
        let rows = stmt.query_map(params![name, scope], |row| {
            Ok((
                TestTarget {
                    name: row.get(0)?,
                    path: row.get::<_, String>(1)?.replace('\\', "/"),
                    line: row.get::<_, i64>(2)? as usize,
                    reason: String::new(),
                },
                row.get::<_, bool>(3)?,
            ))
        })?;
        for row in rows {
            let (user, is_fixture) = row?;
            if is_fixture {
                pending.push((user.name, user.path));
            } else if listed.insert((user.path.clone(), user.line)) {
                users.push(user);
            }
        }
    }
    Ok(users)
}

/// The tests declared in the class `class_id`, without its setup and teardown members.
fn tests_in_class(conn: &Connection, class_id: &str) -> Result<Vec<TestTarget>, QueryError> {
    let not_lifecycle = if has_column(conn, "symbols", "test_lifecycle") {
        "COALESCE(test_lifecycle, 0) = 0"
    } else {
        "1"
    };
    let mut stmt = conn.prepare(&format!(
        "SELECT name, path, start_line FROM symbols
         WHERE parent_symbol_id = ?1 AND is_test = 1 AND {not_lifecycle}
         ORDER BY start_line"
    ))?;
    let rows = stmt.query_map([class_id], |row| {
        Ok(TestTarget {
            name: row.get(0)?,
            path: row.get::<_, String>(1)?.replace('\\', "/"),
            line: row.get::<_, i64>(2)? as usize,
            reason: String::new(),
        })
    })?;
    Ok(rows.collect::<Result<_, _>>()?)
}

/// Compute blast radius and likely tests for given seed symbols or seed file paths.
pub fn compute_blast_radius(
    conn: &Connection,
    seed_symbols: &[&str],
    seed_paths: &[&str],
    max_depth: usize,
    limit: usize,
) -> Result<BlastRadiusResult, QueryError> {
    compute_blast_radius_scoped(conn, seed_symbols, None, seed_paths, max_depth, limit)
}

#[cfg(test)]
mod tests {
    #[test]
    fn a_type_declaration_drops_attributes_and_fits_on_one_short_line() {
        assert_eq!(
            super::type_declaration_line(
                "#[derive(Debug)] #[serde(tag = \"t\")] pub struct Workspace"
            ),
            "pub struct Workspace"
        );
        let record = format!(
            "public sealed record Node(\n{})",
            "    string Id,\n".repeat(20)
        );
        let line = super::type_declaration_line(&record);
        assert!(
            line.starts_with("public sealed record Node( string Id, string Id,"),
            "{line}"
        );
        assert_eq!(line.chars().count(), 120);
        assert!(line.ends_with('…'));
    }

    #[test]
    fn a_mount_names_what_it_mounts_and_where() {
        let mount = |capture: &str, json: &str| {
            super::mount_display(capture, &serde_json::from_str(json).unwrap())
        };
        assert_eq!(
            mount(
                "blueprint_registration",
                r#"{"framework":"flask","mount_target":"auth.bp","query_family":"framework"}"#
            )
            .as_deref(),
            Some("blueprint_registration `auth.bp`")
        );
        assert_eq!(
            mount(
                "mount",
                r#"{"mount_path":"/jobs","mount_target":"Sidekiq::Web","normalized_mount_path":"/admin/jobs","query_family":"framework"}"#
            )
            .as_deref(),
            Some("mount `Sidekiq::Web` at /admin/jobs")
        );
        assert_eq!(
            mount(
                "property",
                r#"{"mount_target":"x","query_family":"metadata"}"#
            ),
            None
        );
    }

    #[test]
    fn a_route_shows_its_source_template_and_keeps_a_joined_prefix() {
        let route = |json: &str| super::route_display(&serde_json::from_str(json).unwrap());
        assert_eq!(
            route(
                r#"{"verb":"GET","route_template":"/<int:id>/update","normalized_route_template":"/:id/update"}"#
            ),
            Some("GET /<int:id>/update".into())
        );
        assert_eq!(
            route(
                r#"{"verb":"POST","route_template":"/register","effective_route_template":"/auth/register","normalized_route_template":"/auth/register"}"#
            ),
            Some("POST /auth/register".into())
        );
        assert_eq!(
            route(
                r#"{"verb":"GET","route_template":"/users/{id}","normalized_route_template":"/api/v1/users/:id"}"#
            ),
            Some("GET /api/v1/users/:id".into())
        );
        assert_eq!(route(r#"{"key":"x"}"#), None);
    }

    #[test]
    fn a_fixture_declared_with_a_name_is_used_by_that_name() {
        assert_eq!(
            super::declared_fixture_name("@pytest.fixture(name=\"async_app\") def _async_app()"),
            Some("async_app".into())
        );
        assert_eq!(
            super::declared_fixture_name(
                "@pytest.fixture(scope='session', name='db') def make_db()"
            ),
            Some("db".into())
        );
        assert_eq!(
            super::declared_fixture_name("@pytest.fixture def app()"),
            None
        );
        assert_eq!(
            super::declared_fixture_name("@pytest.fixture(params=[1]) def f(name='x')"),
            None
        );
    }

    #[test]
    fn result_limit_rejects_values_above_the_shared_ceiling() {
        assert!(validate_result_limit(MAX_RESULT_LIMIT).is_ok());
        assert!(matches!(
            validate_result_limit(usize::MAX),
            Err(QueryError::InvalidResultLimit(usize::MAX))
        ));
    }

    #[test]
    fn find_references_rejects_an_unbounded_limit_before_sql_execution() {
        let conn = Connection::open_in_memory().unwrap();

        assert!(matches!(
            find_references_scoped(&conn, "target", "callers", usize::MAX, false, None),
            Err(QueryError::InvalidResultLimit(usize::MAX))
        ));
    }

    use super::*;
    use crate::db::{ensure_fts_index, open_read_write};

    #[test]
    fn count_parse_diagnostics_counts_rows_for_one_file() {
        let dir = crate::safe_tempdir();
        let conn = open_read_write(&dir.path().join("parse_diagnostics.db")).unwrap();

        assert_eq!(count_parse_diagnostics(&conn, "src/lib.rs"), 0);

        conn.execute_batch(
            "CREATE TABLE parse_diagnostics (
                diagnostic_id TEXT, file_id TEXT, path TEXT, language TEXT, kind TEXT
            );
            INSERT INTO parse_diagnostics VALUES ('d1', 'f1', 'src/lib.rs', 'rust', 'error');
            INSERT INTO parse_diagnostics VALUES ('d2', 'f1', 'src/lib.rs', 'rust', 'error');
            INSERT INTO parse_diagnostics VALUES ('d3', 'f2', 'src/other.rs', 'rust', 'error');",
        )
        .unwrap();

        assert_eq!(count_parse_diagnostics(&conn, "src/lib.rs"), 2);
        assert_eq!(count_parse_diagnostics(&conn, "src\\lib.rs"), 2);
        assert_eq!(count_parse_diagnostics(&conn, "src/clean.rs"), 0);
    }

    #[test]
    fn test_sanitize_fts5_query() {
        let (and_q, or_q) = sanitize_fts5_query("parse tokens");
        assert_eq!(and_q, "(\"parse\"* AND \"tokens\"*) OR \"parsetokens\"*");
        assert_eq!(or_q, "\"parse\"* OR \"tokens\"* OR \"parsetokens\"*");

        let (and_q, or_q) = sanitize_fts5_query("  Option<T>  ");
        assert_eq!(and_q, "(\"Option\"* AND \"T\") OR \"OptionT\"*");
        assert_eq!(or_q, "\"Option\"* OR \"T\" OR \"OptionT\"*");

        let (and_q, or_q) = sanitize_fts5_query("   ");
        assert!(and_q.is_empty());
        assert!(or_q.is_empty());
    }

    #[test]
    fn sanitize_splits_case_boundaries_and_drops_stop_words() {
        let (and_q, or_q) = sanitize_fts5_query("ValidateSyntax");
        assert_eq!(
            and_q,
            "((\"Validate\"* \"Syntax\"*) OR \"ValidateSyntax\"*)"
        );
        assert_eq!(or_q, "\"Validate\"* OR \"Syntax\"* OR \"ValidateSyntax\"*");

        let (and_q, _) = sanitize_fts5_query("find tests related to a symbol");
        assert_eq!(
            and_q,
            "\"find\"* AND \"tests\"* AND \"related\"* AND \"symbol\"*"
        );

        let (and_q, or_q) = sanitize_fts5_query("parseHTTPResponse2");
        assert_eq!(
            and_q,
            "((\"parse\"* \"HTTP\"* \"Response\"* \"2\") OR \"parseHTTPResponse2\"*)"
        );
        assert!(or_q.ends_with("OR \"parseHTTPResponse2\"*"));

        let (and_q, _) = sanitize_fts5_query("validate_syntax");
        assert_eq!(
            and_q,
            "((\"validate\"* \"syntax\"*) OR \"validate_syntax\"*)"
        );

        let (and_q, _) = sanitize_fts5_query("isReady");
        assert_eq!(and_q, "((\"Ready\"*) OR \"isReady\"*)");

        let (and_q, _) = sanitize_fts5_query("before");
        assert_eq!(and_q, "\"before\"*");

        let (and_q, _) = sanitize_fts5_query("fooBar quux");
        assert_eq!(
            and_q,
            "(((\"foo\"* \"Bar\"*) OR \"fooBar\"*) AND \"quux\"*) OR \"fooBarquux\"*"
        );

        let (and_q, _) = sanitize_fts5_query("the for a");
        assert_eq!(and_q, "(\"the\"* AND \"for\"* AND \"a\") OR \"thefora\"*");
    }

    fn search_fixture(rows: &str) -> Connection {
        let conn = Connection::open_in_memory().unwrap();
        conn.execute_batch(&format!(
            "CREATE TABLE symbols (
                symbol_id TEXT PRIMARY KEY, file_id TEXT, path TEXT, language TEXT, name TEXT,
                kind TEXT, signature TEXT, doc_comment TEXT, visibility TEXT, parent_symbol_id TEXT,
                start_line INTEGER, start_column INTEGER, end_line INTEGER, end_column INTEGER,
                start_byte INTEGER, end_byte INTEGER, body_start_line INTEGER,
                body_start_column INTEGER, body_end_line INTEGER, body_end_column INTEGER,
                body_start_byte INTEGER, body_end_byte INTEGER, body_hash TEXT,
                semantic_group TEXT, is_test INTEGER, test_container INTEGER, content_type TEXT
            );
            INSERT INTO symbols VALUES {rows};"
        ))
        .unwrap();
        ensure_fts_index(&conn).unwrap();
        conn
    }

    fn code_row(id: &str, path: &str, language: &str, name: &str, doc: &str) -> String {
        format!(
            "('{id}', 'f_{id}', '{path}', '{language}', '{name}', 'function', 'fn {name}()', '{doc}', 'pub', NULL,
              10, 0, 20, 1, 100, 250, 12, 4, 19, 1, 120, 240, 'h_{id}', NULL, 0, 0, 'code')"
        )
    }

    fn doc_row(id: &str, name: &str, doc: &str) -> String {
        format!(
            "('{id}', 'f_{id}', 'docs/{id}.md', 'markdown', '{name}', 'module', '{name}', '{doc}', NULL, NULL,
              3, 0, 3, 1, 10, 40, NULL, NULL, NULL, NULL, NULL, NULL, 'h_{id}', NULL, 0, 0, 'documentation')"
        )
    }

    #[test]
    fn ranked_owner_context_recovers_members_and_enriches_existing_rows() {
        let mut rows = vec![
            code_row(
                "planner",
                "src/planner.rs",
                "rust",
                "EditPlanner",
                "Plans edit operations for a function body",
            ),
            code_row(
                "target",
                "src/planner.rs",
                "rust",
                "ReplaceBodyImplementation",
                "",
            ),
            code_row("existing", "src/planner.rs", "rust", "Body", ""),
            code_row(
                "unrelated",
                "src/planner.rs",
                "rust",
                "SaveConfiguration",
                "",
            ),
        ];
        // Full-text and short-name hits exhaust both ordinary admission branches.
        rows.extend((0..200).map(|i| {
            code_row(
                &format!("noise{i}"),
                "src/noise.rs",
                "rust",
                &format!("body{i:03}"),
                "edit planner body",
            )
        }));
        let conn = search_fixture(&rows.join(","));
        conn.execute_batch(
            "UPDATE symbols SET kind = 'class' WHERE symbol_id = 'planner';
             UPDATE symbols SET kind = 'method', parent_symbol_id = 'planner'
             WHERE symbol_id IN ('target', 'existing', 'unrelated');",
        )
        .unwrap();
        let ordinary =
            collect_search_candidates(&conn, "edit planner body", None, None, false, 200).unwrap();
        assert!(
            !ordinary
                .iter()
                .any(|c| c.result.symbol.symbol_id == "target")
        );
        assert!(
            ordinary
                .iter()
                .any(|c| c.result.symbol.symbol_id == "existing")
        );
        let found =
            fts_search_symbols_explained(&conn, "edit planner body", None, None, false, 200, true)
                .unwrap();
        for id in ["target", "existing"] {
            let members: Vec<_> = found.iter().filter(|r| r.symbol.symbol_id == id).collect();
            assert_eq!(members.len(), 1);
            let explain = members[0].explain.as_ref().unwrap();
            assert!(explain.branches.iter().any(|b| b == "owner_context"));
            for word in ["edit", "planner"] {
                assert!(
                    explain
                        .terms
                        .contains(&(word.into(), "owner_context".into(), 0.5))
                );
            }
            assert!(explain.terms.contains(&("body".into(), "name".into(), 3.0)));
        }
        assert!(!found.iter().any(|r| r.symbol.symbol_id == "unrelated"));
        for exact in ["EditPlanner", "Body"] {
            assert_eq!(search_names(&conn, exact)[0], exact);
        }
        let scoped = fts_search_symbols_scoped(
            &conn,
            "edit planner body",
            None,
            Some("src/noise.rs"),
            false,
            200,
        )
        .unwrap();
        assert!(scoped.iter().all(|r| r.symbol.path == "src/noise.rs"));
        let methods =
            fts_search_symbols_scoped(&conn, "edit planner body", Some("method"), None, false, 200)
                .unwrap();
        assert!(methods.iter().all(|r| r.symbol.kind == "method"));
        // A method-only query has no owner types in its ordinary top twenty.
        assert!(methods.iter().all(|r| r.explain.is_none()));
    }

