optave-codegraph 3.13.0

package/dist/domain/graph/builder/stages/native-orchestrator.js

@@ -0,0 +1,1763 @@
+/**
+ * Native build orchestrator stage — runs the full Rust pipeline when available,
+ * with WASM fallback for files the native engine drops.
+ *
+ * Extracted from `pipeline.ts` to break the name-collision cycle between
+ * `buildGraph()` (this module's caller) and `ctx.nativeDb.buildGraph()` (the
+ * Rust orchestrator entry point invoked here). Codegraph's name-based call
+ * resolver previously conflated the two and reported a false-positive
+ * function-level cycle (`buildGraph ↔ tryNativeOrchestrator`).
+ *
+ * The orchestrator-selection strategy lives here so `pipeline.ts` stays a thin
+ * top-level controller: detect changes, try native, fall back to JS stages.
+ */
+import { execFileSync } from 'node:child_process';
+import path from 'node:path';
+import { performance } from 'node:perf_hooks';
+import { acquireAdvisoryLock, closeDbPair, openDb, purgeFilesData, releaseAdvisoryLock, setBuildMeta, } from '../../../../db/index.js';
+import { debug, info, warn } from '../../../../infrastructure/logger.js';
+import { loadNative } from '../../../../infrastructure/native.js';
+import { semverCompare } from '../../../../infrastructure/update-check.js';
+import { normalizePath, TS_NATIVE_CONFIDENCE_FLOOR } from '../../../../shared/constants.js';
+import { toErrorMessage } from '../../../../shared/errors.js';
+import { CODEGRAPH_VERSION } from '../../../../shared/version.js';
+import { classifyNativeDrops, formatDropExtensionSummary, getInstalledWasmExtensions, NATIVE_SUPPORTED_EXTENSIONS, parseFilesWasmForBackfill, patchDataflowResult, } from '../../../parser.js';
+import { computeConfidence } from '../../resolve.js';
+import { resolveThisDispatch } from '../cha.js';
+import { batchInsertEdges, batchInsertNodes, CHA_DISPATCH_PENALTY, CHA_TYPED_DISPATCH_CONFIDENCE, collectFiles as collectFilesUtil, fileHash, fileStat, readFileSafe, } from '../helpers.js';
+import { NativeDbProxy } from '../native-db-proxy.js';
+import { closeNativeDb } from './native-db-lifecycle.js';
+// ── Native orchestrator helpers ───────────────────────────────────────
+/** Determine whether the native orchestrator should be skipped. Returns a reason string, or null if it should run. */
+function shouldSkipNativeOrchestrator(ctx) {
+    if (ctx.forceFullRebuild)
+        return 'forceFullRebuild';
+    // v3.9.0 addon had buggy incremental purge (wrong SQL on analysis tables,
+    // scoped removal over-detection). Fixed in v3.9.1 by PR #865. Gate on
+    // < 3.9.1 so v3.9.1+ uses the fast Rust orchestrator path.
+    const orchestratorBuggy = !!ctx.engineVersion && semverCompare(ctx.engineVersion, '3.9.1') < 0;
+    if (orchestratorBuggy)
+        return `buggy addon ${ctx.engineVersion}`;
+    if (ctx.engineName !== 'native')
+        return `engine=${ctx.engineName}`;
+    return null;
+}
+/** Checkpoint WAL through rusqlite, close nativeDb, and reopen better-sqlite3.
+ *  Returns false if the DB reopen fails (caller should return partial result). */
+function handoffWalAfterNativeBuild(ctx) {
+    closeNativeDb(ctx, 'post-native-build');
+    try {
+        ctx.db.close();
+    }
+    catch (e) {
+        debug(`handoffWal JS db close failed: ${toErrorMessage(e)}`);
+    }
+    try {
+        ctx.db = openDb(ctx.dbPath);
+        return true;
+    }
+    catch (reopenErr) {
+        warn(`Failed to reopen DB after native build: ${reopenErr.message}`);
+        return false;
+    }
+}
+/**
+ * Reconstruct fileSymbols from the DB after a native orchestrator build.
+ * When `scopeFiles` is provided, only loads those files (for analysis-only).
+ * When omitted, loads all files (needed for structure rebuilds).
+ */
+function reconstructFileSymbolsFromDb(ctx, scopeFiles) {
+    let query = 'SELECT file, name, kind, line, end_line as endLine FROM nodes WHERE file IS NOT NULL';
+    const params = [];
+    if (scopeFiles && scopeFiles.length > 0) {
+        const placeholders = scopeFiles.map(() => '?').join(',');
+        query += ` AND file IN (${placeholders})`;
+        params.push(...scopeFiles);
+    }
+    query += ' ORDER BY file, line';
+    const rows = ctx.db.prepare(query).all(...params);
+    const fileSymbols = new Map();
+    for (const row of rows) {
+        let entry = fileSymbols.get(row.file);
+        if (!entry) {
+            entry = {
+                definitions: [],
+                calls: [],
+                imports: [],
+                classes: [],
+                exports: [],
+                typeMap: new Map(),
+            };
+            fileSymbols.set(row.file, entry);
+        }
+        entry.definitions.push({
+            name: row.name,
+            kind: row.kind,
+            line: row.line,
+            endLine: row.endLine ?? undefined,
+        });
+    }
+    // Populate import/export counts from DB edges so buildStructure
+    // computes correct import_count/export_count in node_metrics.
+    // The extractor arrays aren't persisted to the DB, so we derive
+    // counts from edge data instead (#804).
+    const importCountRows = ctx.db
+        .prepare(`SELECT n.file, COUNT(*) AS cnt
+       FROM edges e JOIN nodes n ON e.source_id = n.id
+       WHERE e.kind IN ('imports', 'imports-type', 'dynamic-imports')
+         AND n.file IS NOT NULL
+       GROUP BY n.file`)
+        .all();
+    for (const row of importCountRows) {
+        const entry = fileSymbols.get(row.file);
+        if (entry)
+            entry.imports = new Array(row.cnt);
+    }
+    const exportCountRows = ctx.db
+        .prepare(`SELECT n_tgt.file, COUNT(DISTINCT n_tgt.id) AS cnt
+       FROM edges e
+       JOIN nodes n_tgt ON e.target_id = n_tgt.id
+       JOIN nodes n_src ON e.source_id = n_src.id
+       WHERE e.kind IN ('imports', 'imports-type', 'reexports')
+         AND n_tgt.file IS NOT NULL
+         AND n_src.file != n_tgt.file
+       GROUP BY n_tgt.file`)
+        .all();
+    for (const row of exportCountRows) {
+        const entry = fileSymbols.get(row.file);
+        if (entry)
+            entry.exports = new Array(row.cnt);
+    }
+    return fileSymbols;
+}
+/**
+ * Run JS buildStructure() after native orchestrator to fill directory nodes + contains edges.
+ * For full builds, passes changedFiles=null (full rebuild).
+ * For incremental builds, passes the changed file list to scope the update.
+ */
+async function runPostNativeStructure(ctx, allFileSymbols, isFullBuild, changedFiles) {
+    const structureStart = performance.now();
+    try {
+        const directories = new Set();
+        for (const relPath of allFileSymbols.keys()) {
+            const parts = relPath.split('/');
+            for (let i = 1; i < parts.length; i++) {
+                directories.add(parts.slice(0, i).join('/'));
+            }
+        }
+        const lineCountMap = new Map();
+        const cachedLineCounts = ctx.db
+            .prepare(`SELECT n.name AS file, m.line_count
+         FROM node_metrics m JOIN nodes n ON m.node_id = n.id
+         WHERE n.kind = 'file'`)
+            .all();
+        for (const row of cachedLineCounts) {
+            lineCountMap.set(row.file, row.line_count);
+        }
+        // Full builds need null (rebuild everything). Incremental builds pass the
+        // changed file list so buildStructure only updates those files' metrics
+        // and contains edges — matching the JS pipeline's medium-incremental path.
+        const changedFilePaths = isFullBuild || !changedFiles?.length ? null : changedFiles;
+        const { buildStructure: buildStructureFn } = (await import('../../../../features/structure.js'));
+        buildStructureFn(ctx.db, allFileSymbols, ctx.rootDir, lineCountMap, directories, changedFilePaths);
+        debug(`Structure phase completed after native orchestrator${changedFilePaths ? ` (${changedFilePaths.length} files)` : ' (full)'}`);
+    }
+    catch (err) {
+        warn(`Structure phase failed after native build: ${toErrorMessage(err)}`);
+    }
+    return performance.now() - structureStart;
+}
+/**
+ * P6: Build dataflow_vertices and inter-procedural edges after the Rust
+ * orchestrator completes.
+ *
+ * The Rust pipeline writes flows_to/returns/mutates edges directly to the DB
+ * but never writes to dataflow_vertices or dataflow_summary. This pass re-runs
+ * the Rust dataflow visitor (via extractDataflowAnalysis — fast, no re-parse)
+ * to get the DataflowResult and calls buildDataflowVerticesFromMap.
+ *
+ * Languages for which Rust has no dataflow rules return null from
+ * extractDataflowAnalysis and are silently skipped here. A follow-up issue
+ * (#1614 adjacent) will add WASM fallback for those languages.
+ */
+async function runDataflowVertexPass(ctx, changedFiles) {
+    if (ctx.opts.dataflow === false)
+        return;
+    const native = loadNative();
+    if (!native?.extractDataflowAnalysis)
+        return;
+    // Determine which files to process: changed files for incremental, all for full builds.
+    let filesToProcess;
+    if (changedFiles && changedFiles.length > 0) {
+        filesToProcess = changedFiles;
+    }
+    else {
+        // Full build: scope to files that need vertex extraction rather than scanning every
+        // file in the project. Two categories:
+        //   (a) Non-native language files — NATIVE_SUPPORTED_EXTENSIONS doesn't cover them,
+        //       so extractDataflowAnalysis returns null; the wasmStubs path calls buildDataflowEdges
+        //       which writes both edges AND vertices for those files.
+        //   (b) Native-language files with dataflow edges already written by the Rust orchestrator
+        //       (flows_to/returns/mutates) — those need vertex rows to connect them.
+        //
+        // Skipping native-language files with no dataflow edges is safe: extractDataflowAnalysis
+        // would return argFlows=[], assignments=[], mutations=[] for them, producing zero vertices
+        // and zero inter-procedural edges. Excluding them avoids O(n_total_files) re-analysis on
+        // every full build (codegraph itself: ~2000 files, ~50-80% with no dataflow edges).
+        const filesWithDataflow = new Set(ctx.db
+            .prepare(`SELECT DISTINCT n.file
+             FROM dataflow d
+             JOIN nodes n ON n.id = d.source_id
+             WHERE n.file IS NOT NULL`)
+            .all().map((r) => r.file));
+        filesToProcess = ctx.db
+            .prepare(`SELECT DISTINCT file FROM nodes WHERE file IS NOT NULL AND kind != 'directory'`)
+            .all()
+            .map((r) => r.file)
+            .filter((f) => {
+            const ext = path.extname(f).toLowerCase();
+            // Non-native files: always include (WASM handles them via wasmStubs path).
