perf(resolution): batch-loop de-quadratic — keyset reads, changes-based guard, DB-scaled valve caps + resolve profiler (#1339)

The §7a.2 per-ref profile overturned the assumption the whole arc was
built on: resolveOne owns only ~93s of the kernel-scale ~433s batch loop.
Loop-stage attribution (CODEGRAPH_RESOLVE_PROFILE, shipped here) named the
rest: backpressure folds 111.2s, count guard 93.9s, batch reads 54.6s,
deletes/inserts/marks ~84s, settle 85.7s.

- Non-progress guard O(remaining)→O(1): the per-batch COUNT(*) walked every
  remaining pending row (O(N²/batch) per run, 93.9s). The cleanup queries
  now return summed SQLite , and zero-removals-from-claimed-work
  is the guard signal — the DIRECT evidence the count diff inferred (a
  mismatched-name resolver makes keyed cleanup no-op ⇒ changes=0). A real
  COUNT runs only on that suspicious path and arbitrates exactly as before.
- Batch reads OFFSET→keyset (54.6s→O(batch)): OFFSET re-walked the
  accumulated failed-row prefix every read; seeking past the last-seen
  rowid is prefix-independent and enumeration-order identical.
- WAL valve caps scale with DB size (env still wins): every fold re-writes
  hot pages (#1231 in bounded form — 111.2s at the flat 256MB cap);
  soft=clamp(dbSize/4, 256MB, 2GB) trades ~4× fewer folds for a transient
  WAL ≈ project size.
- CODEGRAPH_RESOLVE_PROFILE: per-outcome resolveOne histogram + loop-stage
  attribution, main + workers, off by default.

Gates: dubbo dump byte-identical; suite 2,491 passed / 4 skipped (kernel
required). Kernel-scale payoff run lands in the plan doc next.

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

__tests__/wal-deferral.test.ts

@@ -416,3 +416,14 @@ describe('valve file-size trigger (§7a.1: backfilled WAL still grows the file)'
     db.close();
   });
 });
+
+describe('resolveWalValveMb DB-size scaling (§7a.2 fold-tax reduction)', () => {
+  it('scales soft cap ~dbSize/4 within [256, 2048]MB; env always wins', () => {
+    const GB = 1024 * 1024 * 1024;
+    expect(resolveWalValveMb(undefined, 100 * 1024 * 1024)).toBe(256); // floor
+    expect(resolveWalValveMb(undefined, 4.6 * GB)).toBe(1177); // ~dbSize/4
+    expect(resolveWalValveMb(undefined, 40 * GB)).toBe(2048); // ceiling
+    expect(resolveWalValveMb('64', 40 * GB)).toBe(64); // env override wins
+    expect(resolveWalValveMb(undefined, 0)).toBe(256); // unknown size → default
+  });
+});

src/db/index.ts

@@ -325,6 +325,17 @@ export class DatabaseConnection {
     }
   }
 
+  /** Size of the main DB file in bytes (0 for in-memory/unknown) — the WAL
+   * valve scales its fold caps with it (resolveWalValveMb). */
+  getDbFileSizeBytes(): number {
+    if (!this.dbPath || this.dbPath === ':memory:') return 0;
+    try {
+      return fs.statSync(this.dbPath).size;
+    } catch {
+      return 0;
+    }
+  }
+
   /** Current `wal_autocheckpoint` interval in pages (0 = disabled). */
   getWalAutocheckpoint(): number {
     const v = this.db.pragma('wal_autocheckpoint', { simple: true });

src/db/queries.ts

@@ -230,6 +230,7 @@ export class QueryBuilder {
     getNodesByLowerName?: SqliteStatement;
     getUnresolvedCount?: SqliteStatement;
     getUnresolvedBatch?: SqliteStatement;
+    getUnresolvedBatchAfter?: SqliteStatement;
     deleteRefsByRowIdsFull?: SqliteStatement;
     getAllFilePaths?: SqliteStatement;
     getAllNodeNames?: SqliteStatement;
@@ -2085,6 +2086,34 @@ export class QueryBuilder {
     }));
   }
 