    #[test]
    fn ranked_owner_expansion_is_bounded_and_preserves_member_filters() {
        let mut rows = Vec::new();
        for owner in 0..4 {
            rows.push(code_row(
                &format!("o{owner}"),
                "src/lib.rs",
                "rust",
                &format!("Planner{owner}"),
                "body planner",
            ));
            for member in 0..41 {
                rows.push(code_row(
                    &format!("o{owner}m{member:02}"),
                    "src/lib.rs",
                    "rust",
                    &format!("Body{member:02}"),
                    "",
                ));
            }
        }
        rows.push(doc_row("docowner", "PlannerDocs", "body planner"));
        rows.push(code_row(
            "docchild",
            "src/lib.rs",
            "rust",
            "BodyFromDocs",
            "",
        ));
        rows.push(code_row(
            "testchild",
            "tests/lib.rs",
            "rust",
            "BodyTest",
            "",
        ));
        rows.push(code_row(
            "linkchild",
            "docs/link.md",
            "markdown",
            "BodyLink",
            "",
        ));
        rows.push(code_row(
            "nestedchild",
            "src/lib.rs",
            "rust",
            "BodyNested",
            "",
        ));
        let conn = search_fixture(&rows.join(","));
        for owner in 0..4 {
            conn.execute(
                "UPDATE symbols SET kind = 'class' WHERE symbol_id = ?1",
                [format!("o{owner}")],
            )
            .unwrap();
            conn.execute(
                "UPDATE symbols SET kind = 'method', parent_symbol_id = ?1 WHERE symbol_id LIKE ?2",
                params![format!("o{owner}"), format!("o{owner}m%")],
            )
            .unwrap();
        }
        conn.execute_batch(
            "UPDATE symbols SET kind = 'class' WHERE symbol_id = 'docowner';
             UPDATE symbols SET parent_symbol_id = 'docowner' WHERE symbol_id = 'docchild';
             UPDATE symbols SET parent_symbol_id = 'o0' WHERE symbol_id IN ('testchild', 'linkchild');
             UPDATE symbols SET kind = 'import' WHERE symbol_id = 'linkchild';
             UPDATE symbols SET parent_symbol_id = 'o0m00' WHERE symbol_id = 'nestedchild';",
        ).unwrap();
        let mut candidates =
            collect_search_candidates(&conn, "planner body", None, None, false, 200).unwrap();
        // Isolate owner-driven admission so existing word/name hits cannot hide a cap failure.
        candidates.retain(|c| c.result.symbol.parent_symbol_id.is_none());
        let ranked = rerank_with(candidates.clone(), "planner body", false, None);
        assert!(
            expand_ranked_owner_members(
                &conn,
                &mut candidates,
                &ranked,
                "planner body",
                None,
                None,
                false
            )
            .unwrap()
        );
        let expanded: Vec<_> = candidates
            .iter()
            .filter(|c| c.owner_context.is_some())
            .collect();
        assert_eq!(expanded.len(), 120);
        for owner in 0..3 {
            assert_eq!(
                expanded
                    .iter()
                    .filter(|c| c.result.symbol.parent_symbol_id.as_deref()
                        == Some(&format!("o{owner}")))
                    .count(),
                40
            );
            assert!(
                !expanded
                    .iter()
                    .any(|c| c.result.symbol.symbol_id == format!("o{owner}m40"))
            );
        }
        for excluded in ["docchild", "testchild", "linkchild", "nestedchild"] {
            assert!(
                !expanded
                    .iter()
                    .any(|c| c.result.symbol.symbol_id == excluded)
            );
        }
        // Twenty ordinary non-type rows put a valid type outside the expansion window.
        let mut late_ranked = rerank_with(
            (0..20).map(|i| function(&format!("body{i}"))).collect(),
            "body",
            false,
            None,
        );
        late_ranked.push(ranked[0].clone());
        assert!(
            !expand_ranked_owner_members(
                &conn,
                &mut Vec::new(),
                &late_ranked,
                "planner body",
                None,
                None,
                false
            )
            .unwrap()
        );
    }

    #[test]
    fn lookup_and_search_skip_documentation_links_recorded_as_imports() {
        let conn = search_fixture(
            &[
                "('i1', 'f_i1', 'src/app.py', 'python', 'Flask', 'import', 'from flask import Flask', NULL, NULL, NULL, 1, 0, 1, 20, 0, 20, NULL, NULL, NULL, NULL, NULL, NULL, 'h_i1', NULL, 0, 0, 'code')".to_string(),
                "('i2', 'f_i2', 'README.md', 'markdown', 'Flask', 'import', '[Flask](https://flask.dev)', NULL, NULL, NULL, 1, 0, 1, 20, 0, 20, NULL, NULL, NULL, NULL, NULL, NULL, 'h_i2', NULL, 0, 0, 'documentation')".to_string(),
            ]
            .join(","),
        );

        let lookup: Vec<String> = search_symbols(&conn, "Flask", Some("import"), true, 10)
            .unwrap()
            .into_iter()
            .map(|s| s.path)
            .collect();
        assert_eq!(lookup, vec!["src/app.py"]);
        let search: Vec<String> =
            fts_search_symbols_scoped(&conn, "Flask", Some("import"), None, true, 10)
                .unwrap()
                .into_iter()
                .map(|r| r.symbol.path)
                .collect();
        assert_eq!(search, vec!["src/app.py"]);
    }

    #[test]
    fn lookup_ranks_a_prefix_match_above_a_substring_match() {
        let conn = search_fixture(
            &[
                code_row("c1", "tests/a.rs", "rust", "test_test_client", ""),
                code_row("c2", "tests/b.rs", "rust", "test_client_open_environ", ""),
            ]
            .join(","),
        );

        let names: Vec<String> = search_symbols(&conn, "test_cli", None, true, 10)
            .unwrap()
            .into_iter()
            .map(|s| s.name)
            .collect();

        assert_eq!(names, vec!["test_client_open_environ", "test_test_client"]);
    }

    #[test]
    fn lookup_ranks_a_declared_property_above_a_self_assignment() {
        let property = |id: &str, path: &str, signature: &str| {
            format!(
                "('{id}', 'f_{id}', '{path}', 'python', 'debug', 'property', '{signature}', NULL, NULL, NULL, 1, 0, 1, 20, 0, 20, NULL, NULL, NULL, NULL, NULL, NULL, 'h_{id}', NULL, 0, 0, 'code')"
            )
        };
        let conn = search_fixture(
            &[
                property("p1", "src/flask/app.py", "self.debug = get_debug_flag()"),
                property(
                    "p2",
                    "src/flask/sansio/app.py",
                    "@property def debug(self) -> bool",
                ),
            ]
            .join(","),
        );

        let paths: Vec<String> = search_symbols(&conn, "debug", None, false, 10)
            .unwrap()
            .into_iter()
            .map(|s| s.path)
            .collect();

        assert_eq!(paths, vec!["src/flask/sansio/app.py", "src/flask/app.py"]);
    }

    fn search_names(conn: &Connection, query: &str) -> Vec<String> {
        fts_search_symbols_scoped(conn, query, None, None, false, 10)
            .unwrap()
            .into_iter()
            .map(|r| r.symbol.name)
            .collect()
    }

    const TEST_PATH_CASES: &[(&str, bool)] = &[
        ("tests/foo.py", true),
        ("tests/x.py", true),
        ("tests/tools/test_web.py", true),
        ("src/tests/x.rs", true),
        ("__tests__/a.ts", true),
        ("a/__tests__/b.ts", true),
        ("test/x.java", true),
        ("src/test/Helper.java", true),
        ("src/test_utils.py", true),
        ("lib/test_helper.rb", true),
        ("test_config.py", true),
        ("pkg/test_data/x.json", false),
        ("src/test_detection.rs", false),
        ("autotests/tst_pagerow.qml", true),
        ("autotests/helper.qml", true),
        ("src/autotests/columnview.cpp", true),
        ("tst_foo.qml", true),
        ("src/tst_columnview.qml", true),
        ("autotests\\tst_bar.qml", true),
        ("autotests_helper/x.rs", false),
        ("src/autotest.rs", false),
        ("src/tstamp.rs", false),
        ("src/tst.rs", false),
        ("crates/julie-index/src/analysis/test_quality.rs", false),
        ("x/foo_test.go", true),
        ("x/foo.test.ts", true),
        ("x/foo.spec.js", true),
        ("src/lib_test.rs", true),
        ("src/test.rs", true),
        ("tests.rs", true),
        ("src/tests.rs", true),
        ("Foo.Tests.cs", true),
        ("x/FooTests.cs", true),
        ("src/FooTests.cs", true),
        ("src/Foo.Tests.cs", true),
        ("tests/Foo.cs", true),
        ("x/parser.spec.ts", true),
        ("test_x.py", true),
        ("test.rs", true),
        ("tests\\x.py", true),
        ("src/protocol.spec.v1/parser.rs", false),
        ("pkg/test_support/runtime.py", false),
        ("src/Contests.cs", false),
        ("spec/x.rb", false),
        ("crates/x/src/impact/likely_tests.rs", false),
        ("x/foo_tests.rs", false),
        ("src/latest.rs", false),
        ("x/latest.go", false),
        ("x/manifest.rs", false),
        ("src/attest.rs", false),
        ("contest/x.py", false),
        ("src/testing.rs", false),
        ("src/main.rs", false),
        ("pkg/service.go", false),
    ];

    #[test]
    fn test_path_rule_and_its_sql_mirror_agree_on_every_path() {
        let conn = Connection::open_in_memory().unwrap();
        let sql = format!(
            "SELECT {} FROM (SELECT :path AS path) s",
            test_path_predicate("s")
        );
        let mut stmt = conn.prepare(&sql).unwrap();
        for (path, expected) in TEST_PATH_CASES {
            assert_eq!(is_test_path(path), *expected, "rust rule: {path}");
            let from_sql: bool = stmt
                .query_row(rusqlite::named_params! { ":path": path }, |row| row.get(0))
                .unwrap();
            assert_eq!(from_sql, *expected, "sql mirror: {path}");
        }
    }

    #[test]
    fn test_name_rule_and_its_sql_mirror_agree_on_every_name() {
        let conn = Connection::open_in_memory().unwrap();
        let sql = format!(
            "SELECT {} FROM (SELECT :name AS name) s",
            names_a_test_sql("s")
        );
        let mut stmt = conn.prepare(&sql).unwrap();
        for name in [
            "test_run",
            "TestRoutes",
            "RoutesTest",
            "RoutesTests",
            "runSpec",
            "Specs",
            "create_app",
            "attest_value",
            "inspection",
        ] {
            let from_sql: bool = stmt
                .query_row(rusqlite::named_params! { ":name": name }, |row| row.get(0))
                .unwrap();
            assert_eq!(from_sql, names_a_test(name), "sql mirror: {name}");
        }
    }

    #[test]
    fn unflagged_test_file_rows_are_hidden_unless_tests_are_included() {
        let conn = search_fixture(
            &[
                code_row(
                    "a",
                    "src/parser.rs",
                    "rust",
                    "parse_sidecar",
                    "Parse a sidecar.",
                ),
                code_row(
                    "b",
                    "src/tests/helpers.py",
                    "python",
                    "parse_sidecar_fixture",
                    "Parse a sidecar.",
                ),
            ]
            .join(", "),
        );

        let default_search: Vec<String> =
            fts_search_symbols_scoped(&conn, "parse sidecar", None, None, false, 10)
                .unwrap()
                .into_iter()
                .map(|r| r.symbol.name)
                .collect();
        assert_eq!(default_search, vec!["parse_sidecar".to_string()]);

        let with_tests: Vec<String> =
            fts_search_symbols_scoped(&conn, "parse sidecar", None, None, true, 10)
                .unwrap()
                .into_iter()
                .map(|r| r.symbol.name)
                .collect();
        assert!(with_tests.contains(&"parse_sidecar_fixture".to_string()));

        let default_lookup: Vec<String> =
            search_symbols_scoped(&conn, "parse_sidecar", None, None, false, 10)
                .unwrap()
                .into_iter()
                .map(|s| s.name)
                .collect();
        assert_eq!(default_lookup, vec!["parse_sidecar".to_string()]);

        let lookup_with_tests: Vec<String> =
            search_symbols_scoped(&conn, "parse_sidecar", None, None, true, 10)
                .unwrap()
                .into_iter()
                .map(|s| s.name)
                .collect();
        assert!(lookup_with_tests.contains(&"parse_sidecar_fixture".to_string()));
    }

    #[test]
    fn count_file_symbols_prefers_the_exact_case_path_like_the_loader() {
        let conn = search_fixture(
            &[
                code_row("a", "src/Foo.rs", "rust", "one", ""),
                code_row("b", "src/foo.rs", "rust", "two", ""),
                code_row("c", "src/foo.rs", "rust", "three", ""),
            ]
            .join(", "),
        );
        for path in ["src/Foo.rs", "src/foo.rs", "src/FOO.rs", "src\\foo.rs"] {
            assert_eq!(
                count_file_symbols(&conn, path),
                load_file_symbols(&conn, path).unwrap().len(),
                "{path}"
            );
        }
        assert_eq!(count_file_symbols(&conn, "src/Foo.rs"), 1);
        assert_eq!(count_file_symbols(&conn, "src/FOO.rs"), 3);
    }

    #[test]
    fn lookup_statement_is_not_planned_as_a_multi_index_or() {
        let conn = search_fixture(&code_row("a", "src/lib.rs", "rust", "needle", ""));
        conn.execute_batch(
            "CREATE INDEX idx_symbols_name_kind ON symbols(name, kind);
             CREATE INDEX idx_symbols_test_container ON symbols(test_container);
             CREATE INDEX idx_symbols_is_test ON symbols(is_test);",
        )
        .unwrap();
        let sql = format!(
            "EXPLAIN QUERY PLAN {}",
            search_symbols_sql(false, false, 20)
        );
        let plan: Vec<String> = conn
            .prepare(&sql)
            .unwrap()
            .query_map(
                rusqlite::named_params! {
                    ":query": "needle",
                    ":pattern": "%needle%",
                    ":kind": None::<&str>,
                    ":path": None::<&str>,
                    ":path_like": None::<&str>,
                },
                |row| row.get::<_, String>(3),
            )
            .unwrap()
            .collect::<Result<_, _>>()
            .unwrap();
        assert!(
            !plan.iter().any(|step| step.contains("MULTI-INDEX OR")),
            "{plan:?}"
        );
    }

    #[test]
    fn qualified_lookup_in_a_test_file_returns_the_named_row() {
        let conn = search_fixture(
            &[
                "('c', 'f_c', 'src/tests/helpers.py', 'python', 'Helpers', 'class', 'class Helpers', '', 'pub', NULL,
                  1, 0, 9, 1, 0, 90, 1, 0, 9, 1, 5, 88, 'h_c', NULL, 0, 0, 'code')".to_string(),
                "('d', 'f_d', 'src/tests/helpers.py', 'python', 'load_fixture', 'method', 'def load_fixture()', '', 'pub', 'c',
                  10, 0, 20, 1, 100, 250, 12, 4, 19, 1, 120, 240, 'h_d', NULL, 0, 0, 'code')".to_string(),
            ]
            .join(", "),
        );

        assert_eq!(
            search_symbols_scoped(&conn, "Helpers.load_fixture", None, None, false, 10)
                .unwrap()
                .len(),
            1
        );
        assert_eq!(
            search_symbols_scoped(&conn, "Helpers.load_fixture", None, None, true, 10)
                .unwrap()
                .len(),
            1
        );
        assert!(
            search_symbols_scoped(&conn, "load_fix", None, None, false, 10)
                .unwrap()
                .is_empty()
        );
    }