+            if (!NATIVE_SUPPORTED_EXTENSIONS.has(ext))
+                return true;
+            // Native files: only include when Rust wrote dataflow edges for them.
+            return filesWithDataflow.has(f);
+        });
+    }
+    // Split files into two buckets:
+    //   nativeDataflow — Rust extracted data (vertex-only pass; edges already in DB)
+    //   wasmStubs      — Rust returned null (WASM will handle edges + vertices)
+    const nativeDataflow = new Map();
+    const wasmStubs = new Map();
+    for (const relPath of filesToProcess) {
+        const absPath = path.join(ctx.rootDir, relPath);
+        const source = readFileSafe(absPath);
+        if (!source)
+            continue;
+        let result = null;
+        try {
+            result = native.extractDataflowAnalysis(source, absPath);
+        }
+        catch {
+            // Language-specific parse failure — fall through to WASM.
+        }
+        if (result) {
+            // Normalise the native DataflowResult: Rust emits `bindingType: string | null`
+            // (flat) while the TS dataflow layer expects `binding: { type, index? }` (object).
+            // patchNativeResult handles this via patchDataflow for the full parse path;
+            // extractDataflowAnalysis is a vertex-only fast path that bypasses patchNativeResult,
+            // so we apply the same normalisation here.
+            patchDataflowResult(result);
+            nativeDataflow.set(relPath, result);
+        }
+        else {
+            // Rust has no dataflow rules for this language; WASM fallback will handle
+            // both edge insertion and vertex extraction. Since Rust inserted 0 dataflow
+            // edges for these files, there is no risk of duplicates.
+            wasmStubs.set(relPath, { definitions: [], _langId: null, _tree: null });
+        }
+    }
+    const { buildExtToLangMap } = (await import('../../../../ast-analysis/shared.js'));
+    const { buildDataflowVerticesFromMap, buildDataflowEdges, collectCallerStitchCandidates, collectFuncIdsForFiles, } = (await import('../../../../features/dataflow.js'));
+    // Rust-supported languages: build vertices only (edges already written by Rust orchestrator).
+    if (nativeDataflow.size > 0) {
+        // P4: On incremental builds, unchanged caller files' arg_in edges were deleted when
+        // the changed files' param vertices were purged and recreated. Re-collect stitch
+        // candidates from those caller files so buildInterproceduralStitch can reconnect them.
+        // Skip on full builds (changedFiles absent/empty) — nativeDataflow covers all files.
+        let p4Candidates = [];
+        let p4Captures = [];
+        if (changedFiles && changedFiles.length > 0) {
+            const changedSet = new Set(changedFiles);
+            const totalFilesInDb = ctx.db.prepare(`SELECT COUNT(DISTINCT file) AS n FROM nodes`).get().n;
+            // Only run P4 when this is a real incremental build (not all files changed).
+            if (nativeDataflow.size < totalFilesInDb) {
+                const changedFuncIds = collectFuncIdsForFiles(ctx.db, changedSet);
+                if (changedFuncIds.length > 0) {
+                    const extra = await collectCallerStitchCandidates(ctx.db, changedFuncIds, changedSet, ctx.rootDir, buildExtToLangMap(), null, // parsers — lazily loaded inside collectCallerStitchCandidates
+                    null);
+                    p4Candidates = extra.candidates;
+                    p4Captures = extra.captures;
+                }
+            }
+        }
+        const interCount = buildDataflowVerticesFromMap(ctx.db, nativeDataflow, p4Candidates.length > 0 ? p4Candidates : undefined, p4Captures.length > 0 ? p4Captures : undefined);
+        if (interCount > 0) {
+            info(`Dataflow (native orchestrator): ${interCount} inter-procedural edges inserted${p4Candidates.length > 0 ? ` (P4: ${p4Candidates.length} re-stitch candidate(s) from unchanged callers)` : ''}`);
+        }
+    }
+    // Rust-unsupported languages: run the full WASM extraction (edges + vertices).
+    // wasmStubs entries have no `.dataflow` property, so the native bulk-insert
+    // fast path in buildDataflowEdges is always skipped for them — WASM runs
+    // both edge insertion and vertex extraction end-to-end.
+    if (wasmStubs.size > 0) {
+        await buildDataflowEdges(ctx.db, wasmStubs, ctx.rootDir);
+    }
+}
+/**
+ * JS fallback for AST/complexity/CFG/dataflow analysis after native orchestrator.
+ * Used when the Rust addon doesn't include analysis persistence (older addon
+ * version) or when analysis failed on the Rust side.
+ */
+async function runPostNativeAnalysis(ctx, allFileSymbols, changedFiles) {
+    const timing = { astMs: 0, complexityMs: 0, cfgMs: 0, dataflowMs: 0 };
+    // Scope analysis fileSymbols to changed files only
+    let analysisFileSymbols;
+    if (changedFiles && changedFiles.length > 0) {
+        analysisFileSymbols = new Map();
+        for (const f of changedFiles) {
+            const entry = allFileSymbols.get(f);
+            if (entry)
+                analysisFileSymbols.set(f, entry);
+        }
+    }
+    else {
+        analysisFileSymbols = allFileSymbols;
+    }
+    // Reopen nativeDb for analysis features (suspend/resume WAL pattern).
+    const native = loadNative();
+    if (native?.NativeDatabase) {
+        try {
+            ctx.nativeDb = native.NativeDatabase.openReadWrite(ctx.dbPath);
+            if (ctx.engineOpts)
+                ctx.engineOpts.nativeDb = ctx.nativeDb;
+        }
+        catch {
+            ctx.nativeDb = undefined;
+            if (ctx.engineOpts)
+                ctx.engineOpts.nativeDb = undefined;
+        }
+    }
+    // Flush JS WAL pages once so Rust can see them, then no-op callbacks.
+    // Previously each feature called wal_checkpoint(TRUNCATE) individually
+    // (~68ms each × 3-4 features). One FULL checkpoint suffices.
+    if (ctx.nativeDb && ctx.engineOpts) {
+        ctx.db.pragma('wal_checkpoint(FULL)');
+        ctx.engineOpts.suspendJsDb = () => { };
+        ctx.engineOpts.resumeJsDb = () => { };
+    }
+    try {
+        const { runAnalyses: runAnalysesFn } = (await import('../../../../ast-analysis/engine.js'));
+        const result = await runAnalysesFn(ctx.db, analysisFileSymbols, ctx.rootDir, ctx.opts, ctx.engineOpts);
+        timing.astMs = result.astMs ?? 0;
+        timing.complexityMs = result.complexityMs ?? 0;
+        timing.cfgMs = result.cfgMs ?? 0;
+        timing.dataflowMs = result.dataflowMs ?? 0;
+    }
+    catch (err) {
+        warn(`Analysis phases failed after native build: ${toErrorMessage(err)}`);
+    }
+    // Close nativeDb after analyses — TRUNCATE checkpoint flushes all Rust
+    // WAL writes so JS and external readers can see them. Runs once after
+    // all analysis features complete (not per-feature).
+    if (ctx.nativeDb) {
+        try {
+            ctx.nativeDb.exec('PRAGMA wal_checkpoint(TRUNCATE)');
+        }
+        catch {
+            /* ignore checkpoint errors */
+        }
+        try {
+            ctx.nativeDb.close();
+        }
+        catch {
+            /* ignore close errors */
+        }
+        ctx.nativeDb = undefined;
+        if (ctx.engineOpts) {
+            ctx.engineOpts.nativeDb = undefined;
+            ctx.engineOpts.suspendJsDb = undefined;
+            ctx.engineOpts.resumeJsDb = undefined;
+        }
+    }
+    return timing;
+}
+// ── CHA post-pass helpers ────────────────────────────────────────────────────
+/** Build implementors map: parent/interface name → [child/implementing class names]. */
+function buildChaImplementorsMap(db) {
+    const hierarchyRows = db
+        .prepare(`
+      SELECT src.name AS child_name, tgt.name AS parent_name
+      FROM edges e
+      JOIN nodes src ON e.source_id = src.id
+      JOIN nodes tgt ON e.target_id = tgt.id
+      WHERE e.kind IN ('extends', 'implements')
+    `)
+        .all();
+    const implementors = new Map();
+    for (const row of hierarchyRows) {
+        let list = implementors.get(row.parent_name);
+        if (!list) {
+            list = [];
+            implementors.set(row.parent_name, list);
+        }
+        if (!list.includes(row.child_name))
+            list.push(row.child_name);
+    }
+    return implementors;
+}
+/**
+ * Build RTA set: class names actually instantiated via `new X()`.
+ * Primary query targets `class`-kind nodes (the canonical schema).
+ * Fallback also matches `constructor`/`function`-kind nodes because some native
+ * engine versions record constructor calls against those kinds instead of `class`.
+ * Returns `{ instantiated, noRtaEvidence }` where `noRtaEvidence` means no
+ * constructor-call evidence exists — skip RTA filtering so interface dispatch
+ * still produces edges.
+ */
+function buildChaRtaSet(db) {
+    let rtaRows = db
+        .prepare(`
+      SELECT DISTINCT tgt.name
+      FROM edges e
+      JOIN nodes tgt ON e.target_id = tgt.id
+      WHERE e.kind = 'calls' AND tgt.kind = 'class'
+    `)
+        .all();
+    if (rtaRows.length === 0) {
+        // Fallback: try constructor/function-kind nodes for older native engine schemas
+        rtaRows = db
+            .prepare(`
+        SELECT DISTINCT tgt.name
+        FROM edges e
+        JOIN nodes tgt ON e.target_id = tgt.id
+        WHERE e.kind = 'calls' AND tgt.kind IN ('constructor', 'function')
+        AND INSTR(tgt.name, '.') = 0
+      `)
+            .all();
+    }
+    const instantiated = new Set(rtaRows.map((r) => r.name));
+    const noRtaEvidence = instantiated.size === 0;
+    if (noRtaEvidence) {
+        debug('runPostNativeCha: no constructor-call evidence found — proceeding without RTA filter');
+    }
+    return { instantiated, noRtaEvidence };
+}
+/**
+ * Determine CHA candidate scope for incremental builds.
+ *
+ * Gate A: did a changed file add/change a class hierarchy node?
+ *   A new `extends`/`implements` edge means a previously-untracked implementor
+ *   is now in the hierarchy — unchanged call sites in OTHER files may gain new
+ *   valid expansions, so the full scan is required.
+ *   Note: *removed* class nodes are safe — Rust's `purge_changed_files` runs
+ *   before this post-pass and deletes stale nodes and their hierarchy edges, so
+ *   Gate A queries the post-purge DB. A deleted class returns no row here, which
+ *   is correct: its stale CHA edges were already cleaned up by the Rust purge.
+ *
+ * Gate B: did a changed file add new RTA evidence (`new ConcreteX()`)?
+ *   A new `calls` edge to a class/constructor/function-kind target means the
+ *   instantiated set grew — previously RTA-filtered expansions in unchanged
+ *   caller files become admissible, so the full scan is required.
+ *   (`constructor`/`function` cover the older native engine fallback schema.)
+ *
+ * Returns `true` when the scan should be scoped to changed-file sources only.
+ * Returns `false` (full scan) when changedFiles is null, empty, or either gate fires.