+  /**
+   * Keyset variant of {@link getUnresolvedReferencesBatch} for the batched
+   * resolution loop: seek past the last-seen row id instead of OFFSET-walking.
+   * OFFSET reads re-scan the accumulated failed-row prefix on every batch —
+   * O(failed rows) per read, measured at 54.6s of the kernel-scale batch loop
+   * (§7a.2) — while the seek is O(batch) forever. `id` is the rowid alias, so
+   * the enumeration order is identical to the OFFSET reader's.
+   */
+  getUnresolvedReferencesBatchAfter(afterRowId: number, limit: number): UnresolvedReference[] {
+    if (!this.stmts.getUnresolvedBatchAfter) {
+      this.stmts.getUnresolvedBatchAfter = this.db.prepare(
+        "SELECT * FROM unresolved_refs WHERE status = 'pending' AND id > ? ORDER BY id LIMIT ?"
+      );
+    }
+    const rows = this.stmts.getUnresolvedBatchAfter.all(afterRowId, limit) as UnresolvedRefRow[];
+    return rows.map((row) => ({
+      fromNodeId: row.from_node_id,
+      referenceName: row.reference_name,
+      referenceKind: row.reference_kind as EdgeKind,
+      line: row.line,
+      column: row.col,
+      candidates: row.candidates ? safeJsonParse(row.candidates, undefined) : undefined,
+      filePath: row.file_path,
+      language: row.language as Language,
+      rowId: row.id,
+    }));
+  }
+
   /**
    * Get all tracked file paths (lightweight — no full FileRecord objects)
    */
@@ -2182,17 +2211,22 @@ export class QueryBuilder {
    * Delete specific resolved references by (fromNodeId, referenceName, referenceKind) tuples.
    * More precise than deleteResolvedReferences — only removes refs that were actually resolved.
    */
-  deleteSpecificResolvedReferences(refs: Array<{ fromNodeId: string; referenceName: string; referenceKind: string }>): void {
-    if (refs.length === 0) return;
+  deleteSpecificResolvedReferences(refs: Array<{ fromNodeId: string; referenceName: string; referenceKind: string }>): number {
+    if (refs.length === 0) return 0;
     const stmt = this.db.prepare(
       'DELETE FROM unresolved_refs WHERE from_node_id = ? AND reference_name = ? AND reference_kind = ?'
     );
+    // Returns rows actually removed (SQLite `changes`, summed): the batched
+    // resolution loop's non-progress guard keys on this — zero removals from
+    // a batch that claimed work is the direct runaway signal (§7a.2).
+    let changed = 0;
     const deleteMany = this.db.transaction((items: typeof refs) => {
       for (const ref of items) {
-        stmt.run(ref.fromNodeId, ref.referenceName, ref.referenceKind);
+        changed += stmt.run(ref.fromNodeId, ref.referenceName, ref.referenceKind).changes;
       }
     });
     deleteMany(refs);
+    return changed;
   }
 
   /**
@@ -2203,13 +2237,15 @@ export class QueryBuilder {
    * caller's same-named call sites, the later sites' edges were silently never
    * created (#1269).
    */
-  deleteReferencesByRowIds(rowIds: number[]): void {
-    if (rowIds.length === 0) return;
+  deleteReferencesByRowIds(rowIds: number[]): number {
+    if (rowIds.length === 0) return 0;
     // One transaction for all chunks (each chunk was previously its own
     // implicit transaction = its own WAL commit — measurable on 100k+-ref
     // resolution persists), and the full-size chunk statement is cached so
     // repeat calls skip the re-prepare; only the final partial chunk (if any)
-    // prepares ad hoc.
+    // prepares ad hoc. Returns rows actually removed (summed `changes`) for
+    // the batched loop's non-progress guard (§7a.2).
+    let changed = 0;
     this.db.transaction(() => {
       for (let i = 0; i < rowIds.length; i += SQLITE_PARAM_CHUNK_SIZE) {
         const chunk = rowIds.slice(i, i + SQLITE_PARAM_CHUNK_SIZE);
@@ -2220,13 +2256,14 @@ export class QueryBuilder {
               `DELETE FROM unresolved_refs WHERE id IN (${placeholders})`
             );
           }
-          this.stmts.deleteRefsByRowIdsFull.run(...chunk);
+          changed += this.stmts.deleteRefsByRowIdsFull.run(...chunk).changes;
         } else {
           const placeholders = chunk.map(() => '?').join(',');
-          this.db.prepare(`DELETE FROM unresolved_refs WHERE id IN (${placeholders})`).run(...chunk);
+          changed += this.db.prepare(`DELETE FROM unresolved_refs WHERE id IN (${placeholders})`).run(...chunk).changes;
         }
       }
     })();
+    return changed;
   }
 