    #[test]
    fn concept_query_prefers_partial_code_match_over_full_doc_match() {
        let conn = search_fixture(
            &[
                doc_row(
                    "d1",
                    "Safety guarantees",
                    "Pre-flight syntax validation runs before the edit touches disk",
                ),
                doc_row(
                    "d2",
                    "Audit",
                    "The syntax validation before an edit is the invariant",
                ),
                code_row(
                    "c1",
                    "src/syntax.rs",
                    "rust",
                    "validate_syntax",
                    "Validate the syntax of a file",
                ),
                code_row(
                    "c2",
                    "src/edit.rs",
                    "rust",
                    "replace_symbol_body",
                    "Atomic edit with validation",
                ),
            ]
            .join(","),
        );

        let names = search_names(&conn, "syntax validation before edit");

        assert_eq!(names[0], "validate_syntax");
        assert!(names.contains(&"replace_symbol_body".to_string()));
        assert!(names.contains(&"Safety guarantees".to_string()));
    }

    #[test]
    fn camel_case_query_finds_snake_case_symbol_and_vice_versa() {
        let conn = search_fixture(
            &[
                code_row("c1", "src/syntax.rs", "rust", "validate_syntax", ""),
                code_row("c2", "src/syntax.ts", "typescript", "validateSyntax", ""),
            ]
            .join(","),
        );

        let mut camel = search_names(&conn, "ValidateSyntax");
        camel.sort();
        assert_eq!(camel, vec!["validateSyntax", "validate_syntax"]);
        let mut words = search_names(&conn, "validate syntax");
        words.sort();
        assert_eq!(words, vec!["validateSyntax", "validate_syntax"]);
    }

    #[test]
    fn stop_word_prefixed_camel_case_symbol_is_still_found() {
        let conn = search_fixture(
            &[
                code_row("c1", "src/state.ts", "typescript", "isReady", ""),
                code_row("c2", "src/hooks.rs", "rust", "before", ""),
                code_row(
                    "c3",
                    "src/x.rs",
                    "rust",
                    "fooBar",
                    "has fooBar but not the other word",
                ),
            ]
            .join(","),
        );

        assert_eq!(search_names(&conn, "isReady"), vec!["isReady"]);
        assert_eq!(search_names(&conn, "before"), vec!["before"]);
    }

    #[test]
    fn related_tests_use_the_name_as_typed_without_splitting() {
        let conn = search_fixture(
            &[
                code_row("c1", "src/state.ts", "typescript", "isReady", ""),
                "('t1', 'f_t1', 'tests/ready.rs', 'rust', 'test_ready', 'function', 'fn test_ready()', NULL, NULL, NULL, 1, 0, 5, 1, 0, 50, NULL, NULL, NULL, NULL, NULL, NULL, 'h_t1', NULL, 1, 0, 'code')".to_string(),
                "('t2', 'f_t2', 'tests/state.rs', 'rust', 'isReady_reports_true', 'function', 'fn isReady_reports_true()', NULL, NULL, NULL, 1, 0, 5, 1, 0, 50, NULL, NULL, NULL, NULL, NULL, NULL, 'h_t2', NULL, 1, 0, 'code')".to_string(),
            ]
            .join(","),
        );
        let target = get_symbol_by_name(&conn, "isReady", None).unwrap().unwrap();

        let names: Vec<String> = find_related_tests(&conn, &target, 5)
            .unwrap()
            .into_iter()
            .map(|t| t.name)
            .collect();

        assert_eq!(names, vec!["isReady_reports_true"]);
    }

    #[test]
    fn related_tests_match_underscores_literally_and_skip_dunder_names() {
        let conn = search_fixture(
            &[
                code_row("c1", "src/app.py", "python", "wsgi_app", ""),
                code_row("c2", "src/app.py", "python", "__init__", ""),
                "('t1', 'f_t1', 'tests/test_app.py', 'python', 'test_wsgi_app', 'function', 'def test_wsgi_app()', NULL, NULL, NULL, 1, 0, 5, 1, 0, 50, NULL, NULL, NULL, NULL, NULL, NULL, 'h_t1', NULL, 1, 0, 'code')".to_string(),
                "('t2', 'f_t2', 'tests/test_app.py', 'python', 'test_wsgiXapp', 'function', 'def test_wsgiXapp()', NULL, NULL, NULL, 6, 0, 9, 1, 0, 50, NULL, NULL, NULL, NULL, NULL, NULL, 'h_t2', NULL, 1, 0, 'code')".to_string(),
                "('t3', 'f_t3', 'tests/test_views.py', 'python', 'test_init_once', 'function', 'def test_init_once()', NULL, NULL, NULL, 1, 0, 5, 1, 0, 50, NULL, NULL, NULL, NULL, NULL, NULL, 'h_t3', NULL, 1, 0, 'code')".to_string(),
            ]
            .join(","),
        );
        let names = |name: &str| -> Vec<String> {
            let target = get_symbol_by_name(&conn, name, None).unwrap().unwrap();
            find_related_tests(&conn, &target, 5)
                .unwrap()
                .into_iter()
                .map(|t| t.name)
                .collect()
        };

        assert_eq!(names("wsgi_app"), vec!["test_wsgi_app"]);
        assert!(names("__init__").is_empty());
    }

    #[test]
    fn related_tests_admit_unflagged_test_names_in_test_paths_without_helpers_or_low_signal_rows() {
        let conn = search_fixture(
            "('target', 'f_target', 'src/core.rs', 'rust', 'calculate', 'function', 'fn calculate()', NULL, NULL, NULL, 1, 0, 5, 1, 0, 50, NULL, NULL, NULL, NULL, NULL, NULL, 'h_target', NULL, 0, 0, 'code'),
              ('direct', 'f_direct', 'tests/direct.rs', 'rust', 'test_direct', 'function', 'fn test_direct()', NULL, NULL, NULL, 1, 0, 5, 1, 0, 50, NULL, NULL, NULL, NULL, NULL, NULL, 'h_direct', NULL, 0, 0, 'code'),
              ('pending', 'f_pending', 'autotests/tst_pending.qml', 'qml', 'test_pending', 'function', 'function test_pending() {}', NULL, NULL, NULL, 1, 0, 5, 1, 0, 50, NULL, NULL, NULL, NULL, NULL, NULL, 'h_pending', NULL, 0, 0, 'code'),
              ('name', 'f_name', 'tests/name.rs', 'rust', 'test_calculate_named', 'function', 'fn test_calculate_named()', NULL, NULL, NULL, 1, 0, 5, 1, 0, 50, NULL, NULL, NULL, NULL, NULL, NULL, 'h_name', NULL, 0, 0, 'code'),
              ('fts', 'f_fts', 'tests/fts.rs', 'rust', 'test_fts', 'function', 'fn test_fts()', 'calculate behavior', NULL, NULL, 1, 0, 5, 1, 0, 50, NULL, NULL, NULL, NULL, NULL, NULL, 'h_fts', NULL, 0, 0, 'code'),
              ('local', 'f_local', 'tests/name.rs', 'rust', 'calculate_local', 'variable', 'let calculate_local = 1;', NULL, NULL, NULL, 1, 0, 5, 1, 0, 50, NULL, NULL, NULL, NULL, NULL, NULL, 'h_local', NULL, 0, 0, 'code'),
              ('import', 'f_import', 'tests/name.rs', 'rust', 'calculate_import', 'import', 'use calculate_import;', NULL, NULL, NULL, 1, 0, 5, 1, 0, 50, NULL, NULL, NULL, NULL, NULL, NULL, 'h_import', NULL, 0, 0, 'code'),
              ('production', 'f_production', 'src/testing.rs', 'rust', 'calculate_production', 'function', 'fn calculate_production()', NULL, NULL, NULL, 1, 0, 5, 1, 0, 50, NULL, NULL, NULL, NULL, NULL, NULL, 'h_production', NULL, 0, 0, 'code'),
              ('helper', 'f_helper', 'tests/apps/factory.rs', 'rust', 'build_calculate_app', 'function', 'fn build_calculate_app()', NULL, NULL, NULL, 1, 0, 5, 1, 0, 50, NULL, NULL, NULL, NULL, NULL, NULL, 'h_helper', NULL, 0, 0, 'code')",
        );
        conn.execute_batch(
            "CREATE TABLE relationships (from_symbol_id TEXT, to_symbol_id TEXT, kind TEXT, path TEXT, start_line INTEGER, start_column INTEGER);
             CREATE TABLE pending_relationships (from_symbol_id TEXT, target_terminal_name TEXT, kind TEXT, path TEXT, start_line INTEGER, start_column INTEGER, target_receiver TEXT, target_namespace_json TEXT, target_display_name TEXT);
             CREATE TABLE type_facts (type_fact_id TEXT, symbol_id TEXT, language TEXT, resolved_type TEXT, generic_params_json TEXT);
             INSERT INTO relationships VALUES ('direct', 'target', 'calls', 'tests/direct.rs', 1, 0);
             INSERT INTO pending_relationships VALUES ('pending', 'calculate', 'calls', 'autotests/tst_pending.qml', 1, 0, NULL, '[]', 'calculate');",
        )
        .unwrap();
        let target = get_symbol_by_name(&conn, "calculate", None)
            .unwrap()
            .unwrap();

        let names: Vec<String> = find_related_tests(&conn, &target, 5)
            .unwrap()
            .into_iter()
            .map(|test| test.name)
            .collect();

        assert_eq!(
            names,
            vec![
                "test_direct",
                "test_pending",
                "test_calculate_named",
                "test_fts"
            ]
        );
    }

    #[test]
    fn exact_name_ranks_before_longer_names_with_the_same_tokens() {
        let conn = search_fixture(
            &[
                code_row(
                    "c1",
                    "src/queries.rs",
                    "rust",
                    "fts_search_symbols_scoped",
                    "search symbols scoped with fts",
                ),
                code_row("c2", "src/queries.rs", "rust", "search_symbols_scoped", ""),
            ]
            .join(","),
        );

        assert_eq!(
            search_names(&conn, "search_symbols_scoped")[0],
            "search_symbols_scoped"
        );
    }

    fn sidecar_fixture() -> Connection {
        search_fixture(
            &[
                code_row("c1", "src/sidecar.rs", "rust", "parseSha256Sidecar", ""),
                code_row(
                    "c2",
                    "src/sidecar.rs",
                    "rust",
                    "parse_sidecar_file",
                    "parse the sha256 sidecar file",
                ),
            ]
            .join(","),
        )
    }

    fn candidate<'a>(candidates: &'a [Candidate], name: &str) -> &'a Candidate {
        candidates
            .iter()
            .find(|c| c.result.symbol.name == name)
            .unwrap_or_else(|| panic!("{name} is not a candidate"))
    }

    #[test]
    fn name_substring_admits_a_symbol_the_word_branch_cannot_reach() {
        let conn = sidecar_fixture();

        let candidates = collect_search_candidates(&conn, "sha256", None, None, false, 10).unwrap();

        let target = candidate(&candidates, "parseSha256Sidecar");
        assert!(target.name_match);
        assert!(!target.word_match);
        assert!(!target.exact_name);
        assert!(target.name_terms.contains(&"sha256".to_string()));
    }

    #[test]
    fn a_row_matching_every_word_does_not_hide_a_row_matching_some() {
        let conn = search_fixture(
            &[
                code_row(
                    "c1",
                    "examples/demo.rs",
                    "rust",
                    "demo",
                    "restore offline state",
                ),
                code_row(
                    "c2",
                    "src/replay.rs",
                    "rust",
                    "replay",
                    "restore offline records",
                ),
            ]
            .join(","),
        );

        let candidates =
            collect_search_candidates(&conn, "restore offline state", None, None, false, 10)
                .unwrap();

        assert!(candidate(&candidates, "demo").word_match);
        assert!(candidate(&candidates, "replay").word_match);
        assert_eq!(search_names(&conn, "restore offline state")[0], "replay");
    }

    #[test]
    fn name_branch_admits_the_target_when_word_matches_exceed_the_cap() {
        let mut rows: Vec<String> = (1..=170)
            .map(|i| {
                code_row(
                    &format!("h{i:03}"),
                    "src/sidecar.rs",
                    "rust",
                    &format!("sidecar_helper_{i:03}"),
                    "parse sidecar file",
                )
            })
            .collect();
        rows.push(code_row(
            "c1",
            "src/sidecar.rs",
            "rust",
            "parseSha256Sidecar",
            "",
        ));
        let conn = search_fixture(&rows.join(","));

        let candidates = collect_search_candidates(
            &conn,
            "parse the sha256 sidecar file",
            None,
            None,
            false,
            40,
        )
        .unwrap();

        assert!(candidate(&candidates, "parseSha256Sidecar").name_match);
        assert_eq!(candidates.iter().filter(|c| c.word_match).count(), 160);
    }

    #[test]
    fn the_or_pass_fills_the_word_cap_but_never_exceeds_it() {
        let mut rows: Vec<String> = (1..=20)
            .map(|i| {
                code_row(
                    &format!("a{i:02}"),
                    "src/a.rs",
                    "rust",
                    &format!("both_{i:02}"),
                    "restore offline",
                )
            })
            .collect();
        rows.extend((1..=50).map(|i| {
            code_row(
                &format!("p{i:02}"),
                "src/p.rs",
                "rust",
                &format!("partial_{i:02}"),
                "restore records",
            )
        }));
        let conn = search_fixture(&rows.join(","));

        let candidates =
            collect_search_candidates(&conn, "restore offline", None, None, false, 10).unwrap();

        let word_rows: Vec<&Candidate> = candidates.iter().filter(|c| c.word_match).collect();
        assert_eq!(word_rows.len(), 40);
        assert_eq!(
            word_rows
                .iter()
                .filter(|c| c.result.symbol.name.starts_with("both_"))
                .count(),
            20
        );
    }

    #[test]
    fn exact_name_is_admitted_regardless_of_case() {
        let conn = search_fixture(&code_row("c1", "src/q.rs", "rust", "xyzzy_q", ""));

        let candidates =
            collect_search_candidates(&conn, "XYZZY_Q", None, None, false, 10).unwrap();
        assert!(candidate(&candidates, "xyzzy_q").exact_name);

        conn.execute_batch("DROP TABLE symbol_names_tri").unwrap();
        let candidates =
            collect_search_candidates(&conn, "xyzzy_q", None, None, false, 10).unwrap();
        assert!(candidate(&candidates, "xyzzy_q").exact_name);
    }

    #[test]
    fn exact_name_with_a_quote_is_admitted_through_the_trigram_index() {
        let conn = search_fixture(&code_row(
            "c1",
            "src/say.js",
            "javascript",
            "say \"hi\"",
            "",
        ));

        let candidates =
            collect_search_candidates(&conn, "say \"hi\"", None, None, false, 10).unwrap();

        let target = candidate(&candidates, "say \"hi\"");
        assert!(target.exact_name && target.name_match);
    }