+ */
+function computeChaScope(db, changedFiles) {
+    if (changedFiles === null || changedFiles.length === 0)
+        return false;
+    const CHUNK_SIZE = 500;
+    let gateAFired = false;
+    for (let i = 0; i < changedFiles.length && !gateAFired; i += CHUNK_SIZE) {
+        const chunk = changedFiles.slice(i, i + CHUNK_SIZE);
+        const ph = chunk.map(() => '?').join(',');
+        const row = db
+            .prepare(`SELECT 1 FROM nodes
+         WHERE file IN (${ph})
+         AND kind IN ('class', 'interface', 'trait', 'struct', 'record')
+         LIMIT 1`)
+            .get(...chunk);
+        if (row)
+            gateAFired = true;
+    }
+    let gateBFired = false;
+    if (!gateAFired) {
+        for (let i = 0; i < changedFiles.length && !gateBFired; i += CHUNK_SIZE) {
+            const chunk = changedFiles.slice(i, i + CHUNK_SIZE);
+            const ph = chunk.map(() => '?').join(',');
+            const row = db
+                .prepare(`SELECT 1 FROM edges e
+           JOIN nodes src ON e.source_id = src.id
+           JOIN nodes tgt ON e.target_id = tgt.id
+           WHERE e.kind = 'calls'
+           AND tgt.kind IN ('class', 'interface', 'trait', 'struct', 'record', 'constructor', 'function')
+           AND src.file IN (${ph})
+           LIMIT 1`)
+                .get(...chunk);
+            if (row)
+                gateBFired = true;
+        }
+    }
+    if (!gateAFired && !gateBFired) {
+        debug(`runPostNativeCha: neither gate fired — scoping candidate scan to ${changedFiles.length} changed file(s)`);
+        return true;
+    }
+    debug(`runPostNativeCha: ${gateAFired ? 'Gate A (hierarchy)' : 'Gate B (RTA)'} fired — running full scan`);
+    return false;
+}
+/**
+ * Fetch call→method rows that are candidates for CHA expansion.
+ * When `scopeToChangedFiles` is true, restricts to source nodes in `changedFiles`.
+ */
+function fetchChaCallToMethods(db, changedFiles, scopeToChangedFiles) {
+    if (scopeToChangedFiles && changedFiles && changedFiles.length > 0) {
+        const CHUNK_SIZE = 500;
+        const rows = [];
+        for (let i = 0; i < changedFiles.length; i += CHUNK_SIZE) {
+            const chunk = changedFiles.slice(i, i + CHUNK_SIZE);
+            const ph = chunk.map(() => '?').join(',');
+            const chunkRows = db
+                .prepare(`SELECT e.source_id, src.name AS caller_name, tgt.name AS method_name, src.file AS caller_file
+           FROM edges e
+           JOIN nodes tgt ON e.target_id = tgt.id
+           JOIN nodes src ON e.source_id = src.id
+           WHERE e.kind = 'calls' AND tgt.kind = 'method'
+           AND INSTR(tgt.name, '.') > 0
+           AND (e.technique IS NULL OR e.technique != 'cha-expanded')
+           AND src.file IN (${ph})`)
+                .all(...chunk);
+            rows.push(...chunkRows);
+        }
+        return rows;
+    }
+    return db
+        .prepare(`
+      SELECT e.source_id, src.name AS caller_name, tgt.name AS method_name, src.file AS caller_file
+      FROM edges e
+      JOIN nodes tgt ON e.target_id = tgt.id
+      JOIN nodes src ON e.source_id = src.id
+      WHERE e.kind = 'calls' AND tgt.kind = 'method'
+      AND INSTR(tgt.name, '.') > 0
+      AND (e.technique IS NULL OR e.technique != 'cha-expanded')
+    `)
+        .all();
+}
+/**
+ * BFS-expand CHA call edges and insert new edges into the DB.
+ * Returns `{ newEdgeCount, affectedFiles }` for role re-classification scoping.
+ */
+function expandChaEdges(db, callToMethods, implementors, instantiated, noRtaEvidence) {
+    const affectedFiles = new Set();
+    // Seed seen-pairs only from the source_ids we'll be expanding — avoids loading every
+    // call edge in the DB (which would be O(all edges)) for large codebases.
+    const seen = new Set();
+    if (callToMethods.length > 0) {
+        const sourceIds = [...new Set(callToMethods.map((r) => r.source_id))];
+        const CHUNK_SIZE = 500;
+        for (let i = 0; i < sourceIds.length; i += CHUNK_SIZE) {
+            const chunk = sourceIds.slice(i, i + CHUNK_SIZE);
+            const placeholders = chunk.map(() => '?').join(',');
+            const existingPairs = db
+                .prepare(`SELECT source_id, target_id FROM edges WHERE kind = 'calls' AND source_id IN (${placeholders})`)
+                .all(...chunk);
+            for (const e of existingPairs)
+                seen.add(`${e.source_id}|${e.target_id}`);
+        }
+    }
+    // No LIMIT: multiple files can define the same qualified name in a monorepo.
+    const findMethodStmt = db.prepare(`SELECT id, file AS method_file FROM nodes WHERE name = ? AND kind = 'method'`);
+    const newEdges = [];
+    let newEdgeCount = 0;
+    for (const { source_id, method_name, caller_file } of callToMethods) {
+        const dotIdx = method_name.indexOf('.');
+        if (dotIdx === -1)
+            continue;
+        const typeName = method_name.slice(0, dotIdx);
+        const methodSuffix = method_name.slice(dotIdx + 1);
+        // BFS over the implementors map — handles multi-level hierarchies where
+        // abstract/non-instantiated classes sit between the call-site type and
+        // the concrete leaf implementations (issue #1311).
+        const bfsQueue = [typeName];
+        const bfsVisited = new Set([typeName]);
+        while (bfsQueue.length > 0) {
+            const current = bfsQueue.shift();
+            const children = implementors.get(current);
+            if (!children?.length)
+                continue;
+            for (const cls of children) {
+                if (bfsVisited.has(cls))
+                    continue;
+                bfsVisited.add(cls);
+                if (noRtaEvidence || instantiated.has(cls)) {
+                    const qualifiedName = `${cls}.${methodSuffix}`;
+                    const methodNodes = findMethodStmt.all(qualifiedName);
+                    for (const methodNode of methodNodes) {
+                        if (methodNode.id === source_id)
+                            continue; // skip self-loops
+                        const key = `${source_id}|${methodNode.id}`;
+                        if (seen.has(key))
+                            continue;
+                        seen.add(key);
+                        const conf = CHA_TYPED_DISPATCH_CONFIDENCE;
+                        newEdges.push([source_id, methodNode.id, 'calls', conf, 0, 'cha-expanded']);
+                        newEdgeCount++;
+                        if (caller_file)
+                            affectedFiles.add(caller_file);
+                        if (methodNode.method_file)
+                            affectedFiles.add(methodNode.method_file);
+                    }
+                }
+                // Always traverse children — non-instantiated classes may have instantiated subclasses.
+                bfsQueue.push(cls);
+            }
+        }
+    }
+    if (newEdges.length > 0) {
+        db.transaction(() => batchInsertEdges(db, newEdges))();
+        // Account for post-pass edges excluded from the build summary line (#1452),
+        // mirroring the this/super dispatch post-pass insertion log.
+        debug(`CHA expansion post-pass: inserted ${newEdgeCount} edge(s)`);
+    }
+    return { newEdgeCount, affectedFiles };
+}
+/**
+ * Phase 8.6: CHA expansion post-pass for the native orchestrator path.
+ *
+ * The Rust build pipeline resolves typed receiver calls (e.g. `worker.doWork()`
+ * where `worker: IWorker`) to the interface method declaration only.  This
+ * post-pass reads the class hierarchy (via `implements`/`extends` edges) and
+ * instantiated types (via `calls` edges to class nodes) from the DB and expands
+ * each call to an interface/abstract method to ALL RTA-filtered concrete
+ * implementations.
+ *
+ * Note: `this`/`super` dispatch is handled separately by `runPostNativeThisDispatch`,
+ * which WASM-re-parses JS/TS files to obtain raw call site receiver info.
+ *
+ * `changedFiles` controls candidate scoping on incremental builds:
+ *   - null  → full build; scan all call→method edges (existing behaviour).
+ *   - array → incremental; two cheap gate queries decide scope:
+ *       Gate A: any class/interface/trait/struct/record nodes in changed files?
+ *               If yes, a new implementor may have appeared — full scan required.
+ *       Gate B: any `calls` edges from changed-file sources targeting
+ *               class/constructor/function-kind nodes? If yes, the RTA set may
+ *               have grown (also covers the older-schema fallback where
+ *               constructor calls target `constructor`/`function` nodes instead
+ *               of `class` nodes) — full scan required.
+ *       If neither gate fires: scope `callToMethods` to `src.file IN changedFiles`
+ *       (safe because no hierarchy or RTA evidence changed).
+ *
+ * Returns the count of newly inserted CHA edges plus the set of files containing
+ * the new edges' endpoints, so the caller can scope role re-classification to the
+ * nodes whose fan-in/out actually changed. A zero count means no edges were added
+ * and role re-classification is unnecessary.
+ */
+function runPostNativeCha(db, changedFiles) {
+    const affectedFiles = new Set();
+    const empty = { newEdgeCount: 0, affectedFiles };
+    // Fast guard: no hierarchy edges → no CHA work
+    const hasHierarchy = db
+        .prepare(`SELECT 1 FROM edges WHERE kind IN ('extends', 'implements') LIMIT 1`)
+        .get();
+    if (!hasHierarchy)
+        return empty;
+    const implementors = buildChaImplementorsMap(db);
+    if (implementors.size === 0)
+        return empty;
+    const { instantiated, noRtaEvidence } = buildChaRtaSet(db);
+    const scopeToChangedFiles = computeChaScope(db, changedFiles);
+    const callToMethods = fetchChaCallToMethods(db, changedFiles, scopeToChangedFiles);
+    return expandChaEdges(db, callToMethods, implementors, instantiated, noRtaEvidence);
+}
+// Extensions where `this`/`super` dispatch can occur (JS/TS family)
+const THIS_DISPATCH_EXTS = new Set(['.js', '.ts', '.tsx', '.jsx', '.mjs', '.cjs', '.mts', '.cts']);
+// ── this/super dispatch post-pass helpers ───────────────────────────────────
+/**
+ * Build parents map: child class → direct parent class (from `extends` edges).
+ * May be empty when only func-prop methods exist (no class inheritance) —
+ * resolveThisDispatch handles that case via direct class-prefix lookup.
+ */
+function buildThisDispatchParentsMap(db, hasExtends) {
+    const parents = new Map();
+    if (!hasExtends)
+        return parents;
+    const parentRows = db
+        .prepare(`
+      SELECT src.name AS child_name, tgt.name AS parent_name
+      FROM edges e
+      JOIN nodes src ON e.source_id = src.id
+      JOIN nodes tgt ON e.target_id = tgt.id
+      WHERE e.kind = 'extends'
+    `)
+        .all();
+    for (const row of parentRows) {
+        if (!parents.has(row.child_name))
+            parents.set(row.child_name, row.parent_name);
+    }
+    return parents;
+}
+/**
+ * Determine the set of relative file paths to re-parse for this/super dispatch.