   /**
@@ -2238,17 +2275,19 @@ export class QueryBuilder {
    * is (re)written here so rows inserted before the v8 migration get their
    * tail the first time they're attempted.
    */
-  markReferencesFailed(refs: Array<{ fromNodeId: string; referenceName: string; referenceKind: string }>): void {
-    if (refs.length === 0) return;
+  markReferencesFailed(refs: Array<{ fromNodeId: string; referenceName: string; referenceKind: string }>): number {
+    if (refs.length === 0) return 0;
     const stmt = this.db.prepare(
       "UPDATE unresolved_refs SET status = 'failed', name_tail = ? WHERE from_node_id = ? AND reference_name = ? AND reference_kind = ?"
     );
+    let changed = 0;
     const markMany = this.db.transaction((items: typeof refs) => {
       for (const ref of items) {
-        stmt.run(referenceNameTail(ref.referenceName), ref.fromNodeId, ref.referenceName, ref.referenceKind);
+        changed += stmt.run(referenceNameTail(ref.referenceName), ref.fromNodeId, ref.referenceName, ref.referenceKind).changes;
       }
     });
     markMany(refs);
+    return changed;
   }
 
   /**
@@ -2259,17 +2298,19 @@ export class QueryBuilder {
    * can differ per call site (receiver-type inference reads the ref's line),
    * so a sibling must not inherit this row's failure.
    */
-  markReferencesFailedByRowIds(refs: Array<{ rowId: number; referenceName: string }>): void {
-    if (refs.length === 0) return;
+  markReferencesFailedByRowIds(refs: Array<{ rowId: number; referenceName: string }>): number {
+    if (refs.length === 0) return 0;
     const stmt = this.db.prepare(
       "UPDATE unresolved_refs SET status = 'failed', name_tail = ? WHERE id = ?"
     );
+    let changed = 0;
     const markMany = this.db.transaction((items: typeof refs) => {
       for (const ref of items) {
-        stmt.run(referenceNameTail(ref.referenceName), ref.rowId);
+        changed += stmt.run(referenceNameTail(ref.referenceName), ref.rowId).changes;
       }
     });
     markMany(refs);
+    return changed;
   }
 
   /**

src/db/wal-valve.ts

@@ -61,11 +61,19 @@ const CHECK_INTERVAL_MS = 2000;
  * Resolve the valve's soft threshold from the `CODEGRAPH_WAL_VALVE_MB`
  * override; non-numeric / non-positive values fall back to the default.
  */
-export function resolveWalValveMb(envVal: string | undefined): number {
+export function resolveWalValveMb(envVal: string | undefined, dbSizeBytes?: number): number {
   if (envVal !== undefined && envVal !== '') {
     const n = Number(envVal);
     if (Number.isFinite(n) && n > 0) return Math.floor(n);
   }
+  // Scale with the project when the caller knows the DB size: every fold
+  // re-writes hot B-tree pages into the main file (the #1231 pathology in
+  // bounded form — 111s of a kernel-scale batch loop at the flat 256MB cap,
+  // §7a.2), so a big project affords a proportionally bigger transient WAL
+  // (~dbSize/4 soft ⇒ file cap ≈ dbSize) in exchange for ~4× fewer folds.
+  if (dbSizeBytes !== undefined && dbSizeBytes > 0) {
+    return Math.min(2048, Math.max(DEFAULT_WAL_VALVE_MB, Math.floor(dbSizeBytes / 4 / (1024 * 1024))));
+  }
   return DEFAULT_WAL_VALVE_MB;
 }
 

src/index.ts

@@ -25,7 +25,7 @@ import {
   FindRelevantContextOptions,
 } from './types';
 import { DatabaseConnection, getDatabasePath, removeDatabaseFiles } from './db';
-import { WalCheckpointValve } from './db/wal-valve';
+import { WalCheckpointValve, resolveWalValveMb } from './db/wal-valve';
 import { QueryBuilder } from './db/queries';
 import {
   isInitialized,
@@ -472,7 +472,7 @@ export class CodeGraph {
         this.db.setWalAutocheckpoint(0);
         walValve = new WalCheckpointValve(
           this.db,
-          undefined,
+          resolveWalValveMb(process.env.CODEGRAPH_WAL_VALVE_MB, this.db.getDbFileSizeBytes()),
           undefined,
           options.verbose ? (m) => console.log(`[wal-valve] ${m}`) : undefined
         );
@@ -753,7 +753,7 @@ export class CodeGraph {
         this.db.setWalAutocheckpoint(0);
         walValve = new WalCheckpointValve(
           this.db,
-          undefined,
+          resolveWalValveMb(process.env.CODEGRAPH_WAL_VALVE_MB, this.db.getDbFileSizeBytes()),
           undefined,
           options.verbose ? (m) => console.log(`[wal-valve] ${m}`) : undefined
         );