    #[test]
    fn a_row_matched_by_every_branch_is_one_candidate_with_all_flags() {
        let conn = search_fixture(
            &[
                code_row("c1", "src/a.rs", "rust", "sidecar", ""),
                code_row("c2", "src/b.rs", "rust", "sidecar_helper", ""),
            ]
            .join(","),
        );

        let candidates =
            collect_search_candidates(&conn, "sidecar", None, None, false, 10).unwrap();

        assert_eq!(candidates.len(), 2);
        let target = candidate(&candidates, "sidecar");
        assert!(target.exact_name && target.word_match && target.name_match);
        assert!(target.bm25.is_some());
        let helper = candidate(&candidates, "sidecar_helper");
        assert!(!helper.exact_name && helper.word_match && helper.name_match);
    }

    #[test]
    fn an_index_without_the_trigram_table_returns_word_rows_only() {
        let conn = sidecar_fixture();
        conn.execute_batch("DROP TABLE symbol_names_tri").unwrap();

        let candidates = collect_search_candidates(&conn, "sha256", None, None, false, 10).unwrap();

        let names: Vec<&str> = candidates
            .iter()
            .map(|c| c.result.symbol.name.as_str())
            .collect();
        assert_eq!(names, vec!["parse_sidecar_file"]);
        assert!(candidates.iter().all(|c| c.word_match && !c.name_match));
        assert_eq!(search_names(&conn, "sha256"), vec!["parse_sidecar_file"]);
    }

    #[test]
    fn words_under_three_characters_skip_the_name_branch() {
        let conn = search_fixture(
            &[
                code_row("c1", "src/a.rs", "rust", "ab", ""),
                code_row("c2", "src/b.rs", "rust", "cab", ""),
            ]
            .join(","),
        );

        let candidates = collect_search_candidates(&conn, "ab", None, None, false, 10).unwrap();

        assert!(candidates.iter().all(|c| !c.name_match));
        assert!(candidate(&candidates, "ab").exact_name);
    }

    #[test]
    fn trigram_terms_include_the_identifier_parts_of_each_word() {
        assert_eq!(
            trigram_name_terms("collapse_name"),
            vec!["collapse_name", "collapse", "name"]
        );
        assert_eq!(
            trigram_name_terms("parse the sha256 sidecar"),
            vec!["parse", "sha256", "sha", "256", "sidecar"]
        );
        assert_eq!(trigram_name_terms("isReady"), vec!["isready", "ready"]);
        assert_eq!(trigram_name_terms("the before"), vec!["the", "before"]);
        assert!(trigram_name_terms("ab").is_empty());
    }

    #[test]
    fn snake_case_query_admits_a_pascal_case_name_through_the_name_branch() {
        let conn = search_fixture(
            &[
                code_row("c1", "src/collapse.rs", "rust", "CollapseName", ""),
                code_row("c2", "src/other.rs", "rust", "name_collapsed", ""),
            ]
            .join(","),
        );

        let candidates =
            collect_search_candidates(&conn, "collapse_name", None, None, false, 10).unwrap();

        let target = candidate(&candidates, "CollapseName");
        assert!(target.name_match);
        assert_eq!(target.name_terms, vec!["collapse", "name"]);
        assert_eq!(search_names(&conn, "collapse_name")[0], "CollapseName");
    }

    fn plain_candidate(name: &str, kind: &str, path: &str) -> Candidate {
        Candidate {
            result: SymbolSearchResult {
                symbol: Symbol {
                    symbol_id: format!("{path}:{name}"),
                    file_id: "f".into(),
                    path: path.into(),
                    language: "rust".into(),
                    name: name.into(),
                    kind: kind.into(),
                    signature: None,
                    doc_comment: None,
                    visibility: None,
                    parent_symbol_id: None,
                    start_line: 1,
                    start_column: 0,
                    end_line: 1,
                    end_column: 0,
                    start_byte: 0,
                    end_byte: 0,
                    body_start_line: None,
                    body_start_column: None,
                    body_end_line: None,
                    body_end_column: None,
                    body_start_byte: None,
                    body_end_byte: None,
                    body_hash: None,
                    semantic_group: None,
                    is_test: false,
                    test_container: false,
                },
                score: 0.0,
                snippet: None,
                explain: None,
            },
            bm25: None,
            exact_name: false,
            word_match: false,
            name_match: false,
            owner_match: false,
            name_terms: Vec::new(),
            documentation: false,
            nested: false,
            owner: None,
            owner_context: None,
        }
    }

    fn function(name: &str) -> Candidate {
        plain_candidate(name, "function", "src/lib.rs")
    }

    fn ranked(candidates: Vec<Candidate>, query: &str) -> Vec<(SymbolSearchResult, SearchExplain)> {
        rerank_with(candidates, query, false, None)
    }

    fn ranked_names(candidates: Vec<Candidate>, query: &str) -> Vec<String> {
        ranked(candidates, query)
            .into_iter()
            .map(|(r, _)| r.symbol.name)
            .collect()
    }

    fn documented(name: &str, signature: Option<&str>, doc: &str) -> Candidate {
        let mut candidate = function(name);
        candidate.result.symbol.signature = signature.map(str::to_string);
        candidate.result.symbol.doc_comment = Some(doc.into());
        candidate
    }

    fn strip_ansi_case() -> Vec<Candidate> {
        vec![
            documented("strip_ansi", None, "Remove ANSI escape sequences"),
            documented(
                "_strip_code_fences",
                Some("def _strip_code_fences(text: str)"),
                "The first fenced code block's body, or the stripped text",
            ),
        ]
    }

    #[test]
    fn rerank_words_split_identifiers_and_drop_stop_words_only_beside_content_words() {
        assert_eq!(
            rerank_words("parse the sha256 sidecar file"),
            vec!["parse", "sha", "256", "sidecar", "file"]
        );
        assert_eq!(
            rerank_words("parse_sha256_sidecar"),
            vec!["parse", "sha", "256", "sidecar"]
        );
        assert_eq!(
            rerank_words("ParseHTTPResponse"),
            vec!["parse", "http", "response"]
        );
        assert_eq!(rerank_words("is_ok"), vec!["ok"]);
        assert_eq!(rerank_words("the before"), vec!["the", "before"]);
    }

    #[test]
    fn stop_words_cover_english_function_words_but_not_identifier_directions() {
        for word in ["was", "whether", "another"] {
            assert!(is_stop_word(word), "{word} must be a stop word");
        }
        for word in ["down", "into", "run"] {
            assert!(!is_stop_word(word), "{word} must stay a content word");
        }
        assert_eq!(rerank_words("what was the file"), vec!["file"]);
    }

    #[test]
    fn a_public_name_sorts_before_its_private_twin_at_an_equal_score() {
        let mut private = function("_create_skill");
        private.bm25 = Some(-9.0);
        let mut public = function("create_skill");
        public.bm25 = Some(-1.0);

        let rows = ranked(vec![private, public], "create skill");

        assert_eq!(rows[0].0.score, rows[1].0.score);
        assert_eq!(rows[0].1.name_strength, rows[1].1.name_strength);
        assert_eq!(
            rows.iter()
                .map(|(r, _)| r.symbol.name.as_str())
                .collect::<Vec<_>>(),
            vec!["create_skill", "_create_skill"]
        );
    }

    #[test]
    fn a_whole_name_constant_yields_to_a_function_that_holds_the_word_with_context() {
        let rows = ranked(
            vec![
                plain_candidate("Glob", "constant", "src/glob.rs"),
                function("matches_glob_pattern"),
            ],
            "glob",
        );

        assert_eq!(rows[0].0.symbol.name, "matches_glob_pattern");
        assert_eq!(rows[1].1.name_tier, "whole");
        assert_eq!(rows[1].1.name_bonus, W_NAME_ALL_WORDS);
    }

    #[test]
    fn name_tiers_are_whole_then_all_words_then_partial_then_none() {
        let rows = ranked(
            vec![
                function("validate_everything"),
                function("validate_syntax_now"),
                function("validate_syntax"),
                function("unrelated"),
            ],
            "validate syntax",
        );
        let tiers: Vec<(&str, &str, f64)> = rows
            .iter()
            .map(|(r, e)| (r.symbol.name.as_str(), e.name_tier.as_str(), e.name_bonus))
            .collect();

        assert_eq!(
            tiers,
            vec![
                ("validate_syntax", "whole", W_NAME_WHOLE),
                ("validate_syntax_now", "all", W_NAME_ALL_WORDS),
                ("validate_everything", "partial", 0.0),
                ("unrelated", "none", 0.0),
            ]
        );
        assert_eq!(rows[0].0.score, W_NAME_WHOLE + W_TERMS + W_KIND_DEFINITION);
        assert_eq!(
            rows[1].0.score,
            W_NAME_ALL_WORDS + W_TERMS + W_KIND_DEFINITION
        );
        assert_eq!(rows[2].0.score, W_TERMS / 2.0 + W_KIND_DEFINITION);
        assert_eq!(rows[3].0.score, W_KIND_DEFINITION);
    }

    #[test]
    fn distinct_scoring_prefers_three_terms_covered_once_over_two_terms_repeated() {
        assert_eq!(
            ranked_names(strip_ansi_case(), "strip ansi escape codes"),
            vec!["strip_ansi", "_strip_code_fences"]
        );
    }

    #[test]
    fn distinct_scoring_denies_the_all_words_bonus_to_a_substring_only_name() {
        let candidates = vec![
            function("execute_julie_extract"),
            documented("slice_bytes", None, "cut a byte range"),
        ];

        let rows = ranked(candidates, "cut");
        let bonuses: Vec<(&str, &str, f64)> = rows
            .iter()
            .map(|(r, e)| (r.symbol.name.as_str(), e.name_tier.as_str(), e.name_bonus))
            .collect();

        assert_eq!(
            bonuses,
            vec![
                ("execute_julie_extract", "partial", 0.0),
                ("slice_bytes", "none", 0.0),
            ]
        );
        assert_eq!(rows[0].0.score, rows[1].0.score);
    }

    #[test]
    fn distinct_scoring_keeps_the_whole_name_and_all_words_tiers_in_order() {
        let candidates = vec![
            function("validate_everything"),
            function("validate_syntax_now"),
            function("validate_syntax"),
            function("unrelated"),
        ];

        assert_eq!(
            ranked_names(candidates, "validate syntax"),
            vec![
                "validate_syntax",
                "validate_syntax_now",
                "validate_everything",
                "unrelated"
            ]
        );
    }

    #[test]
    fn idf_weights_rank_a_term_in_one_row_above_a_term_in_most_rows() {
        let mut rows: Vec<String> = (0..10)
            .map(|i| {
                code_row(
                    &format!("s{i}"),
                    &format!("src/f{i}.rs"),
                    "rust",
                    &format!("search_{i}"),
                    "searches the index",
                )
            })
            .collect();
        rows.push(code_row(
            "rare",
            "src/rare.rs",
            "rust",
            "sanitize_input",
            "sanitize the input",
        ));
        let conn = search_fixture(&rows.join(", "));
        let terms = vec!["sanitize".to_string(), "search".to_string()];

        let weights = idf_weights(&conn, &terms);
        assert!(weights[0] > weights[1]);
    }

    #[test]
    fn idf_weights_count_a_term_the_way_the_index_tokenizer_stems_it() {
        let conn = search_fixture(&code_row(
            "n",
            "src/news.rs",
            "rust",
            "fetch_news",
            "fetch the news feed",
        ));
        let terms = vec!["news".to_string(), "unseen".to_string()];

        let weights = idf_weights(&conn, &terms);

        assert!(weights[0] < weights[1]);
    }

    #[test]
    fn a_signature_hit_past_the_head_byte_cap_does_not_credit_its_term() {
        let crediting_field = |padding: usize| {
            let mut row = function("handler");
            row.result.symbol.signature = Some(format!(
                "fn handler({}sidecar: u8)",
                "a: u8, ".repeat(padding)
            ));
            ranked(vec![row], "sidecar")[0].1.terms[0].1.clone()
        };

        assert_eq!(crediting_field(4), "signature");
        assert_eq!(crediting_field(80), "none");
    }

    #[test]
    fn explain_terms_name_the_crediting_field_of_every_query_term() {
        let rows = ranked(strip_ansi_case(), "strip ansi escape codes");
        let terms: Vec<(&str, &str, f64)> = rows[0]
            .1
            .terms
            .iter()
            .map(|(term, field, credit)| (term.as_str(), field.as_str(), *credit))
            .collect();

        assert_eq!(rows[0].0.symbol.name, "strip_ansi");
        assert_eq!(
            terms,
            vec![
                ("strip", "name", 3.0),
                ("ansi", "name", 3.0),
                ("escape", "doc", TEXT_CREDIT),
                ("codes", "none", 0.0),
            ]
        );
    }

    #[test]
    fn credited_fields_require_whole_name_multi_field_evidence_and_query_coverage() {
        let terms = vec![
            ("wsgi".into(), "name".into(), 3.0),
            ("response".into(), "signature".into(), TEXT_CREDIT),
            ("request".into(), "doc".into(), TEXT_CREDIT),
            ("error".into(), "doc".into(), TEXT_CREDIT),
        ];
        assert_eq!(credited_field_bonus(&terms), FIELD_DIVERSITY_BONUS);
        assert_eq!(credited_field_bonus(&terms[..2]), 0.0);
        assert_eq!(
            credited_field_bonus(&[
                ("wsgi".into(), "name".into(), 2.0),
                ("response".into(), "signature".into(), TEXT_CREDIT),
                ("request".into(), "doc".into(), TEXT_CREDIT),
                ("error".into(), "doc".into(), TEXT_CREDIT),
            ]),
            0.0
        );
        let sparse_terms = [
            ("terminal".into(), "name".into(), 3.0),
            ("docker".into(), "signature".into(), TEXT_CREDIT),
            ("local".into(), "doc".into(), TEXT_CREDIT),
            ("sandbox".into(), "doc".into(), TEXT_CREDIT),
            ("seven".into(), "none".into(), 0.0),
            ("ssh".into(), "none".into(), 0.0),
            ("singularity".into(), "none".into(), 0.0),
            ("modal".into(), "none".into(), 0.0),
            ("daytona".into(), "none".into(), 0.0),
            ("vercel".into(), "none".into(), 0.0),
        ];
        assert_eq!(credited_field_bonus(&sparse_terms), 0.0);
        assert_eq!(
            credited_field_bonus(&[
                ("response".into(), "name".into(), 3.0),
                ("request".into(), "name".into(), 3.0),
            ]),
            0.0
        );
    }

    #[test]
    fn credited_field_bonus_requires_two_distinct_document_terms() {
        let terms = [
            ("strip".into(), "name".into(), 3.0),
            ("path".into(), "signature".into(), TEXT_CREDIT),
            ("windows".into(), "doc".into(), TEXT_CREDIT),
        ];
        assert_eq!(credited_field_bonus(&terms), 0.0);
    }