+ *
+ * On a full build we do NOT re-parse every JS/TS file — that would WASM-parse
+ * the entire project on top of the native pass, causing a massive regression
+ * (measured: +358% ms/file on codegraph itself). Instead we restrict to files
+ * that are part of the class inheritance hierarchy OR that contain dot-named
+ * method nodes (func-prop assignments whose bodies may call `this.sibling()`).
+ */
+function selectThisDispatchFiles(db, changedFiles, isFullBuild) {
+    if (isFullBuild || !changedFiles) {
+        const rows = db
+            .prepare(`
+        SELECT DISTINCT file FROM (
+          SELECT src.file AS file
+          FROM edges e
+          JOIN nodes src ON e.source_id = src.id
+          WHERE e.kind = 'extends' AND src.file IS NOT NULL
+          UNION
+          SELECT tgt.file AS file
+          FROM edges e
+          JOIN nodes tgt ON e.target_id = tgt.id
+          WHERE e.kind = 'extends' AND tgt.file IS NOT NULL
+          UNION
+          -- Files with func-prop method definitions (e.g. f.h = function(){this.g()}).
+          -- Only include files where the method's owner prefix is NOT a known class name —
+          -- this keeps the re-parse set small (func-prop files only, not all class-method files).
+          -- AND name IS NOT NULL guards the NOT IN sub-select: if any class node had a NULL
+          -- name the entire NOT IN clause would silently return no rows (SQL NULL semantics).
+          SELECT n.file AS file
+          FROM nodes n
+          WHERE n.kind = 'method'
+          AND INSTR(n.name, '.') > 0
+          AND n.file IS NOT NULL
+          AND SUBSTR(n.name, 1, INSTR(n.name, '.') - 1) NOT IN (
+            SELECT name FROM nodes WHERE kind IN ('class', 'struct', 'interface', 'type')
+            AND name IS NOT NULL
+          )
+        )
+      `)
+            .all();
+        return rows
+            .map((r) => r.file)
+            .filter((f) => THIS_DISPATCH_EXTS.has(path.extname(f).toLowerCase()));
+    }
+    // NOTE: Only files explicitly listed in changedFiles are re-parsed.
+    // If a parent-class method is replaced (new node ID) but the child file is
+    // unchanged, the stale super.method() edge is not refreshed here. A full
+    // rebuild (isFullBuild=true) is required to recover in that scenario.
+    return changedFiles.filter((f) => THIS_DISPATCH_EXTS.has(path.extname(f).toLowerCase()));
+}
+/**
+ * Re-parse files via native (preferred) + WASM fallback to obtain call sites
+ * with receiver info. Returns a map of relPath → calls array.
+ *
+ * The native engine is preferred: this pass only runs after a native
+ * orchestrator build, so the addon is already loaded and re-parses the
+ * hierarchy file set in single-digit milliseconds with the same
+ * receiver-annotated call sites as the WASM extractor. Booting the WASM
+ * runtime here instead cost ~40–110ms per full build (in-process
+ * web-tree-sitter + grammar init dominated) — part of the v3.12.0
+ * publish-gate regression. Files the native engine cannot parse (extension
+ * outside NATIVE_SUPPORTED_EXTENSIONS, e.g. .mts/.cts) and native parse
+ * failures fall back to the WASM backfill path so the sweep stays complete.
+ */
+async function parseFilesForThisDispatch(absFiles, rootDir) {
+    const callsByRel = new Map();
+    const nativeAbs = absFiles.filter((f) => NATIVE_SUPPORTED_EXTENSIONS.has(path.extname(f).toLowerCase()));
+    // Track native-supported files that returned null (per-file parse error) so
+    // they can be included in the WASM fallback set below, ensuring no file's
+    // this/super call sites are silently discarded.
+    const nativeNullFiles = new Set();
+    let nativeParsed = false;
+    if (nativeAbs.length > 0) {
+        const native = loadNative();
+        if (native) {
+            try {
+                const results = native.parseFiles(nativeAbs, rootDir, false, false);
+                for (let i = 0; i < results.length; i++) {
+                    const r = results[i];
+                    if (!r) {
+                        // Per-file parse failure — fall back to WASM for this file.
+                        const abs = nativeAbs[i];
+                        if (abs)
+                            nativeNullFiles.add(abs);
+                        continue;
+                    }
+                    callsByRel.set(normalizePath(path.relative(rootDir, r.file)), r.calls ?? []);
+                }
+                nativeParsed = true;
+            }
+            catch (e) {
+                debug(`this-dispatch native re-parse failed, falling back to WASM: ${toErrorMessage(e)}`);
+            }
+        }
+    }
+    // WASM handles: (a) non-native extensions (e.g. .mts/.cts), (b) the entire
+    // file list when the native batch threw, and (c) individual files where the
+    // native addon returned null (per-file parse error).
+    const wasmAbs = nativeParsed
+        ? [
+            ...absFiles.filter((f) => !NATIVE_SUPPORTED_EXTENSIONS.has(path.extname(f).toLowerCase())),
+            ...nativeNullFiles,
+        ]
+        : absFiles;
+    const wasmResults = wasmAbs.length > 0
+        ? await parseFilesWasmForBackfill(wasmAbs, rootDir, { symbolsOnly: true })
+        : new Map();
+    for (const [relPath, symbols] of wasmResults) {
+        callsByRel.set(relPath, symbols.calls ?? []);
+    }
+    return { callsByRel, wasmResults };
+}
+/** Emit this/super dispatch edges from re-parsed call sites. */
+function emitThisDispatchEdges(db, callsByRel, chaCtx, lookup, seen) {
+    // Find the innermost containing method/function for a call at `line` in `file`.
+    // COALESCE maps NULL end_line to a large sentinel so unbounded nodes sort last
+    // (SQLite ASC orders NULLs first, so a raw `end_line - line` would pick them first).
+    const findCallerByLineStmt = db.prepare(`
+    SELECT id, name FROM nodes
+    WHERE file = ? AND kind IN ('method', 'function')
+    AND line <= ? AND (end_line IS NULL OR end_line >= ?)
+    ORDER BY COALESCE(end_line - line, 999999999) ASC
+    LIMIT 1
+  `);
+    const newEdges = [];
+    const targetIds = new Set();
+    const affectedFiles = new Set();
+    for (const [relPath, calls] of callsByRel) {
+        for (const call of calls) {
+            // Only 'this' and 'super' are class-instance receivers in JS/TS.
+            // 'self' refers to WindowOrWorkerGlobalScope — not a class instance — so
+            // filtering it here prevents spurious dispatch edges from Worker call sites.
+            if (call.receiver !== 'this' && call.receiver !== 'super')
+                continue;
+            const callerRow = findCallerByLineStmt.get(relPath, call.line, call.line);
+            if (!callerRow)
+                continue;
+            const targets = resolveThisDispatch(call.name, callerRow.name, call.receiver, chaCtx, lookup, relPath);
+            for (const t of targets) {
+                if (t.id === callerRow.id)
+                    continue; // skip self-loops
+                const key = `${callerRow.id}|${t.id}`;
+                if (seen.has(key))
+                    continue;
+                seen.add(key);
+                const conf = computeConfidence(relPath, t.file, null) - CHA_DISPATCH_PENALTY;
+                if (conf <= 0)
+                    continue;
+                // Tag super-dispatch edges distinctly so runPostNativeCha can exclude them
+                // from further CHA expansion (super calls are not virtual dispatch).
+                const technique = call.receiver === 'super' ? 'super-dispatch' : 'cha';
+                newEdges.push([callerRow.id, t.id, 'calls', conf, 0, technique]);
+                targetIds.add(t.id);
+                affectedFiles.add(relPath);
+                if (t.file)
+                    affectedFiles.add(t.file);
+            }
+        }
+    }
+    return { newEdges, targetIds, affectedFiles };
+}
+/** Free WASM parse trees after this-dispatch post-pass to prevent memory leaks. */
+function cleanupThisDispatchWasmTrees(wasmResults) {
+    for (const [, symbols] of wasmResults) {
+        const tree = symbols._tree;
+        if (tree && typeof tree.delete === 'function') {
+            try {
+                tree.delete();
+            }
+            catch {
+                /* ignore cleanup errors */
+            }
+        }
+        symbols._tree = undefined;
+        symbols._langId = undefined;
+    }
+}
+/**
+ * Phase 8.5: this/super dispatch post-pass for the native orchestrator path.
+ *
+ * The Rust build pipeline resolves typed receiver calls but does NOT persist raw
+ * unresolved call site receiver info (e.g. `this`, `super`) to the DB. This
+ * hybrid post-pass re-parses JS/TS/TSX files via WASM to collect call sites with
+ * `this`/`super` receivers, then resolves them through the class hierarchy stored
+ * in DB `extends` edges — mirroring what `buildChaPostPass` does on the WASM path.
+ *
+ * Also handles function-as-object-property methods (`f.h = function() { this.g() }`):
+ * these use `this` to reference sibling properties on the same object (`f`), so
+ * `resolveThisDispatch` resolves them by treating the dot-prefix of the caller name
+ * (`f` from `f.h`) as the class and looking up `f.g` directly — no `extends` edge needed.
+ *
+ * Runs when either `extends` edges exist (class inheritance) OR dot-named `method`
+ * nodes exist (func-prop assignments); skips only when neither is present.
+ */
+async function runPostNativeThisDispatch(db, rootDir, changedFiles, isFullBuild) {
+    const t0 = performance.now();
+    // Fast guard: need at least one extends edge (class inheritance) OR a dot-named
+    // method node (func-prop assignment: `f.h = function() { this.g() }`) for
+    // this/super dispatch to produce any edges.
+    const hasExtends = db.prepare(`SELECT 1 FROM edges WHERE kind = 'extends' LIMIT 1`).get();
+    const hasFuncPropMethod = db
+        .prepare(`SELECT 1 FROM nodes WHERE kind = 'method' AND INSTR(name, '.') > 0 LIMIT 1`)
+        .get();
+    const emptyResult = {
+        elapsedMs: 0,
+        targetIds: new Set(),
+        affectedFiles: new Set(),
+    };
+    if (!hasExtends && !hasFuncPropMethod)
+        return emptyResult;
+    const parents = buildThisDispatchParentsMap(db, hasExtends);
+    // Note: parents may be empty when hasFuncPropMethod but !hasExtends — that is
+    // intentional. resolveThisDispatch still resolves `this.g()` inside `f.h` by
+    // treating `f` (the dot-prefix of callerName `f.h`) as the class and looking
+    // up `f.g` directly via lookup.byName(), without traversing the parents chain.