src/resolution/index.ts

@@ -637,7 +637,7 @@ export class ReferenceResolver {
 
     for (let i = 0; i < refs.length; i++) {
       const ref = refs[i]!; // Array index is guaranteed to be in bounds
-      const result = this.resolveOne(ref);
+      const result = this.resolveOneTimed(ref);
 
       if (result) {
         resolved.push(result);
@@ -1233,7 +1233,7 @@ export class ReferenceResolver {
         language: raw.language || this.getLanguageFromNodeId(raw.fromNodeId),
         rowId: raw.rowId,
       };
-      const result = this.resolveOne(ref);
+      const result = this.resolveOneTimed(ref);
       if (result) {
         resolved.push(result);
         byMethod[result.resolvedBy] = (byMethod[result.resolvedBy] || 0) + 1;
@@ -1266,6 +1266,47 @@ export class ReferenceResolver {
    * of resolveBatchYielding, minus the main-thread yields (worker threads have
    * no watchdog heartbeat to starve). Results are in input order.
    */
+  /**
+   * CODEGRAPH_RESOLVE_PROFILE=1: per-outcome wall-clock histogram of
+   * resolveOne, keyed by the winning strategy (`resolvedBy`) or
+   * `fail:<referenceKind>` — the §7a.2 "profile the per-ref path" probe. The
+   * kernel-scale batch loop is ~430s and CORE-INVARIANT (835.9s pooled-4-on-8
+   * ≈ 812.5s sequential-on-2 for the whole superphase), so the next lever is
+   * which CLASS of ref the time belongs to, not more parallelism. Off by
+   * default: the hrtime pair costs ~100ns/ref only when the env is set.
+   */
+  private resolveProfile: Map<string, { n: number; ns: bigint }> | null =
+    process.env.CODEGRAPH_RESOLVE_PROFILE ? new Map() : null;
+
+  private resolveOneTimed(ref: UnresolvedRef): ResolvedRef | null {
+    if (!this.resolveProfile) return this.resolveOne(ref);
+    const t0 = process.hrtime.bigint();
+    const result = this.resolveOne(ref);
+    const dt = process.hrtime.bigint() - t0;
+    const key = result ? result.resolvedBy : `fail:${ref.referenceKind}`;
+    const slot = this.resolveProfile.get(key);
+    if (slot) {
+      slot.n++;
+      slot.ns += dt;
+    } else {
+      this.resolveProfile.set(key, { n: 1, ns: dt });
+    }
+    return result;
+  }
+
+  /** Dump the CODEGRAPH_RESOLVE_PROFILE histogram to stderr (no-op when off). */
+  dumpResolveProfile(label: string): void {
+    if (!this.resolveProfile || this.resolveProfile.size === 0) return;
+    const rows = [...this.resolveProfile.entries()]
+      .map(([k, v]) => ({ k, n: v.n, ms: Number(v.ns / 1_000_000n) }))
+      .sort((a, b) => b.ms - a.ms);
+    for (const r of rows) {
+      console.error(
+        `[resolve-profile] ${label} ${r.k}: n=${r.n} total=${(r.ms / 1000).toFixed(1)}s avg=${((r.ms * 1000) / Math.max(1, r.n)).toFixed(0)}µs`
+      );
+    }
+  }
+
   resolveListForAdmission(refs: UnresolvedReference[]): {
     resolved: ResolvedRef[];
     unresolved: UnresolvedRef[];
@@ -1288,7 +1329,7 @@ export class ReferenceResolver {
         language: raw.language || this.getLanguageFromNodeId(raw.fromNodeId),
         rowId: raw.rowId,
       };
-      const result = this.resolveOne(ref);
+      const result = this.resolveOneTimed(ref);
       if (result) {
         resolved.push(result);
         byMethod[result.resolvedBy] = (byMethod[result.resolvedBy] || 0) + 1;
@@ -1362,6 +1403,16 @@ export class ReferenceResolver {
       console.error(`[pool-timing] backpressure hook: ${parallel?.backpressure ? 'present' : 'absent'}`);
     }
 