    #[test]
    fn name_coverage_accepts_token_runs_substrings_and_stems() {
        let strengths = |name: &str, query: &str| {
            let stemmer = Stemmer::create(Algorithm::English);
            let words = stemmed_query_words(query, &stemmer);
            name_hits(name, &words, &stemmer)
        };

        assert_eq!(strengths("parseSha256Sidecar", "sha 256"), vec![3, 3]);
        assert_eq!(strengths("parseSha256Sidecar", "sha256"), vec![3, 3]);
        assert_eq!(strengths("parseSha256Sidecar", "esha"), vec![1]);
        assert_eq!(strengths("validate_syntax", "validation"), vec![2]);
        assert_eq!(strengths("is_ok", "ok"), vec![3]);
        assert_eq!(strengths("isReady", "is"), vec![3]);
        assert_eq!(strengths("größe_berechnen", "größe"), vec![3]);
        assert_eq!(
            strengths("parseSha256Sidecar", "sidecar checksum"),
            vec![3, 0]
        );
        assert_eq!(
            strengths("parseSha256Sidecar", "checksum digest"),
            vec![0, 0]
        );
    }

    #[test]
    fn a_rarer_term_moves_the_score_more_than_a_common_one() {
        let weights = [4.0, 1.0];
        let rows = rerank_with(
            vec![function("rare_helper"), function("common_helper")],
            "rare common",
            false,
            Some(&weights),
        );

        assert_eq!(
            rows[0].1.word_weights,
            vec![("rare".to_string(), 4.0), ("common".to_string(), 1.0)]
        );
        assert_eq!(rows[0].0.symbol.name, "rare_helper");
        assert_eq!(rows[0].0.score, W_TERMS * 4.0 / 5.0 + W_KIND_DEFINITION);
        assert_eq!(rows[1].0.score, W_TERMS * 1.0 / 5.0 + W_KIND_DEFINITION);
    }

    #[test]
    fn any_name_hit_outranks_a_zero_coverage_definition_for_long_queries() {
        let rows = ranked(
            vec![
                function("render_mode"),
                plain_candidate("retry_count", "constant", "src/scan.rs"),
            ],
            "how many times a failed download is tried again retry limit",
        );

        assert_eq!(rows[0].0.symbol.name, "retry_count");
        assert_eq!(rows[0].1.name_tier, "partial");
        assert!(rows[0].0.score > rows[1].0.score);
        assert_eq!(rows[1].0.score, W_KIND_DEFINITION);
    }

    #[test]
    fn a_doc_hit_past_the_head_byte_cap_does_not_credit_its_term() {
        let mut row = function("load");
        row.result.symbol.signature = Some("fn load(config: &Config) -> Loaded".into());
        row.result.symbol.doc_comment = Some(format!("{}settings", "é".repeat(321)));
        let (result, explain) = ranked(vec![row], "config settings").remove(0);

        assert_eq!(
            explain.terms,
            vec![
                ("config".into(), "signature".into(), TEXT_CREDIT),
                ("settings".into(), "none".into(), 0.0),
            ]
        );
        assert_eq!(result.score, explain.term_score + W_KIND_DEFINITION);
    }

    #[test]
    fn text_coverage_matches_whole_tokens_by_word_or_stem_prefix() {
        let crediting_field =
            |row: Candidate, query: &str| ranked(vec![row], query).remove(0).1.terms[0].1.clone();
        let doc_field = |doc: &str, query: &str| {
            let mut row = function("row");
            row.result.symbol.doc_comment = Some(doc.into());
            crediting_field(row, query)
        };
        let sig_field = |signature: &str, query: &str| {
            let mut row = function("row");
            row.result.symbol.signature = Some(signature.into());
            crediting_field(row, query)
        };

        assert_eq!(doc_field("The system runs.", "stemming"), "none");
        assert_eq!(doc_field("The stemmer runs.", "stemming"), "doc");
        assert_eq!(doc_field("Compares stems.", "stemming"), "doc");
        assert_eq!(doc_field("An important port.", "porter"), "none");
        assert_eq!(
            sig_field("fn sha256sum(data: &[u8]) -> String", "sha256"),
            "signature"
        );
        assert_eq!(sig_field("fn is_ok()", "ok"), "signature");
        assert_eq!(sig_field("fn okay()", "ok"), "none");
        assert_eq!(
            sig_field("fn parseSha256Sidecar(text)", "sidecar"),
            "signature"
        );
    }

    #[test]
    fn text_tokens_split_like_query_words_then_identifiers() {
        fn two_pass(text: &str) -> Vec<&str> {
            query_words(text)
                .into_iter()
                .flat_map(split_identifier)
                .collect()
        }
        fn one_pass(text: &str) -> Vec<&str> {
            let mut out = Vec::new();
            text_tokens_into(text, &mut out);
            out
        }
        let ascii = "fn parseHTTPResponse2(raw: &str, _id: u8) -> Vec<&str> // sha256_sum";
        let unicode = "Berechnet die Größe: größe_berechnen(pfad) -> ÜberGroß2x";

        assert_eq!(one_pass(ascii), two_pass(ascii));
        assert_eq!(
            one_pass(ascii),
            vec![
                "fn", "parse", "HTTP", "Response", "2", "raw", "str", "id", "u", "8", "Vec", "str",
                "sha", "256", "sum",
            ]
        );
        assert_eq!(one_pass(unicode), two_pass(unicode));
        assert!(one_pass("").is_empty());
        assert!(one_pass("_ __ ...").is_empty());
    }

    #[test]
    fn a_doc_credits_a_term_by_its_stem() {
        let mut row = function("check");
        row.result.symbol.doc_comment = Some("Validates the input.".into());
        let explain = ranked(vec![row], "validation").remove(0).1;

        assert_eq!(
            explain.terms,
            vec![("validation".into(), "doc".into(), TEXT_CREDIT)]
        );
    }

    #[test]
    fn kind_prior_orders_definitions_over_members_over_imports() {
        let rows = ranked(
            vec![
                plain_candidate("Scan", "import", "src/a.rs"),
                plain_candidate("Scan", "enum_member", "src/b.rs"),
                plain_candidate("Scan", "function", "src/c.rs"),
            ],
            "scan",
        );
        let order: Vec<(&str, f64)> = rows
            .iter()
            .map(|(r, e)| (r.symbol.path.as_str(), e.kind_prior))
            .collect();

        assert_eq!(
            order,
            vec![
                ("src/c.rs", W_KIND_DEFINITION),
                ("src/b.rs", W_KIND_MEMBER),
                ("src/a.rs", W_KIND_IMPORT),
            ]
        );
    }

    #[test]
    fn a_partial_name_match_on_a_member_beats_the_kind_prior_of_a_function() {
        let names = ranked_names(
            vec![
                function("RenderMode"),
                plain_candidate("MaxRetryCount", "constant", "pkg/scan.go"),
            ],
            "retry download limit timeout",
        );

        assert_eq!(names[0], "MaxRetryCount");
    }

    #[test]
    fn path_role_demotes_role_directories_unless_the_query_names_them() {
        let rows = |query: &str| {
            ranked(
                vec![
                    plain_candidate("verifyChecksum", "function", "scripts/launcher.ts"),
                    plain_candidate("verify_checksum", "function", "src/archive.rs"),
                ],
                query,
            )
        };

        let plain = rows("verify checksum");
        assert_eq!(plain[0].0.symbol.path, "src/archive.rs");
        assert_eq!(plain[1].1.path_role, W_PATH_ROLE);

        let named = rows("launcher script verify checksum");
        assert!(named.iter().all(|(_, e)| e.path_role == 0.0));

        let only_launcher = rows("launcher verify checksum");
        assert_eq!(only_launcher[0].0.symbol.path, "src/archive.rs");
        assert_eq!(only_launcher[1].1.path_role, W_PATH_ROLE);

        let windows = ranked(
            vec![plain_candidate(
                "verifyChecksum",
                "function",
                "scripts\\launcher.ts",
            )],
            "verify checksum",
        );
        assert_eq!(windows[0].1.path_role, W_PATH_ROLE);
    }

    #[test]
    fn documentation_rows_sort_after_every_code_row() {
        let mut heading = plain_candidate("Verify checksum", "heading", "README.md");
        heading.documentation = true;
        heading.result.symbol.language = "markdown".into();
        heading.result.symbol.signature = Some("Verify checksum".into());
        heading.result.symbol.doc_comment = Some("Verify the checksum of the archive.".into());
        let rows = ranked(
            vec![
                heading,
                plain_candidate("unrelated", "variable", "src/a.rs"),
            ],
            "verify checksum",
        );

        assert_eq!(rows[0].0.symbol.name, "unrelated");
        assert_eq!(rows[1].1.documentation, W_DOCUMENTATION_ROW);
        assert_eq!(rows[1].1.name_tier, "whole");
        assert!(rows[1].0.score < 0.0);
    }

    #[test]
    fn a_helper_nested_in_a_function_ranks_below_an_equal_module_function() {
        let mut helper = plain_candidate("run_test", "function", "tests/test_basic.py");
        helper.nested = true;
        let module_function = plain_candidate("run_dotenv", "function", "tests/test_basic.py");

        let rows = rerank_with(vec![helper, module_function], "run", true, None);

        assert_eq!(rows[0].0.symbol.name, "run_dotenv");
        assert_eq!(rows[1].1.nested, W_NESTED);
    }

    #[test]
    fn test_intent_boosts_test_rows_only_when_tests_are_included_and_named() {
        let rows = |query: &str, include_tests: bool| {
            let mut test_row = plain_candidate("payment_flow", "function", "tests/payment.rs");
            test_row.result.symbol.is_test = true;
            let plain_row = plain_candidate("payment_flow", "function", "src/payment.rs");
            rerank_with(vec![plain_row, test_row], query, include_tests, None)
        };

        let boosted = rows("payment flow tests", true);
        assert_eq!(boosted[0].0.symbol.path, "tests/payment.rs");
        assert_eq!(boosted[0].1.test_intent, W_TEST_INTENT);
        assert_eq!(boosted[1].1.test_intent, 0.0);

        assert!(
            rows("payment flow tests", false)
                .iter()
                .all(|(_, e)| e.test_intent == 0.0)
        );
        assert!(
            rows("payment flow", true)
                .iter()
                .all(|(_, e)| e.test_intent == 0.0)
        );
    }

    #[test]
    fn ties_break_by_bm25_then_name_length_then_path() {
        let mut word_row = plain_candidate("payment", "function", "src/z.rs");
        word_row.word_match = true;
        word_row.bm25 = Some(-4.0);
        let mut weaker_word_row = plain_candidate("payment", "function", "src/a.rs");
        weaker_word_row.word_match = true;
        weaker_word_row.bm25 = Some(-2.0);
        let mut name_only = plain_candidate("payment", "function", "src/b.rs");
        name_only.name_match = true;
        let rows = ranked(
            vec![
                plain_candidate("payment", "function", "src/y.rs"),
                name_only,
                weaker_word_row,
                word_row,
            ],
            "payment",
        );
        let paths: Vec<&str> = rows.iter().map(|(r, _)| r.symbol.path.as_str()).collect();

        assert_eq!(paths, vec!["src/z.rs", "src/a.rs", "src/b.rs", "src/y.rs"]);

        let by_length = ranked_names(
            vec![
                function("payment_gateway_client"),
                function("payment_gateway"),
            ],
            "gateway",
        );
        assert_eq!(by_length, vec!["payment_gateway", "payment_gateway_client"]);
    }

    #[test]
    fn a_whole_token_name_outranks_a_substring_name_with_better_bm25() {
        let mut token = function("csr");
        token.bm25 = Some(-1.0);
        let mut substring = function("action_csrf_token");
        substring.bm25 = Some(-5.0);

        let rows = ranked(vec![substring, token], "csr adjacency");

        assert!(rows[0].0.score > rows[1].0.score);
        assert_eq!(rows[0].0.symbol.name, "csr");
    }

    #[test]
    fn an_acronym_token_outranks_a_name_that_only_contains_it() {
        let mut token = function("http_client");
        token.bm25 = Some(-1.0);
        let mut substring = function("shttpd_config");
        substring.bm25 = Some(-5.0);

        let rows = ranked(vec![substring, token], "http");

        assert!(rows[0].0.score > rows[1].0.score);
        assert_eq!(rows[0].0.symbol.name, "http_client");
    }

    #[test]
    fn explain_reports_the_sum_of_the_name_strengths() {
        let rows = ranked(vec![function("action_csrf_token")], "csr token");

        assert_eq!(rows[0].1.name_strength, 4);
    }

    #[test]
    fn snippets_follow_the_admitting_branch() {
        let mut word_row = function("parse_sidecar_file");
        word_row.word_match = true;
        word_row.result.snippet = Some("parse the [sha256] sidecar file".into());
        let mut name_row = function("parseSha256Sidecar");
        name_row.name_match = true;
        name_row.name_terms = vec!["sha".into(), "sha256".into(), "256".into()];
        let mut exact_row = function("sha256");
        exact_row.exact_name = true;
        let rows = ranked(vec![word_row, name_row, exact_row], "sha256");
        let snippets: Vec<(&str, &str)> = rows
            .iter()
            .map(|(r, _)| (r.symbol.name.as_str(), r.snippet.as_deref().unwrap()))
            .collect();

        assert_eq!(
            snippets,
            vec![
                ("sha256", "sha256"),
                ("parseSha256Sidecar", "parse[Sha256]Sidecar"),
                ("parse_sidecar_file", "parse the [sha256] sidecar file"),
            ]
        );
        assert_eq!(rows[1].1.branches, vec!["name"]);
        assert_eq!(rows[0].1.branches, vec!["exact"]);
    }

    #[test]
    fn explain_is_attached_only_when_requested() {
        let conn = sidecar_fixture();
        let query = "sha256";

        let silent = fts_search_symbols_scoped(&conn, query, None, None, false, 10).unwrap();
        assert!(silent.iter().all(|r| r.explain.is_none()));
        assert!(silent[0].score > 0.0);
        assert_eq!(
            serde_json::to_value(&silent[0]).unwrap().get("explain"),
            None
        );

        let explained =
            fts_search_symbols_explained(&conn, query, None, None, false, 10, true).unwrap();
        let by_name = |name: &str| {
            explained
                .iter()
                .find(|r| r.symbol.name == name)
                .and_then(|r| r.explain.as_ref())
                .unwrap()
        };
        let name_only = by_name("parseSha256Sidecar");
        assert_eq!(name_only.candidates, 2);
        assert_eq!(name_only.branches, vec!["name"]);
        assert_eq!(name_only.bm25, None);
        let word_row = by_name("parse_sidecar_file");
        assert!(word_row.bm25.unwrap() < 0.0);
        assert_eq!(word_row.candidates, 2);
        assert!(
            serde_json::to_value(&explained[0])
                .unwrap()
                .get("explain")
                .is_some()
        );
    }