+    const chaCtx = {
+        implementors: new Map(), // not needed for this/super resolution
+        parents,
+        instantiatedTypes: new Set(), // not needed for this/super resolution
+    };
+    const relFiles = selectThisDispatchFiles(db, changedFiles, isFullBuild);
+    if (relFiles.length === 0)
+        return emptyResult;
+    // DB-backed CallNodeLookup — resolveThisDispatch only calls byName()
+    const findByNameStmt = db.prepare(`SELECT id, file, kind FROM nodes WHERE name = ?`);
+    const lookup = {
+        byName: (name) => findByNameStmt.all(name),
+        byNameAndFile: (name, file) => findByNameStmt.all(name).filter((n) => n.file === file),
+        isBarrel: () => false,
+        resolveBarrel: () => null,
+        nodeId: () => undefined,
+    };
+    // Seed seen-pairs from existing call edges on source nodes in our file set
+    const seen = new Set();
+    const CHUNK = 500;
+    for (let i = 0; i < relFiles.length; i += CHUNK) {
+        const chunk = relFiles.slice(i, i + CHUNK);
+        const ph = chunk.map(() => '?').join(',');
+        const rows = db
+            .prepare(`SELECT e.source_id, e.target_id
+         FROM edges e
+         JOIN nodes n ON e.source_id = n.id
+         WHERE e.kind = 'calls' AND n.file IN (${ph})`)
+            .all(...chunk);
+        for (const r of rows)
+            seen.add(`${r.source_id}|${r.target_id}`);
+    }
+    const absFiles = relFiles.map((f) => path.join(rootDir, f));
+    const { callsByRel, wasmResults } = await parseFilesForThisDispatch(absFiles, rootDir);
+    const { newEdges, targetIds, affectedFiles } = emitThisDispatchEdges(db, callsByRel, chaCtx, lookup, seen);
+    if (newEdges.length > 0) {
+        db.transaction(() => batchInsertEdges(db, newEdges))();
+        debug(`this/super dispatch post-pass: inserted ${newEdges.length} edge(s)`);
+    }
+    cleanupThisDispatchWasmTrees(wasmResults);
+    return { elapsedMs: performance.now() - t0, targetIds, affectedFiles };
+}
+/** Format timing result from native orchestrator phases + JS post-processing. */
+function formatNativeTimingResult(p, structurePatchMs, analysisTiming, postPass) {
+    return {
+        phases: {
+            setupMs: +(p.setupMs ?? 0).toFixed(1),
+            collectMs: +(p.collectMs ?? 0).toFixed(1),
+            detectMs: +(p.detectMs ?? 0).toFixed(1),
+            parseMs: +(p.parseMs ?? 0).toFixed(1),
+            insertMs: +(p.insertMs ?? 0).toFixed(1),
+            resolveMs: +(p.resolveMs ?? 0).toFixed(1),
+            edgesMs: +(p.edgesMs ?? 0).toFixed(1),
+            structureMs: +((p.structureMs ?? 0) + structurePatchMs).toFixed(1),
+            rolesMs: +(p.rolesMs ?? 0).toFixed(1),
+            gapDetectMs: +postPass.gapDetectMs.toFixed(1),
+            chaMs: +postPass.chaMs.toFixed(1),
+            thisDispatchMs: +postPass.thisDispatchMs.toFixed(1),
+            reclassifyMs: +postPass.reclassifyMs.toFixed(1),
+            techniqueBackfillMs: +postPass.techniqueBackfillMs.toFixed(1),
+            astMs: +(analysisTiming.astMs ?? 0).toFixed(1),
+            complexityMs: +(analysisTiming.complexityMs ?? 0).toFixed(1),
+            cfgMs: +(analysisTiming.cfgMs ?? 0).toFixed(1),
+            dataflowMs: +(analysisTiming.dataflowMs ?? 0).toFixed(1),
+            finalizeMs: +(p.finalizeMs ?? 0).toFixed(1),
+        },
+    };
+}
+/**
+ * Compute the WASM-only files present in the DB but missing from disk (#1073).
+ *
+ * Returns relative paths that:
+ *   - appear in `existingNodes` or `existingHashes` (in DB),
+ *   - are absent from `expected` (not on disk),
+ *   - have an extension installed for WASM, AND
+ *   - have an extension NOT covered by `nativeSupported` — Rust's
+ *     `purge_changed_files` handles deletion for natively-supported extensions
+ *     via its own `detect_removed_files`, so the caller must not double-purge.
+ *
+ * Extensions are lowercased before lookup to match the registry and Rust's
+ * `LanguageKind::from_extension` (which normalises case for the languages
+ * where both cases are conventional, e.g. R's `.r` / `.R`).
+ *
+ * DB paths are forced to forward slashes before comparison with `expected`
+ * (which is always normalised). The on-disk invariant is that DB rows are
+ * written with forward slashes, but a stale row written by older code on
+ * Windows could carry back-slashes — normalising here makes the comparison
+ * platform-safe and prevents false-positive purges of live rows. We replace
+ * `\\` explicitly (rather than calling `normalizePath`, which only touches
+ * `path.sep`) so the defence works when running on POSIX against a DB that
+ * was migrated from Windows.
+ *
+ * Exported for unit testing.
+ */
+export function computeWasmOnlyStaleFiles(input) {
+    const { existingNodes, existingHashes, expected, installedExts, nativeSupported } = input;
+    const stale = [];
+    const seen = new Set();
+    const consider = (rawRel) => {
+        const rel = rawRel.replace(/\\/g, '/');
+        if (expected.has(rel) || seen.has(rel))
+            return;
+        const ext = path.extname(rel).toLowerCase();
+        if (nativeSupported.has(ext))
+            return;
+        if (!installedExts.has(ext))
+            return;
+        seen.add(rel);
+        // Push the ORIGINAL raw path (not the normalised form) so the eventual
+        // `DELETE FROM nodes WHERE file = ?` predicate in `purgeFilesData`
+        // matches the actual stored row. The dedup `seen` set keeps the
+        // normalised form so a file written once with `\` and once with `/`
+        // is still treated as one entry — but the value the SQL sees has to
+        // be byte-identical to what's on disk in the DB.
+        stale.push(rawRel);
+    };
+    for (const rel of existingNodes)
+        consider(rel);
+    for (const rel of existingHashes)
+        consider(rel);
+    return stale;
+}
+/**
+ * Group relative paths by their lowercased extension. Shape matches the bucket
+ * type that `formatDropExtensionSummary` consumes, so callers can render a
+ * log-friendly per-extension summary without going through `classifyNativeDrops`
+ * when the reason is already known (e.g. the stale-purge path where every path
+ * is guaranteed `unsupported-by-native`).
+ */
+function groupByExtension(relPaths) {
+    const buckets = new Map();
+    for (const rel of relPaths) {
+        const ext = path.extname(rel).toLowerCase();
+        let list = buckets.get(ext);
+        if (!list) {
+            list = [];
+            buckets.set(ext, list);
+        }
+        list.push(rel);
+    }
+    return buckets;
+}
+/**
+ * Return the subset of relative paths that are gitignored in `rootDir`.
+ *
+ * Runs `git check-ignore --stdin` with all candidate paths piped in. Any
+ * path that git echoes back is gitignored. Fails silently (returns an empty
+ * set) when git is unavailable, the directory is not a git repo, or the
+ * check-ignore call throws — the gap-detection logic handles those cases
+ * gracefully without this filter.
+ *
+ * Uses relative paths (forward-slash separated) as both input and output so
+ * the result set can be matched directly against the `expected` set in
+ * `detectDroppedLanguageGap` without any further path manipulation.
+ */
+function queryGitIgnoredFiles(rootDir, relPaths) {
+    const ignored = new Set();
+    const paths = [...relPaths];
+    if (paths.length === 0)
+        return ignored;
+    try {
+        const stdin = paths.join('\n');
+        const output = execFileSync('git', ['check-ignore', '--stdin'], {
+            cwd: rootDir,
+            input: stdin,
+            encoding: 'utf-8',
+            maxBuffer: 100 * 1024 * 1024,
+            // git check-ignore exits with 1 when none of the paths are ignored —
+            // that is not an error for our purposes. stdio: 'pipe' lets us capture
+            // stdout without swallowing stderr, and the try/catch handles the
+            // non-zero exit from execFileSync when ALL paths are non-ignored
+            // (exit code 1 from git check-ignore means "no matches").
+            stdio: ['pipe', 'pipe', 'pipe'],
+        });
+        for (const line of output.split('\n')) {
+            const trimmed = normalizePath(line.trim());
+            if (trimmed)
+                ignored.add(trimmed);
+        }
+    }
+    catch (e) {
+        // Exit code 1 means no paths were ignored — not an error. Any other
+        // failure (git unavailable, not a repo, etc.) is silently swallowed
+        // so the caller proceeds with the unfiltered set.
+        const exitCode = e.status;
+        if (exitCode !== 1) {
+            debug(`queryGitIgnoredFiles: git check-ignore failed: ${toErrorMessage(e)}`);
+        }
+        // On exit code 1, output is empty so ignored stays empty — correct.
+        // On other errors we also proceed with the empty set (safe degradation).
+    }
+    return ignored;
+}
+/**
+ * Detect files the native orchestrator silently dropped.
+ *
+ * Walks the filesystem and compares against `nodes` + `file_hashes`. A file
+ * is "missing" if it's absent from EITHER table — both must be present for
+ * the fast-skip pre-flight (#1054) to work, and the two can diverge (e.g.
+ * legacy DBs where `nodes` was populated but `file_hashes` was not).
+ *
+ * Restricted to files with an installed WASM grammar; extensions in
+ * `LANGUAGE_REGISTRY` without a shipped grammar (e.g. groovy on minimal
+ * installs) can't be parsed by either engine, so they're not a native
+ * regression — excluding them keeps the warn count in
+ * `backfillNativeDroppedFiles` meaningful.
+ *
+ * Also detects WASM-only files deleted from disk (#1073). Rust's
+ * `detect_removed_files` filter (#1070) skips files outside its supported
+ * extensions, so deletions of WASM-only languages don't reach the native
+ * purge path; the rest of the backfill only inserts rows, so without this
+ * step stale `nodes`/`file_hashes` rows would linger across incremental
+ * rebuilds until the next full rebuild.
+ *
+ * Cheap (no DB handoff, no parsing): used both to gate the backfill call
+ * and as its working set. NativeDbProxy supports `.prepare().all()`, so
+ * this works whether `ctx.db` is a proxy or a real better-sqlite3
+ * connection — letting us skip the close-native / reopen-better-sqlite3
+ * cost when there's nothing to backfill.
+ */
+function detectDroppedLanguageGap(ctx) {
+    const collected = collectFilesUtil(ctx.rootDir, [], ctx.config, new Set());
+    const expectedRaw = collected.files.map((f) => normalizePath(path.relative(ctx.rootDir, f)));
+    // The native Rust engine uses the `ignore` crate with git_ignore(true), so it
+    // respects .gitignore and never processes gitignored files. The JS collectFiles
+    // walker has no gitignore awareness, so without this filter gitignored files
+    // (e.g. NAPI-RS generated crates/codegraph-core/index.js / index.d.ts) appear
+    // in `expected` but not in the DB, causing a spurious "native extractor bug"
+    // WARN and triggering an unnecessary WASM backfill (#1626).
+    const gitIgnored = queryGitIgnoredFiles(ctx.rootDir, expectedRaw);
+    const expected = new Set(gitIgnored.size > 0 ? expectedRaw.filter((r) => !gitIgnored.has(r)) : expectedRaw);
+    const existingNodeRows = ctx.db
+        .prepare("SELECT DISTINCT file FROM nodes WHERE kind = 'file'")
+        .all();
+    const existingNodes = new Set(existingNodeRows.map((r) => r.file));
+    let existingHashes = new Set();
+    try {
+        const existingHashRows = ctx.db
+            .prepare('SELECT DISTINCT file FROM file_hashes')
+            .all();
+        existingHashes = new Set(existingHashRows.map((r) => r.file));
+    }
+    catch (e) {
+        // file_hashes table may not exist on legacy DBs; treat as fully missing
+        // so the backfill writes rows on the upsert path below.
+        debug(`detectDroppedLanguageGap: file_hashes read failed (table may not exist): ${toErrorMessage(e)}`);
+    }
+    const installedExts = getInstalledWasmExtensions();
+    const missingRel = [];
+    const missingAbs = [];
+    for (const rel of expected) {
+        if (existingNodes.has(rel) && existingHashes.has(rel))
+            continue;
+        const ext = path.extname(rel).toLowerCase();
+        if (!installedExts.has(ext))
+            continue;
+        missingRel.push(rel);
+        missingAbs.push(path.join(ctx.rootDir, rel));
+    }
+    const staleRel = computeWasmOnlyStaleFiles({
+        existingNodes,
+        existingHashes,
+        expected,
+        installedExts,
+        nativeSupported: NATIVE_SUPPORTED_EXTENSIONS,
+    });
+    return { missingRel, missingAbs, staleRel };
+}
+// ── backfillNativeDroppedFiles helpers ───────────────────────────────────────
+/** Purge stale WASM-only files deleted from disk (#1073). */
+function purgeStaleWasmOnlyFiles(db, staleRel) {
+    // `computeWasmOnlyStaleFiles` guarantees every path here has an extension
+    // outside NATIVE_SUPPORTED_EXTENSIONS, so `classifyNativeDrops` would
+    // always bucket 100% into `unsupported-by-native`. Build the extension
+    // summary directly to avoid a redundant classification pass.
+    const staleByExt = groupByExtension(staleRel);
+    info(`Detected ${staleRel.length} deleted WASM-only file(s) across ${staleByExt.size} extension(s) the native orchestrator skipped; purging stale rows:${formatDropExtensionSummary(staleByExt)}`);
+    purgeFilesData(db, staleRel);
+}
+/**
+ * Classify and log dropped file buckets.
+ * Three-way split of native-extractor-failure files:
+ *   realFailureBuckets  — WASM found symbols → real Rust extractor bug (WARN)
+ *   emptyFileBuckets    — WASM parsed but found 0 symbols → gitignored/empty (debug)
+ *   wasmSkipBuckets     — WASM skipped entirely → no file-node insert (debug)
+ */
+function classifyAndLogDroppedFiles(missingRel, wasmParsedFiles, wasmFoundSymbols) {
+    const { byReason, totals } = classifyNativeDrops(missingRel);
+    if (totals['unsupported-by-native'] > 0) {
+        const buckets = byReason['unsupported-by-native'];
+        info(`Native orchestrator skipped ${totals['unsupported-by-native']} file(s) across ${buckets.size} extension(s) in languages without a Rust extractor; backfilling via WASM:${formatDropExtensionSummary(buckets)}`);
+    }
+    if (totals['native-extractor-failure'] > 0) {
+        const allFailurePaths = byReason['native-extractor-failure'];
+        const realFailureBuckets = new Map();
+        const emptyFileBuckets = new Map();
+        const wasmSkipBuckets = new Map();
+        for (const [ext, paths] of allFailurePaths) {
+            for (const relPath of paths) {
+                let bucket;
+                if (wasmFoundSymbols.has(relPath)) {
+                    bucket = realFailureBuckets;
+                }
+                else if (wasmParsedFiles.has(relPath)) {
+                    bucket = emptyFileBuckets;
+                }
+                else {
+                    bucket = wasmSkipBuckets;
+                }
+                let list = bucket.get(ext);
+                if (!list) {
+                    list = [];
+                    bucket.set(ext, list);
+                }
+                list.push(relPath);
+            }
+        }
+        if (realFailureBuckets.size > 0) {
+            const realCount = [...realFailureBuckets.values()].reduce((s, a) => s + a.length, 0);
+            warn(`Native orchestrator dropped ${realCount} file(s) across ${realFailureBuckets.size} extension(s) in natively-supported languages — likely a Rust extractor bug. Backfilling via WASM:${formatDropExtensionSummary(realFailureBuckets)}`);
+        }
+        if (emptyFileBuckets.size > 0) {
+            const emptyCount = [...emptyFileBuckets.values()].reduce((s, a) => s + a.length, 0);
+            debug(`Native orchestrator skipped ${emptyCount} file(s) in natively-supported languages that also produced 0 symbols via WASM (likely gitignored or empty); backfilling file nodes:${formatDropExtensionSummary(emptyFileBuckets)}`);
+        }
+        if (wasmSkipBuckets.size > 0) {
+            const skipCount = [...wasmSkipBuckets.values()].reduce((s, a) => s + a.length, 0);
+            debug(`Native orchestrator skipped ${skipCount} file(s) in natively-supported languages that WASM also could not parse (unregistered extension or parse error); no file-node inserted:${formatDropExtensionSummary(wasmSkipBuckets)}`);
+        }
+    }
+}
+/** Insert node rows for all backfilled files and mark exported symbols. */
+function insertBackfilledNodes(db, wasmResults) {
+    const rows = [];
+    const exportKeys = [];
+    for (const [relPath, symbols] of wasmResults) {
+        // File row — mirrors insertDefinitionsAndExports: qualified_name is null.
+        rows.push([relPath, 'file', relPath, 0, null, null, null, null, null]);
+        for (const def of symbols.definitions ?? []) {
+            // Populate qualified_name/scope the same way the JS fallback does so
+            // downstream queries (cross-file references, "go to definition") find
+            // these symbols.
+            const dotIdx = def.name.lastIndexOf('.');
+            const scope = dotIdx !== -1 ? def.name.slice(0, dotIdx) : null;
+            rows.push([
+                def.name,
+                def.kind,
+                relPath,
+                def.line,
+                def.endLine ?? null,
+                null,
+                def.name,
+                scope,
+                def.visibility ?? null,
+            ]);
+        }
+        // Exports: insert the row (INSERT OR IGNORE — a matching definition row
+        // is a no-op) and queue a key for the second-pass exported=1 update, so
+        // queries filtering on exported=1 find backfilled symbols (#970).
+        for (const exp of symbols.exports ?? []) {
+            rows.push([exp.name, exp.kind, relPath, exp.line, null, null, exp.name, null, null]);
+            exportKeys.push([exp.name, exp.kind, relPath, exp.line]);
+        }
+    }
+    batchInsertNodes(db, rows);
+    // Mark exported symbols in batches — mirrors insertDefinitionsAndExports.
+    if (exportKeys.length > 0) {
+        const EXPORT_CHUNK = 500;
+        const exportStmtCache = new Map();
+        for (let i = 0; i < exportKeys.length; i += EXPORT_CHUNK) {
+            const end = Math.min(i + EXPORT_CHUNK, exportKeys.length);
+            const chunkSize = end - i;
+            let updateStmt = exportStmtCache.get(chunkSize);
+            if (!updateStmt) {
+                const conditions = Array.from({ length: chunkSize }, () => '(name = ? AND kind = ? AND file = ? AND line = ?)').join(' OR ');
+                updateStmt = db.prepare(`UPDATE nodes SET exported = 1 WHERE ${conditions}`);
+                exportStmtCache.set(chunkSize, updateStmt);
+            }
+            const vals = [];
+            for (let j = i; j < end; j++) {
+                const k = exportKeys[j];
+                vals.push(k[0], k[1], k[2], k[3]);
+            }
+            updateStmt.run(...vals);
+        }
+    }
+}
+/**
+ * Persist file_hashes rows for every backfilled file.
+ *
+ * The Rust orchestrator only hashes files it parsed itself, so without this
+ * step files in optional-language extensions (e.g. .clj when no Rust extractor
+ * exists) would be missing from `file_hashes` — permanently breaking the JS-side
+ * fast-skip pre-flight (#1054), which rejects on `collected file missing
+ * from file_hashes` and forces every no-op rebuild back through the full
+ * ~2s native pipeline (#1068).
+ *
+ * Iterates `missingRel` (every collected file the Rust orchestrator dropped),
+ * not `wasmResults`, so files that produced zero symbols still get a row.
+ */
+function backfillFileHashes(db, missingRel, missingAbs) {
+    try {
+        const upsertHash = db.prepare('INSERT OR REPLACE INTO file_hashes (file, hash, mtime, size) VALUES (?, ?, ?, ?)');
+        const writeHashes = db.transaction(() => {
+            for (let i = 0; i < missingRel.length; i++) {
+                const relPath = missingRel[i];
+                const absPath = missingAbs[i];
+                if (!relPath || !absPath)
+                    continue;
+                let code;
+                try {
+                    code = readFileSafe(absPath);
+                }
+                catch (e) {
+                    debug(`backfillNativeDroppedFiles: read failed for ${relPath}: ${toErrorMessage(e)}`);
+                    continue;
+                }
+                if (code === null)
+                    continue;
+                const stat = fileStat(absPath);
+                const mtime = stat ? stat.mtime : 0;
+                const size = stat ? stat.size : 0;
+                upsertHash.run(relPath, fileHash(code), mtime, size);
+            }
+        });
+        writeHashes();
+    }
+    catch (e) {
+        debug(`backfillNativeDroppedFiles: file_hashes write failed (table may not exist): ${toErrorMessage(e)}`);
+    }
+}
+/**
+ * Backfill files that the native orchestrator silently dropped during parse.
+ * Falls back to WASM + inserts file/symbol nodes so engine counts match (#967).
+ *
+ * Also purges stale rows for WASM-only files deleted from disk (#1073), which
+ * Rust's `detect_removed_files` filter (#1070) skips.
+ *
+ * Accepts a pre-computed `gap` from `detectDroppedLanguageGap` so the caller
+ * can use the same scan for both gating and the actual backfill — avoiding
+ * a redundant fs walk when the orchestrator's signals already triggered.
+ */
+async function backfillNativeDroppedFiles(ctx, gap) {
+    const { missingRel, missingAbs, staleRel } = gap;
+    if (missingAbs.length === 0 && staleRel.length === 0)
+        return;
+    // Now that we know there's work to do, hand off to better-sqlite3 (needed
+    // for the INSERT path below).
+    if (ctx.nativeFirstProxy) {
+        closeNativeDb(ctx, 'pre-parity-backfill');
+        ctx.db = openDb(ctx.dbPath);
+        ctx.nativeFirstProxy = false;
+    }
+    const dbConn = ctx.db;
+    // Purge WASM-only files that were deleted from disk (#1073). Rust's
+    // detect_removed_files skips them and the insert path below never visits
+    // them, so without this their rows would persist across rebuilds until the
+    // next full rebuild reset the DB.
+    if (staleRel.length > 0) {
+        purgeStaleWasmOnlyFiles(dbConn, staleRel);
+    }
+    if (missingAbs.length === 0)
+        return;
+    // Parse all missing files via WASM first so we can distinguish real native
+    // extractor failures (WASM finds symbols but native didn't) from files the
+    // Rust engine legitimately skipped (gitignored artifacts, empty declaration
+    // files, etc. where WASM also produces 0 symbols). Both categories are
+    // backfilled — only the former triggers a WARN (#1566).