    #[test]
    fn search_symbols_treats_like_wildcards_as_literals() {
        let dir = crate::safe_tempdir();
        let db_path = dir.path().join("search_symbols_treats_like_wildcards.db");
        let conn = open_read_write(&db_path).unwrap();
        conn.execute_batch(
            "CREATE TABLE symbols (
                symbol_id TEXT, file_id TEXT, path TEXT, language TEXT, name TEXT, kind TEXT,
                signature TEXT, doc_comment TEXT, visibility TEXT, parent_symbol_id TEXT,
                start_line INTEGER, start_column INTEGER, end_line INTEGER, end_column INTEGER,
                start_byte INTEGER, end_byte INTEGER, body_start_line INTEGER,
                body_start_column INTEGER, body_end_line INTEGER, body_end_column INTEGER,
                body_start_byte INTEGER, body_end_byte INTEGER, body_hash TEXT,
                semantic_group TEXT, is_test INTEGER, test_container INTEGER
            );
            INSERT INTO symbols VALUES (
                's', 'f', 'src/lib.rs', 'rust', 'ordinary', 'function', NULL, NULL, NULL, NULL,
                1, 0, 1, 0, 0, 0, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, 0, 0
            );
            INSERT INTO symbols VALUES (
                'p', 'f', 'src/lib.rs', 'rust', 'literal%name', 'function', NULL, NULL, NULL, NULL,
                1, 0, 1, 0, 0, 0, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, 0, 0
            );
            INSERT INTO symbols VALUES (
                'u', 'f', 'src/lib.rs', 'rust', 'literal_name', 'function', NULL, NULL, NULL, NULL,
                1, 0, 1, 0, 0, 0, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, 0, 0
            );
            CREATE TABLE files (
                file_id TEXT, path TEXT, language TEXT, content_hash TEXT,
                content_bytes INTEGER, line_count INTEGER, indexed_at TEXT
            );
            INSERT INTO files VALUES ('f1', 'src/literal_path/lib.rs', 'rust', 'hash', 0, 0, 'now');
            INSERT INTO files VALUES ('f2', 'src/literalXpath/lib.rs', 'rust', 'hash', 0, 0, 'now'
            );",
        )
        .unwrap();

        assert_eq!(
            search_symbols(&conn, "%", None, false, 10).unwrap()[0].name,
            "literal%name"
        );
        assert_eq!(
            search_symbols(&conn, "_", None, false, 10).unwrap()[0].name,
            "literal_name"
        );
        assert_eq!(
            load_scoped_files(&conn, Some("src/literal_path"))
                .unwrap()
                .len(),
            1
        );
    }

    #[test]
    fn find_references_for_symbol_limits_callees_by_symbol_id() {
        let dir = crate::safe_tempdir();
        let db_path = dir.path().join("find_references_for_symbol.db");
        let conn = open_read_write(&db_path).unwrap();
        conn.execute_batch(
            "CREATE TABLE symbols (
                symbol_id TEXT, file_id TEXT, path TEXT, language TEXT, name TEXT, kind TEXT,
                signature TEXT, doc_comment TEXT, visibility TEXT, parent_symbol_id TEXT,
                start_line INTEGER, start_column INTEGER, end_line INTEGER, end_column INTEGER,
                start_byte INTEGER, end_byte INTEGER, body_start_line INTEGER,
                body_start_column INTEGER, body_end_line INTEGER, body_end_column INTEGER,
                body_start_byte INTEGER, body_end_byte INTEGER, body_hash TEXT,
                semantic_group TEXT, is_test INTEGER, test_container INTEGER
            );
            CREATE TABLE relationships (
                from_symbol_id TEXT, to_symbol_id TEXT, kind TEXT, path TEXT,
                start_line INTEGER, start_column INTEGER
            );
            CREATE TABLE pending_relationships (
                from_symbol_id TEXT, target_terminal_name TEXT, kind TEXT, path TEXT,
                start_line INTEGER, start_column INTEGER
            );
            INSERT INTO symbols VALUES
                ('wanted', 'f', 'a.rs', 'rust', 'new', 'method', NULL, NULL, NULL, NULL, 1, 0, 1, 0, 0, 0, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, 0, 0),
                ('other', 'f', 'b.rs', 'rust', 'new', 'method', NULL, NULL, NULL, NULL, 1, 0, 1, 0, 0, 0, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, 0, 0),
                ('wanted-callee', 'f', 'a.rs', 'rust', 'wanted_dep', 'function', NULL, NULL, NULL, NULL, 1, 0, 1, 0, 0, 0, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, 0, 0),
                ('other-callee', 'f', 'b.rs', 'rust', 'other_dep', 'function', NULL, NULL, NULL, NULL, 1, 0, 1, 0, 0, 0, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, 0, 0);
            INSERT INTO relationships VALUES
                ('other', 'other-callee', 'calls', 'b.rs', 1, 0),
                ('wanted', 'wanted-callee', 'calls', 'a.rs', 1, 0);",
        )
        .unwrap();

        let references = find_references_for_symbol(&conn, "new", "callees", 1, "wanted").unwrap();
        assert_eq!(references.len(), 1);
        assert_eq!(references[0].to_symbol_name, "wanted_dep");
    }

    #[test]
    fn test_fts_search_symbols_and_porter_stemming() {
        let dir = crate::safe_tempdir();
        let db_path = dir.path().join("fts_search_symbols.db");
        let conn = open_read_write(&db_path).unwrap();

        conn.execute_batch(
            "CREATE TABLE symbols (
                symbol_id TEXT PRIMARY KEY,
                file_id TEXT,
                path TEXT,
                language TEXT,
                name TEXT,
                kind TEXT,
                signature TEXT,
                doc_comment TEXT,
                visibility TEXT,
                parent_symbol_id TEXT,
                start_line INTEGER,
                start_column INTEGER,
                end_line INTEGER,
                end_column INTEGER,
                start_byte INTEGER,
                end_byte INTEGER,
                body_start_line INTEGER,
                body_start_column INTEGER,
                body_end_line INTEGER,
                body_end_column INTEGER,
                body_start_byte INTEGER,
                body_end_byte INTEGER,
                body_hash TEXT,
                semantic_group TEXT,
                is_test INTEGER,
                test_container INTEGER
            );
            INSERT INTO symbols VALUES (
                's1', 'f1', 'src/payment.rs', 'rust', 'PaymentGateway', 'trait',
                'pub trait PaymentGateway', 'Core payment provider interface for transactions',
                'pub', NULL, 10, 0, 20, 1, 100, 250, 12, 4, 19, 1, 120, 240, 'hash1', 'type', 0, 0
            );
            INSERT INTO symbols VALUES (
                's2', 'f1', 'src/payment.rs', 'rust', 'StripeClient', 'struct',
                'pub struct StripeClient', 'Handles HTTP requests to stripe payment API',
                'pub', NULL, 25, 0, 35, 1, 300, 450, 27, 4, 34, 1, 320, 440, 'hash2', 'type', 0, 0
            );
            INSERT INTO symbols VALUES (
                's3', 'f2', 'src/parser.rs', 'rust', 'parse_tokens', 'function',
                'pub fn parse_tokens(stream: &TokenStream) -> Result<Vec<Token>>', 'Parses syntax tokens from stream',
                'pub', NULL, 5, 0, 15, 1, 50, 200, 7, 4, 14, 1, 70, 190, 'hash3', 'function', 0, 0
            );
            INSERT INTO symbols VALUES (
                's4', 'f3', 'tests/payment_test.rs', 'rust', 'test_payment_flow', 'function',
                'fn test_payment_flow()', 'Tests payment charge workflow',
                NULL, NULL, 5, 0, 15, 1, 50, 200, 7, 4, 14, 1, 70, 190, 'hash4', 'function', 1, 0
            );",
        )
        .unwrap();

        ensure_fts_index(&conn).unwrap();

        // 1. Porter stemming match: 'parsing' matches 'parse_tokens' and 'Parses' docstring
        let results =
            fts_search_symbols_scoped(&conn, "parsing tokens", None, None, false, 10).unwrap();
        assert_eq!(results.len(), 1);
        assert_eq!(results[0].symbol.name, "parse_tokens");
        assert!(results[0].snippet.is_some());

        // 2. Docstring conceptual search: 'transactions' matches 'PaymentGateway'
        let results =
            fts_search_symbols_scoped(&conn, "transactions", None, None, false, 10).unwrap();
        assert_eq!(results.len(), 1);
        assert_eq!(results[0].symbol.name, "PaymentGateway");

        // 3. Test filter: searching 'payment' with include_tests=false ignores 'test_payment_flow'
        let results = fts_search_symbols_scoped(&conn, "payment", None, None, false, 10).unwrap();
        assert_eq!(results.len(), 2);
        assert!(results.iter().all(|r| !r.symbol.is_test));

        // 4. Test filter: searching 'payment' with include_tests=true includes 'test_payment_flow'
        let results = fts_search_symbols_scoped(&conn, "payment", None, None, true, 10).unwrap();
        assert_eq!(results.len(), 3);

        // 5. Fallback OR matching: multi-term where only some match
        let results =
            fts_search_symbols_scoped(&conn, "stripe kafka redis", None, None, false, 10).unwrap();
        assert_eq!(results.len(), 1);
        assert_eq!(results[0].symbol.name, "StripeClient");
    }

    #[test]
    fn find_related_tests_returns_each_test_once_under_the_limit() {
        let dir = crate::safe_tempdir();
        let conn = open_read_write(&dir.path().join("related_tests_limit.db")).unwrap();
        conn.execute_batch(
            "CREATE TABLE symbols (
                symbol_id TEXT PRIMARY KEY, file_id TEXT, path TEXT, language TEXT,
                name TEXT, kind TEXT, signature TEXT, doc_comment TEXT,
                visibility TEXT, parent_symbol_id TEXT, start_line INTEGER,
                start_column INTEGER, end_line INTEGER, end_column INTEGER,
                start_byte INTEGER, end_byte INTEGER, body_start_line INTEGER,
                body_start_column INTEGER, body_end_line INTEGER, body_end_column INTEGER,
                body_start_byte INTEGER, body_end_byte INTEGER, body_hash TEXT,
                semantic_group TEXT, is_test INTEGER, test_container INTEGER
            );
            CREATE TABLE relationships (
                from_symbol_id TEXT, to_symbol_id TEXT, kind TEXT, path TEXT,
                start_line INTEGER, start_column INTEGER
            );
            CREATE TABLE pending_relationships (
                from_symbol_id TEXT, target_terminal_name TEXT, kind TEXT, path TEXT,
                start_line INTEGER, start_column INTEGER,
                target_receiver TEXT, target_namespace_json TEXT, target_display_name TEXT
            );
            CREATE TABLE type_facts (
                type_fact_id TEXT, symbol_id TEXT, language TEXT, resolved_type TEXT, generic_params_json TEXT
            );
            INSERT INTO symbols VALUES
                ('s_target', 'f1', 'src/lib.rs', 'rust', 'compute', 'function', 'pub fn compute()', NULL, 'pub', NULL, 1, 0, 5, 1, 0, 50, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, 0, 0),
                ('t_a', 'f2', 'tests/a.rs', 'rust', 'first_case', 'function', 'fn first_case()', NULL, NULL, NULL, 1, 0, 20, 1, 0, 300, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, 1, 0),
                ('t_b', 'f3', 'tests/b.rs', 'rust', 'second_case', 'function', 'fn second_case()', NULL, NULL, NULL, 1, 0, 10, 1, 0, 100, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, 1, 0);
            INSERT INTO pending_relationships (from_symbol_id, target_terminal_name, kind, path, start_line, start_column, target_receiver, target_namespace_json, target_display_name) VALUES
                ('t_a', 'compute', 'calls', 'tests/a.rs', 3, 4, NULL, NULL, 'compute'),
                ('t_a', 'compute', 'calls', 'tests/a.rs', 5, 4, NULL, NULL, 'compute'),
                ('t_a', 'compute', 'calls', 'tests/a.rs', 7, 4, NULL, NULL, 'compute'),
                ('t_a', 'compute', 'calls', 'tests/a.rs', 9, 4, NULL, NULL, 'compute'),
                ('t_a', 'compute', 'calls', 'tests/a.rs', 11, 4, NULL, NULL, 'compute'),
                ('t_b', 'compute', 'calls', 'tests/b.rs', 3, 4, NULL, NULL, 'compute');",
        )
        .unwrap();
        let target = get_symbol_by_name(&conn, "compute", None).unwrap().unwrap();

        let tests = find_related_tests(&conn, &target, 5).unwrap();

        let mut names: Vec<&str> = tests.iter().map(|t| t.name.as_str()).collect();
        names.sort();
        assert_eq!(names, vec!["first_case", "second_case"]);
    }

    #[test]
    fn documentation_rows_rank_after_code_in_search() {
        let dir = crate::safe_tempdir();
        let conn = open_read_write(&dir.path().join("doc_rank.db")).unwrap();
        conn.execute_batch(
            "CREATE TABLE symbols (
                symbol_id TEXT PRIMARY KEY, file_id TEXT, path TEXT, language TEXT, name TEXT,
                kind TEXT, signature TEXT, doc_comment TEXT, visibility TEXT, parent_symbol_id TEXT,
                start_line INTEGER, start_column INTEGER, end_line INTEGER, end_column INTEGER,
                start_byte INTEGER, end_byte INTEGER, body_start_line INTEGER,
                body_start_column INTEGER, body_end_line INTEGER, body_end_column INTEGER,
                body_start_byte INTEGER, body_end_byte INTEGER, body_hash TEXT,
                semantic_group TEXT, is_test INTEGER, test_container INTEGER, content_type TEXT
            );
            INSERT INTO symbols VALUES
                ('s_doc', 'f1', 'docs/plans/018.adoc', 'asciidoc', 'Reconcile offline edits',
                 'heading', 'Reconcile offline edits', NULL, NULL, NULL,
                 3, 0, 3, 1, 10, 40, 3, 0, 3, 1, 10, 40, 'hash_doc', NULL, 0, 0, 'documentation'),
                ('s_code', 'f2', 'src/sync.rs', 'rust', 'reconcile_offline_edits', 'function',
                 'fn reconcile_offline_edits()', 'Reconcile offline edits at startup', 'pub', NULL,
                 10, 0, 20, 1, 100, 250, 12, 4, 19, 1, 120, 240, 'hash_code', NULL, 0, 0, 'code');",
        )
        .unwrap();
        ensure_fts_index(&conn).unwrap();

        let results =
            fts_search_symbols_scoped(&conn, "reconcile offline edits", None, None, false, 10)
                .unwrap();

        assert_eq!(results.len(), 2);
        assert_eq!(results[0].symbol.name, "reconcile_offline_edits");
        assert_eq!(results[1].symbol.name, "Reconcile offline edits");
    }