+    const wasmResults = await parseFilesWasmForBackfill(missingAbs, ctx.rootDir);
+    // Build two sets from wasmResults:
+    //   wasmParsedFiles  — rel-paths present in wasmResults (WASM succeeded, even 0 symbols)
+    //   wasmFoundSymbols — subset where WASM found ≥1 symbol
+    // Files absent from wasmParsedFiles were skipped by WASM entirely (extension
+    // not in _extToLang, wasmExtractSymbols returned null, or a read error).
+    // Those files do NOT end up in the batchInsertNodes loop below.
+    const wasmParsedFiles = new Set();
+    const wasmFoundSymbols = new Set();
+    for (const [relPath, symbols] of wasmResults) {
+        wasmParsedFiles.add(relPath);
+        if ((symbols.definitions?.length ?? 0) > 0 || (symbols.exports?.length ?? 0) > 0) {
+            wasmFoundSymbols.add(relPath);
+        }
+    }
+    // Classify drops so users see per-extension reasons instead of just a count
+    // (#1011). `unsupported-by-native` is a legitimate parser limit (no Rust
+    // extractor); `native-extractor-failure` indicates a real native bug since
+    // the language IS supported by the addon yet WASM found symbols the native
+    // engine should have extracted. Files where both engines produce 0 symbols
+    // are legitimately empty (e.g. gitignored napi-generated declaration stubs)
+    // and logged at debug level only.
+    classifyAndLogDroppedFiles(missingRel, wasmParsedFiles, wasmFoundSymbols);
+    insertBackfilledNodes(dbConn, wasmResults);
+    backfillFileHashes(dbConn, missingRel, missingAbs);
+    // Free WASM parse trees from the inline backfill path (#1058).
+    // `parseFilesWasmInline` sets `symbols._tree` (a live web-tree-sitter Tree
+    // backed by WASM linear memory) on every result, but these symbols are
+    // consumed locally for DB row construction and never added to
+    // `ctx.allSymbols`, so the finalize-stage `releaseWasmTrees` sweep never
+    // sees them. Without this, trees leak WASM memory until process exit —
+    // bounded per run but cumulative across in-process integration tests.
+    // Mirrors the cleanup discipline established for #931.
+    cleanupThisDispatchWasmTrees(wasmResults);
+}
+/**
+ * Backfill the `technique` column on `calls` edges written by the native Rust
+ * orchestrator, which does not write the column itself.  Also lifts any
+ * resolved ts-native edge whose confidence is below TS_NATIVE_CONFIDENCE_FLOOR
+ * to that floor value so that the name-lookup quality of the native resolver is
+ * reflected in the call-confidence metric.
+ *
+ * For full builds, all `calls` edges in the DB are new so a global UPDATE is
+ * correct.  For incremental builds, only changed-file source nodes are updated
+ * to avoid overwriting previously-set technique values on unchanged edges.
+ */
+function backfillEdgeTechniquesAfterNativeOrchestrator(db, isFullBuild, changedFiles) {
+    // Quiet incremental: no files changed → no new edges inserted, nothing to tag.
+    // Running the global UPDATE here would mis-tag pre-migration NULL-technique edges
+    // from unchanged files as 'ts-native'.
+    if (!isFullBuild && changedFiles && changedFiles.length === 0) {
+        return;
+    }
+    if (isFullBuild || !changedFiles) {
+        db.prepare("UPDATE edges SET technique = 'ts-native' WHERE kind = 'calls' AND technique IS NULL").run();
+        // Lift resolved ts-native edges below the confidence floor.
+        db.prepare(`UPDATE edges SET confidence = ?
+       WHERE kind = 'calls' AND technique = 'ts-native'
+         AND confidence > 0 AND confidence < ?`).run(TS_NATIVE_CONFIDENCE_FLOOR, TS_NATIVE_CONFIDENCE_FLOOR);
+        return;
+    }
+    // Incremental: scope to source nodes whose file is one of the changed files.
+    // Chunk to stay within SQLite's SQLITE_LIMIT_VARIABLE_NUMBER (999 on older builds).
+    const CHUNK_SIZE = 500;
+    const tx = db.transaction(() => {
+        for (let i = 0; i < changedFiles.length; i += CHUNK_SIZE) {
+            const chunk = changedFiles.slice(i, i + CHUNK_SIZE);
+            const placeholders = chunk.map(() => '?').join(',');
+            db.prepare(`UPDATE edges SET technique = 'ts-native'
+         WHERE kind = 'calls' AND technique IS NULL
+         AND source_id IN (
+           SELECT id FROM nodes WHERE file IN (${placeholders})
+         )`).run(...chunk);
+            // Lift resolved ts-native edges below the confidence floor for this chunk.
+            db.prepare(`UPDATE edges SET confidence = ?
+         WHERE kind = 'calls' AND technique = 'ts-native'
+           AND confidence > 0 AND confidence < ?
+           AND source_id IN (
+             SELECT id FROM nodes WHERE file IN (${placeholders})
+           )`).run(TS_NATIVE_CONFIDENCE_FLOOR, TS_NATIVE_CONFIDENCE_FLOOR, ...chunk);
+        }
+    });
+    tx();
+}
+// ── tryNativeOrchestrator helpers ────────────────────────────────────────────
+/**
+ * Open NativeDatabase on demand — deferred from setupPipeline to skip the
+ * ~60ms cost on no-op/early-exit builds.
+ *
+ * Closes the better-sqlite3 connection first to avoid dual-connection WAL
+ * corruption. On setup failure, falls back to reopening better-sqlite3 and
+ * leaves ctx.nativeDb undefined so the caller falls through to the JS pipeline.
+ */
+function openNativeDatabase(ctx) {
+    if (ctx.nativeDb || !ctx.nativeAvailable)
+        return;
+    const native = loadNative();
+    if (!native?.NativeDatabase)
+        return;
+    try {
+        // Close better-sqlite3 before opening rusqlite to avoid WAL conflicts.
+        // Uses raw close() instead of closeDb() intentionally — the advisory lock
+        // is kept and transferred to the NativeDbProxy below, not released here.
+        ctx.db.close();
+        acquireAdvisoryLock(ctx.dbPath);
+        ctx.nativeDb = native.NativeDatabase.openReadWrite(ctx.dbPath);
+        ctx.nativeDb.initSchema();
+        // Replace ctx.db with a NativeDbProxy so post-native JS fallback
+        // (structure, analysis) can use it without reopening better-sqlite3.
+        const proxy = new NativeDbProxy(ctx.nativeDb);
+        proxy.__lockPath = `${ctx.dbPath}.lock`;
+        ctx.db = proxy;
+        ctx.nativeFirstProxy = true;
+    }
+    catch (err) {
+        warn(`NativeDatabase setup failed, falling back to JS: ${toErrorMessage(err)}`);
+        try {
+            ctx.nativeDb?.close();
+        }
+        catch (e) {
+            debug(`tryNativeOrchestrator: close failed during fallback: ${toErrorMessage(e)}`);
+        }
+        ctx.nativeDb = undefined;
+        ctx.nativeFirstProxy = false; // defensive: reset in case future refactors move the assignment above throwing lines
+        releaseAdvisoryLock(`${ctx.dbPath}.lock`);
+        // Reopen better-sqlite3 for JS pipeline fallback
+        ctx.db = openDb(ctx.dbPath);
+    }
+}
+/**
+ * Coordinate all post-native edge-writing post-passes, role re-classification,
+ * and technique backfill. Returns timing data for the build result.
+ *
+ * Post-passes run before structure/analysis so role classification sees the
+ * complete graph including CHA + this/super dispatch edges.
+ */
+async function runPostNativePasses(ctx, result) {
+    // Engine parity: the native orchestrator silently drops files whose
+    // Rust extractor/grammar is missing or fails (e.g. HCL, Scala, Swift on
+    // stale native binaries). WASM handles those — backfill via WASM so both
+    // engines process the same file set (#967).
+    //
+    // Detect the gap once (fs walk + 2 DB queries) and use it for both gating
+    // and the backfill itself. On quiet incrementals we still pay the walk so
+    // we can detect brand-new files in dropped-language extensions — a gap that
+    // the orchestrator's `detect_removed_files` filter (#1070) leaves open
+    // (#1083, #1091). The pre-check is cheap because the expensive part (WASM
+    // re-parse of the missing set) is gated below.
+    const gapDetectStart = performance.now();
+    const gap = detectDroppedLanguageGap(ctx);
+    const backfillHappened = gap.missingAbs.length > 0 || gap.staleRel.length > 0;
+    if (backfillHappened) {
+        await backfillNativeDroppedFiles(ctx, gap);
+    }
+    const gapDetectMs = performance.now() - gapDetectStart;
+    // Phase 8.5: this/super dispatch — hybrid WASM re-parse to resolve call sites
+    // whose raw receiver info the Rust pipeline does not persist to DB.
+    // Runs BEFORE the CHA expansion pass so that super.method() → Parent.method edges
+    // (technique='cha') are in the DB when runPostNativeCha expands them to sibling
+    // class overrides (e.g. PostMixin.m → B.m when PostMixin and B both extend A).
+    const { elapsedMs: thisDispatchMs, targetIds: thisDispatchTargetIds, affectedFiles: thisDispatchAffectedFiles, } = await runPostNativeThisDispatch(ctx.db, ctx.rootDir, result.changedFiles, !!result.isFullBuild);
+    // Phase 8.6: expand CHA call edges (interface dispatch → concrete implementations).
+    // Returns the affected files so role re-classification below can be scoped to
+    // the nodes whose fan-in/out actually changed.
+    //
+    // Runs AFTER this/super dispatch so super.method() edges are already in the DB.
+    // The 'cha-expanded' technique tag on this pass's own output prevents re-expansion
+    // of those edges in subsequent incremental builds, while 'cha'-tagged edges from
+    // this/super dispatch remain eligible for expansion here.
+    //
+    // Function-as-object-property methods (`fn.method = function() {}`) are extracted
+    // natively by the Rust engine (#1432) and resolved in-build by its edge builder, so
+    // no WASM re-parse post-pass is needed for them. `Foo.prototype.bar = fn` likewise.
+    const chaStart = performance.now();
+    const { newEdgeCount: chaEdgeCount, affectedFiles: chaAffectedFiles } = runPostNativeCha(ctx.db, 
+    // null = full build (scan all call→method edges); array = incremental (gate queries decide scope)
+    result.isFullBuild ? null : (result.changedFiles ?? null));
+    const chaMs = performance.now() - chaStart;
+    // Role re-classification after JS edge-writing post-passes.
+    // The Rust orchestrator classifies roles before these post-passes (CHA,
+    // this-dispatch) add edges, so roles for the edge endpoints are stale.
+    // Scoped to the files containing those endpoints: a new edge only changes
+    // fan-in/out for its own source and target nodes, so re-classifying their
+    // files restores correctness without re-running the classifier over the
+    // whole graph (which cost ~130ms per build on codegraph itself and was a
+    // major part of the v3.12.0 native full-build benchmark regression).