    #[test]
    fn test_queries_nocase_and_path_normalization() {
        let conn = Connection::open_in_memory().unwrap();
        conn.execute_batch(
            "CREATE TABLE files (
                file_id TEXT PRIMARY KEY,
                path TEXT NOT NULL,
                language TEXT,
                content_hash TEXT,
                content_bytes INTEGER,
                line_count INTEGER,
                indexed_at INTEGER
            );
            CREATE TABLE symbols (
                symbol_id TEXT PRIMARY KEY,
                file_id TEXT,
                path TEXT NOT NULL,
                language TEXT,
                name TEXT,
                kind TEXT,
                signature TEXT,
                doc_comment TEXT,
                visibility TEXT,
                parent_symbol_id TEXT,
                start_line INTEGER,
                start_column INTEGER,
                end_line INTEGER,
                end_column INTEGER,
                start_byte INTEGER,
                end_byte INTEGER,
                body_start_line INTEGER,
                body_start_column INTEGER,
                body_end_line INTEGER,
                body_end_column INTEGER,
                body_start_byte INTEGER,
                body_end_byte INTEGER,
                body_hash TEXT,
                semantic_group TEXT,
                is_test INTEGER,
                test_container INTEGER
            );
            -- Insert with backslashes and mixed casing to verify defensive normalization and COLLATE NOCASE
            INSERT INTO files VALUES ('f1', 'src\\Payment.rs', 'rust', 'hash1', 100, 10, '2026-09-14T00:00:00Z');
            INSERT INTO symbols VALUES (
                's1', 'f1', 'src\\Payment.rs', 'rust', 'ProcessPayment', 'function',
                'pub fn ProcessPayment()', NULL, 'pub', NULL, 1, 0, 5, 0, 0, 50,
                2, 4, 4, 1, 10, 45, 'bhash', 'function', 0, 0
            );",
        )
        .unwrap();

        // 1. get_file: query with uppercase, lowercase, and forward slashes
        let file = get_file(&conn, "SRC/PAYMENT.RS")
            .unwrap()
            .expect("File should be found");
        assert_eq!(
            file.path, "src/Payment.rs",
            "Path should be normalized to forward slashes"
        );

        let file2 = get_file(&conn, "src/payment.rs")
            .unwrap()
            .expect("File should be found");
        assert_eq!(file2.path, "src/Payment.rs");

        // 2. load_file_symbols: query with uppercase and forward slashes
        let syms = load_file_symbols(&conn, "SRC/PAYMENT.RS").unwrap();
        assert_eq!(syms.len(), 1);
        assert_eq!(
            syms[0].path, "src/Payment.rs",
            "Symbol path should be normalized to forward slashes"
        );

        // 3. get_symbol_by_name with path filter
        let sym = get_symbol_by_name(&conn, "ProcessPayment", Some("SRC/PAYMENT.RS"))
            .unwrap()
            .expect("Symbol should be found with case-insensitive path filter");
        assert_eq!(sym.path, "src/Payment.rs");
    }

    #[test]
    fn test_exact_case_prioritized_over_nocase() {
        let conn = Connection::open_in_memory().unwrap();
        conn.execute_batch(
            "CREATE TABLE files (
                file_id TEXT PRIMARY KEY,
                path TEXT NOT NULL,
                language TEXT,
                content_hash TEXT,
                content_bytes INTEGER,
                line_count INTEGER,
                indexed_at TEXT
            );
            CREATE TABLE symbols (
                symbol_id TEXT PRIMARY KEY,
                file_id TEXT,
                path TEXT NOT NULL,
                language TEXT,
                name TEXT NOT NULL,
                kind TEXT NOT NULL,
                signature TEXT,
                doc_comment TEXT,
                visibility TEXT,
                parent_symbol_id TEXT,
                start_line INTEGER,
                start_column INTEGER,
                end_line INTEGER,
                end_column INTEGER,
                start_byte INTEGER,
                end_byte INTEGER,
                body_start_line INTEGER,
                body_start_column INTEGER,
                body_end_line INTEGER,
                body_end_column INTEGER,
                body_start_byte INTEGER,
                body_end_byte INTEGER,
                body_hash TEXT,
                semantic_group TEXT,
                is_test INTEGER,
                test_container INTEGER
            );
            INSERT INTO files VALUES ('f1', 'src/Payment.rs', 'rust', 'h1', 100, 10, '2026-09-14T00:00:00Z');
            INSERT INTO files VALUES ('f2', 'src/payment.rs', 'rust', 'h2', 100, 10, '2026-09-14T00:00:00Z');
            INSERT INTO symbols VALUES (
                's1', 'f1', 'src/Payment.rs', 'rust', 'pay', 'function',
                'pub fn pay()', NULL, 'pub', NULL, 1, 0, 5, 0, 0, 50,
                2, 4, 4, 1, 10, 45, 'b1', 'function', 0, 0
            );
            INSERT INTO symbols VALUES (
                's2', 'f2', 'src/payment.rs', 'rust', 'pay', 'function',
                'pub fn pay()', NULL, 'pub', NULL, 1, 0, 5, 0, 0, 50,
                2, 4, 4, 1, 10, 45, 'b2', 'function', 0, 0
            );",
        )
        .unwrap();

        // Exact match should return exact file, not conflate with sibling differing only by case
        let f_lower = get_file(&conn, "src/payment.rs").unwrap().unwrap();
        assert_eq!(f_lower.path, "src/payment.rs");
        assert_eq!(f_lower.file_id, "f2");

        let f_upper = get_file(&conn, "src/Payment.rs").unwrap().unwrap();
        assert_eq!(f_upper.path, "src/Payment.rs");
        assert_eq!(f_upper.file_id, "f1");

        let syms_lower = load_file_symbols(&conn, "src/payment.rs").unwrap();
        assert_eq!(syms_lower.len(), 1);
        assert_eq!(syms_lower[0].file_id, "f2");

        let syms_upper = load_file_symbols(&conn, "src/Payment.rs").unwrap();
        assert_eq!(syms_upper.len(), 1);
        assert_eq!(syms_upper[0].file_id, "f1");
    }

    #[test]
    fn test_conservative_pending_resolution_ignores_unmatched_namespace() {
        let dir = crate::safe_tempdir();
        let db_path = dir.path().join("conservative_resolution.db");
        let conn = open_read_write(&db_path).unwrap();

        conn.execute_batch(
            "CREATE TABLE symbols (
                symbol_id TEXT PRIMARY KEY, file_id TEXT, path TEXT, language TEXT,
                name TEXT, kind TEXT, signature TEXT, doc_comment TEXT,
                visibility TEXT, parent_symbol_id TEXT, start_line INTEGER,
                start_column INTEGER, end_line INTEGER, end_column INTEGER,
                start_byte INTEGER, end_byte INTEGER, body_start_line INTEGER,
                body_start_column INTEGER, body_end_line INTEGER, body_end_column INTEGER,
                body_start_byte INTEGER, body_end_byte INTEGER, body_hash TEXT,
                semantic_group TEXT, is_test INTEGER, test_container INTEGER
            );
            CREATE TABLE relationships (
                from_symbol_id TEXT, to_symbol_id TEXT, kind TEXT, path TEXT,
                start_line INTEGER, start_column INTEGER
            );
            CREATE TABLE pending_relationships (
                from_symbol_id TEXT, target_terminal_name TEXT, kind TEXT, path TEXT,
                start_line INTEGER, start_column INTEGER,
                target_receiver TEXT, target_namespace_json TEXT, target_display_name TEXT
            );
            CREATE TABLE type_facts (
                type_fact_id TEXT, symbol_id TEXT, language TEXT, resolved_type TEXT, generic_params_json TEXT
            );
            -- Workspace struct Workspace and method Workspace::new
            INSERT INTO symbols VALUES
                ('s_ws', 'f1', 'src/workspace.rs', 'rust', 'Workspace', 'struct', 'pub struct Workspace', NULL, 'pub', NULL, 1, 0, 10, 0, 0, 100, NULL, NULL, NULL, NULL, NULL, NULL, NULL, 'struct', 0, 0),
                ('s_ws_new', 'f1', 'src/workspace.rs', 'rust', 'new', 'method', 'pub fn new() -> Workspace', NULL, 'pub', 's_ws', 2, 4, 4, 5, 20, 50, 2, 4, 4, 5, 20, 50, 'h1', 'method', 0, 0),
                ('s_caller', 'f2', 'src/caller.rs', 'rust', 'my_func', 'function', 'pub fn my_func()', NULL, 'pub', NULL, 1, 0, 10, 0, 0, 100, NULL, NULL, NULL, NULL, NULL, NULL, NULL, 'function', 0, 0);

            -- my_func calls Vec::new() (external namespace 'Vec')
            INSERT INTO pending_relationships (from_symbol_id, target_terminal_name, kind, path, start_line, start_column, target_receiver, target_namespace_json, target_display_name) VALUES
                ('s_caller', 'new', 'calls', 'src/caller.rs', 3, 8, NULL, '[\"Vec\"]', 'Vec::new');",
        )
        .unwrap();

        // When include_external is false, calling Vec::new() should NOT resolve to Workspace::new()
        let sigs = find_callee_signatures(&conn, "my_func", "s_caller", 10, false).unwrap();
        assert!(sigs.is_empty(), "Expected 0 signatures, got: {:?}", sigs);

        let refs = find_references_for_symbol(&conn, "my_func", "callees", 10, "s_caller").unwrap();
        assert!(refs.is_empty(), "Expected 0 references, got: {:?}", refs);

        // Caller references for Workspace::new should NOT list my_func
        let callers = find_references_for_symbol(&conn, "new", "callers", 10, "s_ws_new").unwrap();
        assert!(
            callers.is_empty(),
            "Expected 0 callers for Workspace::new, got: {:?}",
            callers
        );

        // Blast radius for Workspace::new should NOT impact my_func (which only called Vec::new)
        let blast = compute_blast_radius(&conn, &["new"], &["src/workspace.rs"], 2, 20).unwrap();
        assert!(
            !blast.impacted_symbols.iter().any(|s| s.name == "my_func"),
            "my_func should not be impacted before calling Workspace::new: {:?}",
            blast.impacted_symbols
        );

        // Now add a call to Workspace::new()
        conn.execute(
            "INSERT INTO pending_relationships (from_symbol_id, target_terminal_name, kind, path, start_line, start_column, target_receiver, target_namespace_json, target_display_name) VALUES ('s_caller', 'new', 'calls', 'src/caller.rs', 5, 8, NULL, '[\"Workspace\"]', 'Workspace::new')",
            [],
        )
        .unwrap();

        let sigs2 = find_callee_signatures(&conn, "my_func", "s_caller", 10, false).unwrap();
        assert_eq!(
            sigs2.len(),
            1,
            "Expected 1 signature for Workspace::new, got: {:?}",
            sigs2
        );
        assert!(sigs2[0].contains("pub fn new() -> Workspace"));

        // Blast radius for Workspace::new should now include my_func
        let blast2 = compute_blast_radius(&conn, &["new"], &["src/workspace.rs"], 2, 20).unwrap();
        assert!(
            blast2.impacted_symbols.iter().any(|s| s.name == "my_func"),
            "my_func should be impacted after calling Workspace::new: {:?}",
            blast2.impacted_symbols
        );

        // Add a bare call to new() from an unrelated caller s_other
        conn.execute(
            "INSERT INTO symbols VALUES
                ('s_other', 'f3', 'src/other.rs', 'rust', 'other_func', 'function', 'pub fn other_func()', NULL, 'pub', NULL, 1, 0, 10, 0, 0, 100, NULL, NULL, NULL, NULL, NULL, NULL, NULL, 'function', 0, 0);",
            [],
        )
        .unwrap();
        conn.execute(
            "INSERT INTO pending_relationships (from_symbol_id, target_terminal_name, kind, path, start_line, start_column, target_receiver, target_namespace_json, target_display_name) VALUES ('s_other', 'new', 'calls', 'src/other.rs', 2, 8, NULL, NULL, 'new')",
            [],
        )
        .unwrap();

        // Bare call from unrelated function should NOT resolve to Workspace::new
        let sigs_other = find_callee_signatures(&conn, "other_func", "s_other", 10, false).unwrap();
        assert!(
            sigs_other.is_empty(),
            "Bare call to new() from outside Workspace should not resolve to Workspace::new: {:?}",
            sigs_other
        );

        // A sibling method inside Workspace calling bare new() SHOULD resolve to Workspace::new
        conn.execute(
            "INSERT INTO symbols VALUES
                ('s_ws_helper', 'f1', 'src/workspace.rs', 'rust', 'helper', 'method', 'pub fn helper()', NULL, 'pub', 's_ws', 5, 4, 7, 5, 60, 90, 5, 4, 7, 5, 60, 90, 'h2', 'method', 0, 0);",
            [],
        )
        .unwrap();
        conn.execute(
            "INSERT INTO pending_relationships (from_symbol_id, target_terminal_name, kind, path, start_line, start_column, target_receiver, target_namespace_json, target_display_name) VALUES ('s_ws_helper', 'new', 'calls', 'src/workspace.rs', 6, 8, NULL, NULL, 'new')",
            [],
        )
        .unwrap();

        let sigs_sibling =
            find_callee_signatures(&conn, "helper", "s_ws_helper", 10, false).unwrap();
        assert_eq!(
            sigs_sibling.len(),
            1,
            "Sibling method calling bare new() should resolve to Workspace::new: {:?}",
            sigs_sibling
        );

        // With include_external: true, external calls should be returned
        let ext_sigs = find_callee_signatures(&conn, "my_func", "s_caller", 10, true).unwrap();
        assert!(
            ext_sigs.iter().any(|s| s.contains("Vec")),
            "include_external: true should include external Vec::new: {:?}",
            ext_sigs
        );
    }

    #[test]
    fn test_find_structural_facts_and_literals_scoped() {
        let dir = crate::safe_tempdir();
        let db_path = dir.path().join("facts_test.db");
        let conn = open_read_write(&db_path).unwrap();
        conn.execute_batch(
            "CREATE TABLE symbols (
                symbol_id TEXT PRIMARY KEY, file_id TEXT, path TEXT, language TEXT, name TEXT, kind TEXT,
                signature TEXT, doc_comment TEXT, visibility TEXT, parent_symbol_id TEXT,
                start_line INTEGER, start_column INTEGER, end_line INTEGER, end_column INTEGER,
                start_byte INTEGER, end_byte INTEGER, body_start_line INTEGER,
                body_start_column INTEGER, body_end_line INTEGER, body_end_column INTEGER,
                body_start_byte INTEGER, body_end_byte INTEGER, body_hash TEXT,
                semantic_group TEXT, is_test INTEGER, test_container INTEGER
            );
            CREATE TABLE structural_facts (
                structural_fact_id TEXT PRIMARY KEY, file_id TEXT, path TEXT NOT NULL, language TEXT,
                pattern_id TEXT, capture_name TEXT, node_kind TEXT, containing_symbol_id TEXT,
                start_line INTEGER, end_line INTEGER, confidence REAL, metadata_json TEXT
            );
            CREATE TABLE literals (
                literal_id TEXT PRIMARY KEY, file_id TEXT, path TEXT NOT NULL, language TEXT,
                kind TEXT, literal_text TEXT, carrier TEXT, containing_symbol_id TEXT,
                start_line INTEGER, start_column INTEGER, end_line INTEGER, end_column INTEGER,
                start_byte INTEGER, end_byte INTEGER
            );
            INSERT INTO structural_facts VALUES
                ('sf_toml', 'f1', 'Cargo.toml', 'toml', 'toml.key_value.v1', 'key_value', 'table', NULL, 1, 2, 1.0, '{\"key\":\"command\",\"key_path\":\"mcp_servers.code-kb.command\"}'),
                ('sf_yaml', 'f6', '.github/workflows/ci.yml', 'yaml', 'yaml.key_value.v1', 'key_value', 'block_mapping_pair', NULL, 3, 3, 1.0, '{\"key\":\"name\",\"key_path\":\"$.on.name\"}'),
                ('sf_route', 'f2', 'src/routes/api.rs', 'rust', 'axum.route.v1', 'get_users', 'function', NULL, 10, 20, 1.0, '{\"verb\":\"GET\",\"route_template\":\"/api/v1/users/{id}\",\"normalized_route_template\":\"/api/v1/users/:id\"}'),
                ('sf_sql', 'f3', 'src/db/queries.rs', 'rust', 'sql.select_query.v1', 'select_users', 'function', NULL, 30, 40, 1.0, NULL),
                ('sf_model', 'f4', 'src/models/user.rs', 'rust', 'sql.table_definition.v1', 'User', 'struct', NULL, 50, 60, 1.0, NULL),
                ('sf_css', 'f7', 'web/site.css', 'css', 'css.media_query.v1', 'media', 'media_statement', NULL, 1, 1, 1.0, NULL),
                ('sf_custom', 'f5', 'src/custom.rs', 'rust', 'my_custom_pattern', 'custom_name', 'item', NULL, 70, 80, 1.0, NULL);
            INSERT INTO literals VALUES
                ('lit_toml', 'f1', 'Cargo.toml', 'toml', 'toml_key', '\"version\"', 'key', NULL, 3, 0, 3, 9, 20, 29),
                ('lit_route', 'f2', 'src/routes/api.rs', 'rust', 'http_route', '\"/api/v1/users\"', 'string', NULL, 12, 0, 12, 15, 100, 115),
                ('lit_sql', 'f3', 'src/db/queries.rs', 'rust', 'sql_query', '\"SELECT * FROM users\"', 'string', NULL, 32, 0, 32, 21, 200, 221),
                ('lit_model', 'f4', 'src/models/user.rs', 'rust', 'model_table', '\"users_table\"', 'string', NULL, 52, 0, 52, 13, 300, 313);",
        )
        .unwrap();

        // 1. "config" alias
        let facts_config = find_structural_facts_scoped(&conn, "config", None, 10).unwrap();
        assert_eq!(facts_config.len(), 2);
        assert_eq!(facts_config[0].pattern_id, "yaml.key_value.v1");
        assert_eq!(facts_config[0].key.as_deref(), Some("on.name"));
        assert_eq!(facts_config[1].pattern_id, "toml.key_value.v1");
        assert_eq!(
            facts_config[1].key.as_deref(),
            Some("mcp_servers.code-kb.command")
        );
        let lits_config = find_literals_scoped(&conn, "config", None, 10).unwrap();
        assert_eq!(lits_config.len(), 1);
        assert_eq!(lits_config[0].kind, "toml_key");

        // 2. "route" and "routes" aliases
        let facts_route = find_structural_facts_scoped(&conn, "route", None, 10).unwrap();
        assert_eq!(facts_route.len(), 1);
        assert_eq!(facts_route[0].pattern_id, "axum.route.v1");
        assert_eq!(
            facts_route[0].key.as_deref(),
            Some("GET /api/v1/users/{id}")
        );
        let facts_routes = find_structural_facts_scoped(&conn, "routes", None, 10).unwrap();
        assert_eq!(facts_routes.len(), 1);
        let lits_route = find_literals_scoped(&conn, "route", None, 10).unwrap();
        assert_eq!(lits_route.len(), 1);
        assert_eq!(lits_route[0].kind, "http_route");

        // 3. "query", "queries", "sql" aliases
        for q in &["query", "queries", "sql"] {
            let facts = find_structural_facts_scoped(&conn, q, None, 10).unwrap();
            assert_eq!(facts.len(), 2, "Failed for {}", q);
            assert!(facts.iter().all(|f| f.pattern_id.starts_with("sql.")));
            let lits = find_literals_scoped(&conn, q, None, 10).unwrap();
            assert_eq!(lits.len(), 1, "Failed for {}", q);
            assert_eq!(lits[0].kind, "sql_query");
        }

        // 4. "model" and "models" aliases
        for m in &["model", "models"] {
            let facts = find_structural_facts_scoped(&conn, m, None, 10).unwrap();
            assert_eq!(facts.len(), 1, "Failed for {}", m);
            assert_eq!(facts[0].pattern_id, "sql.table_definition.v1");
            let lits = find_literals_scoped(&conn, m, None, 10).unwrap();
            assert_eq!(lits.len(), 1, "Failed for {}", m);
            assert_eq!(lits[0].kind, "model_table");
        }

        // 5. Custom / unknown category
        let facts_custom = find_structural_facts_scoped(&conn, "custom_pattern", None, 10).unwrap();
        assert_eq!(facts_custom.len(), 1);
        assert_eq!(facts_custom[0].pattern_id, "my_custom_pattern");
        assert_eq!(facts_custom[0].key, None);

        // 6. Path filter: exact file match
        let facts_exact =
            find_structural_facts_scoped(&conn, "config", Some("Cargo.toml"), 10).unwrap();
        assert_eq!(facts_exact.len(), 1);
        let facts_miss =
            find_structural_facts_scoped(&conn, "config", Some("src/routes/api.rs"), 10).unwrap();
        assert_eq!(facts_miss.len(), 0);

        // 7. Path filter: directory prefix
        let facts_dir =
            find_structural_facts_scoped(&conn, "route", Some("src/routes"), 10).unwrap();
        assert_eq!(facts_dir.len(), 1);
        let facts_dir_miss =
            find_structural_facts_scoped(&conn, "route", Some("src/db"), 10).unwrap();
        assert_eq!(facts_dir_miss.len(), 0);

        // 8. Delegating find_structural_facts and find_literals
        let f_del = find_structural_facts(&conn, "config", 10).unwrap();
        assert_eq!(f_del.len(), 2);
        let l_del = find_literals(&conn, "config", 10).unwrap();
        assert_eq!(l_del.len(), 1);
    }

    fn local_variable_fixture() -> Connection {
        let conn = Connection::open_in_memory().unwrap();
        conn.execute_batch(
            "CREATE TABLE symbols (
                symbol_id TEXT PRIMARY KEY, file_id TEXT, path TEXT, language TEXT, name TEXT,
                kind TEXT, signature TEXT, doc_comment TEXT, visibility TEXT,
                parent_symbol_id TEXT, start_line INTEGER, start_column INTEGER,
                end_line INTEGER, end_column INTEGER, start_byte INTEGER, end_byte INTEGER,
                body_start_line INTEGER, body_start_column INTEGER, body_end_line INTEGER,
                body_end_column INTEGER, body_start_byte INTEGER, body_end_byte INTEGER,
                body_hash TEXT, semantic_group TEXT, is_test INTEGER, test_container INTEGER
            );
            INSERT INTO symbols (symbol_id, file_id, path, language, name, kind, signature,
                                 parent_symbol_id, start_line, start_column, end_line, end_column,
                                 start_byte, end_byte, is_test, test_container)
            VALUES
                ('func', 'f1', 'src/db.rs', 'rust', 'open_conn', 'function',
                 'fn open_conn() -> sqlite Connection', NULL, 1, 0, 9, 1, 0, 100, 0, 0),
                ('local', 'f1', 'src/db.rs', 'rust', 'conn', 'variable',
                 'let conn: sqlite Connection', 'func', 2, 4, 2, 30, 10, 40, 0, 0),
                ('pool', 'f1', 'src/db.rs', 'rust', 'Pool', 'struct',
                 'struct Pool sqlite', NULL, 12, 0, 16, 1, 120, 200, 0, 0),
                ('field', 'f1', 'src/db.rs', 'rust', 'conn', 'variable',
                 'conn: sqlite Connection', 'pool', 13, 4, 13, 28, 130, 160, 0, 0),
                ('global', 'f1', 'src/db.rs', 'rust', 'conn', 'variable',
                 'static conn: sqlite Connection', NULL, 20, 0, 20, 30, 210, 240, 0, 0),
                ('closure', 'f1', 'src/db.rs', 'rust', 'with_conn', 'variable',
                 'let with_conn = |c: sqlite Connection|', 'func', 4, 4, 6, 5, 50, 90, 0, 0),
                ('nested', 'f1', 'src/db.rs', 'rust', 'conn', 'variable',
                 'let conn = c sqlite', 'closure', 5, 8, 5, 24, 60, 80, 0, 0);",
        )
        .unwrap();
        conn
    }

    fn matched_symbol_ids(conn: &Connection, query: &str) -> Vec<String> {
        let mut stmt = conn
            .prepare(
                "SELECT s.symbol_id FROM symbols_fts f
                 JOIN symbols s ON s.rowid = f.rowid
                 WHERE f.symbols_fts MATCH ?1 ORDER BY s.symbol_id",
            )
            .unwrap();
        let mut ids = stmt
            .query_map(params![query], |row| row.get::<_, String>(0))
            .unwrap()
            .collect::<Result<Vec<_>, _>>()
            .unwrap();
        ids.sort();
        ids
    }

    #[test]
    fn fts_index_excludes_locals_and_rebuilds_a_stale_index() {
        let conn = local_variable_fixture();
        conn.execute_batch(
            "CREATE VIRTUAL TABLE symbols_fts USING fts5(
                name, signature, doc_comment,
                content='symbols', content_rowid='rowid', tokenize='porter unicode61'
            );
            INSERT INTO symbols_fts(rowid, name, signature, doc_comment)
            SELECT rowid, name, signature, doc_comment FROM symbols;",
        )
        .unwrap();

        ensure_fts_index(&conn).unwrap();

        assert_eq!(
            matched_symbol_ids(&conn, "sqlite"),
            vec!["field", "func", "global", "pool"]
        );
    }

    #[test]
    fn lookup_excludes_locals_and_parameters() {
        let conn = local_variable_fixture();

        let ids: Vec<String> = search_symbols_scoped(&conn, "conn", None, None, false, 10)
            .unwrap()
            .into_iter()
            .map(|s| s.symbol_id)
            .collect();

        assert!(!ids.contains(&"local".to_string()));
        assert!(!ids.contains(&"nested".to_string()));
        assert!(ids.contains(&"field".to_string()));
        assert!(ids.contains(&"global".to_string()));
    }

    #[test]
    fn search_excludes_locals_and_parameters() {
        let conn = local_variable_fixture();
        ensure_fts_index(&conn).unwrap();

        let ids: Vec<String> = fts_search_symbols_scoped(&conn, "sqlite", None, None, false, 10)
            .unwrap()
            .into_iter()
            .map(|r| r.symbol.symbol_id)
            .collect();

        assert!(!ids.contains(&"local".to_string()));
        assert!(ids.contains(&"func".to_string()));
    }

    #[test]
    fn variable_kind_search_keeps_full_text_matching() {
        let conn = local_variable_fixture();
        ensure_fts_index(&conn).unwrap();

        let ids: Vec<String> = fts_search_symbols_scoped(
            &conn,
            "sqlite connection",
            Some("variable"),
            None,
            false,
            10,
        )
        .unwrap()
        .into_iter()
        .map(|r| r.symbol.symbol_id)
        .collect();

        assert!(ids.contains(&"global".to_string()));
        assert!(ids.contains(&"field".to_string()));
    }

    #[test]
    fn qualified_lookup_returns_the_named_local_variable() {
        let conn = local_variable_fixture();

        let ids: Vec<String> =
            search_symbols_scoped(&conn, "open_conn::conn", None, None, false, 10)
                .unwrap()
                .into_iter()
                .map(|s| s.symbol_id)
                .collect();

        assert_eq!(ids, vec!["local".to_string()]);
    }

    #[test]
    fn exact_local_variable_outranks_a_partial_global_match_within_the_limit() {
        let conn = Connection::open_in_memory().unwrap();
        conn.execute_batch(
            "CREATE TABLE symbols (
                symbol_id TEXT PRIMARY KEY, file_id TEXT, path TEXT, language TEXT, name TEXT,
                kind TEXT, signature TEXT, doc_comment TEXT, visibility TEXT,
                parent_symbol_id TEXT, start_line INTEGER, start_column INTEGER,
                end_line INTEGER, end_column INTEGER, start_byte INTEGER, end_byte INTEGER,
                body_start_line INTEGER, body_start_column INTEGER, body_end_line INTEGER,
                body_end_column INTEGER, body_start_byte INTEGER, body_end_byte INTEGER,
                body_hash TEXT, semantic_group TEXT, is_test INTEGER, test_container INTEGER
            );
            INSERT INTO symbols (symbol_id, file_id, path, language, name, kind, signature,
                                 parent_symbol_id, start_line, start_column, end_line, end_column,
                                 start_byte, end_byte, is_test, test_container)
            VALUES
                ('func', 'f1', 'src/sum.rs', 'rust', 'digest', 'function',
                 'fn digest()', NULL, 1, 0, 9, 1, 0, 100, 0, 0),
                ('local', 'f1', 'src/sum.rs', 'rust', 'checksum', 'variable',
                 'let checksum', 'func', 2, 4, 2, 30, 10, 40, 0, 0),
                ('global', 'f1', 'src/sum.rs', 'rust', 'getChecksum', 'variable',
                 'const getChecksum', NULL, 20, 0, 20, 30, 210, 240, 0, 0);",
        )
        .unwrap();
        ensure_fts_index(&conn).unwrap();

        let rows =
            fts_search_symbols_explained(&conn, "checksum", Some("variable"), None, false, 1, true)
                .unwrap();

        assert_eq!(rows.len(), 1);
        assert_eq!(rows[0].symbol.symbol_id, "local");
        let explain = rows[0].explain.as_ref().unwrap();
        assert_eq!(explain.name_tier, "whole");
        assert_eq!(explain.branches, vec!["exact", "name"]);
        assert_eq!(explain.candidates, 2);
    }

    #[test]
    fn variable_kind_filter_returns_locals_and_parameters() {
        let conn = local_variable_fixture();
        ensure_fts_index(&conn).unwrap();

        let lookup_ids: Vec<String> =
            search_symbols_scoped(&conn, "conn", Some("variable"), None, false, 10)
                .unwrap()
                .into_iter()
                .map(|s| s.symbol_id)
                .collect();
        assert!(lookup_ids.contains(&"local".to_string()));
        assert!(lookup_ids.contains(&"nested".to_string()));

        let search_ids: Vec<String> =
            fts_search_symbols_scoped(&conn, "conn", Some("variable"), None, false, 10)
                .unwrap()
                .into_iter()
                .map(|r| r.symbol.symbol_id)
                .collect();
        assert!(search_ids.contains(&"local".to_string()));
    }
}