+    let reclassifyMs = 0;
+    if (chaEdgeCount > 0 || thisDispatchTargetIds.size > 0) {
+        const affectedFiles = [...new Set([...chaAffectedFiles, ...thisDispatchAffectedFiles])];
+        // When edges were inserted but all their endpoint nodes have null `file`
+        // columns (rare but possible), affectedFiles stays empty even though
+        // fan-in/out changed. Fall back to full-graph re-classification in that
+        // case — scoped classification with an empty set would be a no-op, leaving
+        // roles stale for those nodes.
+        const scopedFiles = affectedFiles.length > 0 ? affectedFiles : null;
+        const reclassifyStart = performance.now();
+        try {
+            const { classifyNodeRoles } = (await import('../../../../features/structure.js'));
+            classifyNodeRoles(ctx.db, scopedFiles);
+            debug(scopedFiles
+                ? `Post-pass role re-classification complete (${scopedFiles.length} file(s))`
+                : 'Post-pass role re-classification complete (full graph — null-file endpoints)');
+        }
+        catch (err) {
+            debug(`Post-pass role re-classification failed: ${toErrorMessage(err)}`);
+        }
+        reclassifyMs = performance.now() - reclassifyStart;
+    }
+    // Backfill the `technique` column on `calls` edges written by the Rust
+    // orchestrator, which does not write the column. Runs after all edge-writing
+    // phases (including the WASM dropped-language backfill, CHA post-pass, and
+    // this/super dispatch) so every new edge in this build cycle gets a label.
+    const techniqueBackfillStart = performance.now();
+    backfillEdgeTechniquesAfterNativeOrchestrator(ctx.db, !!result.isFullBuild, result.changedFiles);
+    const techniqueBackfillMs = performance.now() - techniqueBackfillStart;
+    // Re-count nodes/edges now that all edge-writing post-passes have run: the
+    // Rust orchestrator captured its counts before the JS post-passes added
+    // edges, so both its summary and build_meta under-report (#1452).
+    //
+    // Fast path: skip the COUNT(*) scan when no post-pass wrote any edges.
+    // COUNT(*) on large tables (50K+ edges) is non-trivial, especially via the
+    // NativeDbProxy napi-rs round-trip. When all post-passes were no-ops, the
+    // Rust orchestrator's counts are still accurate — no re-count needed.
+    let finalNodeCount = result.nodeCount ?? 0;
+    let finalEdgeCount = result.edgeCount ?? 0;
+    const postPassWroteData = backfillHappened || chaEdgeCount > 0 || thisDispatchTargetIds.size > 0;
+    if (postPassWroteData) {
+        try {
+            const counts = ctx.db
+                .prepare('SELECT (SELECT COUNT(*) FROM nodes) AS n, (SELECT COUNT(*) FROM edges) AS e')
+                .get();
+            if (counts.n !== finalNodeCount || counts.e !== finalEdgeCount) {
+                finalNodeCount = counts.n;
+                finalEdgeCount = counts.e;
+                setBuildMeta(ctx.db, { node_count: finalNodeCount, edge_count: finalEdgeCount });
+            }
+        }
+        catch (err) {
+            debug(`Post-pass node/edge re-count failed: ${toErrorMessage(err)}`);
+        }
+    }
+    info(`Native build orchestrator completed: ${finalNodeCount} nodes, ${finalEdgeCount} edges, ${result.fileCount ?? 0} files`);
+    return {
+        gapDetectMs,
+        chaMs,
+        thisDispatchMs,
+        reclassifyMs,
+        techniqueBackfillMs,
+        backfillHappened,
+    };
+}
+/**
+ * Try the native build orchestrator.
+ *
+ * Returns:
+ *   - `BuildResult` on success (caller should return it directly).
+ *   - `'early-exit'` when the orchestrator detected no changes (caller should return undefined).
+ *   - `undefined` when native is unavailable or skipped (caller should fall through to the JS pipeline).
+ *
+ * Encapsulates the orchestrator-selection strategy: open `NativeDatabase`,
+ * invoke `nativeDb.buildGraph()` (the Rust pipeline), and run post-native
+ * structure + analysis fallbacks. Lives in its own file to keep the Rust
+ * orchestrator entry point separated from the JS-side `buildGraph()` driver
+ * in `pipeline.ts`.
+ */
+export async function tryNativeOrchestrator(ctx) {
+    const skipReason = shouldSkipNativeOrchestrator(ctx);
+    if (skipReason) {
+        debug(`Skipping native orchestrator: ${skipReason}`);
+        return undefined;
+    }
+    openNativeDatabase(ctx);
+    if (!ctx.nativeDb?.buildGraph)
+        return undefined;
+    const resultJson = ctx.nativeDb.buildGraph(ctx.rootDir, JSON.stringify(ctx.config), JSON.stringify(ctx.aliases), JSON.stringify(ctx.opts));
+    const result = JSON.parse(resultJson);
+    if (result.earlyExit) {
+        info('No changes detected');
+        // Even on no-op rebuilds, dropped-language files added since the last
+        // full build are still missing from `nodes`/`file_hashes` (#1083), and
+        // WASM-only files deleted from disk leave stale rows behind (#1073).
+        // The orchestrator's collect_files skipped them, so its earlyExit
+        // doesn't imply DB consistency. Run the gap repair before returning.
+        const gap = detectDroppedLanguageGap(ctx);
+        if (gap.missingAbs.length > 0 || gap.staleRel.length > 0) {
+            await backfillNativeDroppedFiles(ctx, gap);
+        }
+        closeDbPair({ db: ctx.db, nativeDb: ctx.nativeDb });
+        return 'early-exit';
+    }
+    // Log incremental status to match JS pipeline output
+    const changed = result.changedCount ?? 0;
+    const removed = result.removedCount ?? 0;
+    if (!result.isFullBuild && (changed > 0 || removed > 0)) {
+        info(`Incremental: ${changed} changed, ${removed} removed`);
+    }
+    const p = result.phases;
+    // Sync build_meta so JS-side version/engine checks work on next build.
+    // Use the binary's CARGO_PKG_VERSION (ctx.nativeBinaryVersion), not the
+    // platform package.json version (ctx.engineVersion). The Rust side's
+    // check_version_mismatch compares against CARGO_PKG_VERSION; writing
+    // the package.json value would create a permanent mismatch whenever
+    // the binary and platform package.json diverge — e.g., CI hot-swap
+    // via ci-install-native.mjs (#1066) — forcing every subsequent build
+    // to be a full rebuild.
+    //
+    // When the native addon doesn't expose engineVersion() (older addon),
+    // fall back to CODEGRAPH_VERSION — same fallback used by both
+    // checkEngineSchemaMismatch (read path) and persistBuildMetadata
+    // (the JS-pipeline write path in finalize.ts). Using ctx.engineVersion
+    // here would re-introduce the asymmetry this PR fixes for that case.
+    const nativeVersionForMeta = ctx.nativeBinaryVersion || CODEGRAPH_VERSION;
+    setBuildMeta(ctx.db, {
+        engine: ctx.engineName,
+        engine_version: nativeVersionForMeta,
+        codegraph_version: nativeVersionForMeta,
+        schema_version: String(ctx.schemaVersion),
+        built_at: new Date().toISOString(),
+    });
+    // The build summary is logged after the JS edge-writing post-passes below
+    // (dropped-language backfill, CHA, this/super dispatch) so the reported
+    // counts include their edges (#1452).
+    // ── Post-native structure + analysis ──────────────────────────────
+    let analysisTiming = {
+        astMs: +(p.astMs ?? 0),
+        complexityMs: +(p.complexityMs ?? 0),
+        cfgMs: +(p.cfgMs ?? 0),
+        dataflowMs: +(p.dataflowMs ?? 0),
+    };
+    let structurePatchMs = 0;
+    // Skip JS structure when the Rust pipeline's small-incremental fast path
+    // already handled it. For full builds and large incrementals where Rust
+    // skipped structure, we must run the JS fallback.
+    const needsStructure = !result.structureHandled;
+    // When the Rust addon doesn't include analysis persistence (older addon
+    // version or analysis failed), fall back to JS-side analysis.
+    const needsAnalysisFallback = !result.analysisComplete &&
+        (ctx.opts.ast !== false ||
+            ctx.opts.complexity !== false ||
+            ctx.opts.cfg !== false ||
+            ctx.opts.dataflow !== false);
+    // ── DB handoff ────────────────────────────────────────────────────────────
+    // Ensure a proper better-sqlite3 connection is open before any post-pass that
+    // writes edges (dropped-language backfill, CHA) and before structure/analysis.
+    // When analysis fallback is needed the handoff already happened above; when
+    // neither structure nor analysis is needed the proxy conversion is deferred to
+    // here so CHA and technique-backfill can still write rows.
+    if (needsStructure || needsAnalysisFallback) {
+        if (needsAnalysisFallback && ctx.nativeFirstProxy) {
+            closeNativeDb(ctx, 'pre-analysis-fallback');
+            ctx.db = openDb(ctx.dbPath);
+            ctx.nativeFirstProxy = false;
+        }
+        else if (!ctx.nativeFirstProxy && !handoffWalAfterNativeBuild(ctx)) {
+            // DB reopen failed — return partial result (no post-pass phases completed)
+            return formatNativeTimingResult(p, 0, analysisTiming, {
+                gapDetectMs: 0,
+                chaMs: 0,
+                thisDispatchMs: 0,
+                reclassifyMs: 0,
+                techniqueBackfillMs: 0,
+            });
+        }
+    }
+    const postPassTimings = await runPostNativePasses(ctx, result);
+    // ── Structure and analysis fallback (run after edge-writing so roles see full graph) ──
+    // Reconstruct fileSymbols once for both structure and analysis to avoid two
+    // expensive DB scans. The DB handoff above already ensured ctx.db is a proper
+    // better-sqlite3 connection when either flag is set.
+    if (needsStructure || needsAnalysisFallback) {
+        const fileSymbols = reconstructFileSymbolsFromDb(ctx);
+        if (needsStructure) {
+            structurePatchMs = await runPostNativeStructure(ctx, fileSymbols, !!result.isFullBuild, result.changedFiles);
+        }
+        if (needsAnalysisFallback) {
+            analysisTiming = await runPostNativeAnalysis(ctx, fileSymbols, result.changedFiles);
+        }
+    }
+    // P6: Vertex extraction for the analysisComplete=true path.
+    // When needsAnalysisFallback=false (the normal native case), runPostNativeAnalysis
+    // was skipped, so buildDataflowEdges never ran and dataflow_vertices were never
+    // populated. Re-run the Rust dataflow visitor per file (fast — no re-parse) to
+    // get the DataflowResult, then build vertices and inter-procedural edges.
+    // Languages where Rust has no dataflow rules are silently skipped; a WASM
+    // fallback for those is tracked in issue #1614.
+    if (ctx.opts.dataflow !== false && !needsAnalysisFallback) {
+        await runDataflowVertexPass(ctx, result.changedFiles);
+    }
+    closeDbPair({ db: ctx.db, nativeDb: ctx.nativeDb });
+    return formatNativeTimingResult(p, structurePatchMs, analysisTiming, postPassTimings);
+}
+//# sourceMappingURL=native-orchestrator.js.map