+    // CODEGRAPH_RESOLVE_PROFILE loop-stage attribution: the §7a.2 kernel-scale
+    // histogram showed resolveOne owns only ~93s of the ~436s batch loop —
+    // these counters name where the other ~340s goes (reads, edge build+insert,
+    // deletes/marks, the per-batch count guard).
+    const loopProf: Record<string, number> | null = process.env.CODEGRAPH_RESOLVE_PROFILE
+      ? { read: 0, settle: 0, backpressure: 0, createEdges: 0, insertEdges: 0, deletes: 0, marks: 0, countGuard: 0 }
+      : null;
+    const lp = (k: string, t0: number): void => { if (loopProf) loopProf[k] = (loopProf[k] ?? 0) + (Date.now() - t0); };
+    let tLp = 0;
+
     await this.warmCachesYielding(maybeYield);
 
     const total = this.queries.getUnresolvedReferencesCount();
@@ -1476,18 +1527,24 @@ export class ReferenceResolver {
 
     try {
     try {
-    let batch = this.queries.getUnresolvedReferencesBatch(0, batchSize);
+    tLp = Date.now();
+    let batch = this.queries.getUnresolvedReferencesBatchAfter(0, batchSize);
+    lp('read', tLp);
     let inFlight: InFlight | null = batch.length > 0 ? beginBatch(batch) : null;
     while (batch.length > 0 && inFlight) {
       // Prefetch the NEXT batch before this one persists: this batch's rows
       // are still pending (nothing has mutated the table since they were
-      // read), so skipping exactly batch.length rows in the same rowid
-      // enumeration yields the following batch.
-      const nextBatch = this.queries.getUnresolvedReferencesBatch(batch.length, batchSize);
+      // read), so seeking past this batch's last row id in the same rowid
+      // enumeration yields the following batch (keyset — OFFSET re-walked the
+      // accumulated failed prefix every read, 54.6s at kernel scale, §7a.2).
+      tLp = Date.now();
+      const nextBatch = this.queries.getUnresolvedReferencesBatchAfter(batch[batch.length - 1]!.rowId!, batchSize);
+      lp('read', tLp);
 
       const tBatch = Date.now();
       const result = await settleBatch(inFlight, batch);
       if (process.env.CODEGRAPH_SYNTH_TIMINGS) console.error(`[pool-timing] batch ${inFlight.mode}: ${batch.length} refs in ${Date.now() - tBatch}ms`);
+      lp('settle', tBatch);
 
       // WAL-valve backstop at the ONE pool-idle boundary of the double-buffer
       // (this batch settled, the next not yet fanned out): past the hard cap
@@ -1495,8 +1552,10 @@ export class ReferenceResolver {
       // are all between statements — so the backfill completes, readers
       // re-enter at SQLite's backfilled mark, and the next persist commit
       // WRAPS the WAL instead of growing it. No-op (one fstat) under the cap.
+      tLp = Date.now();
       const bp = parallel?.backpressure?.();
       if (bp) await bp;
+      lp('backpressure', tLp);
 
       // Persist in bounded sub-transactions with yields between: a whole
       // batch's edge insert / keyed deletes are otherwise one solid
@@ -1515,11 +1574,15 @@ export class ReferenceResolver {
       // base class if those edges are visible. (Validated on dubbo: fanning
       // out first downgraded exactly those supertype-method resolutions from
       // the 0.9 typed-receiver path to the 0.65 word-overlap fallback.)
+      tLp = Date.now();
       const edges = this.createEdges(result.resolved);
+      lp('createEdges', tLp);
+      tLp = Date.now();
       for (let i = 0; i < edges.length; i += PERSIST_CHUNK) {
         this.queries.insertEdges(edges.slice(i, i + PERSIST_CHUNK));
         await maybeYield();
       }
+      lp('insertEdges', tLp);
 
       // NOW fan the next batch out — workers see exactly the edge state the
       // sequential baseline would (every batch ≤ this one committed), while
@@ -1531,29 +1594,34 @@ export class ReferenceResolver {
       // by row id, so a same-key sibling ref in a LATER batch (same caller
       // calling the same callee at another line) is left pending for its own
       // attempt instead of being swept out with this batch's rows (#1269).
+      tLp = Date.now();
+      let removedThisBatch = 0;
       const resolvedCleanup = ReferenceResolver.partitionResolvedCleanup(result.resolved);
       for (let i = 0; i < resolvedCleanup.rowIds.length; i += PERSIST_CHUNK) {
-        this.queries.deleteReferencesByRowIds(resolvedCleanup.rowIds.slice(i, i + PERSIST_CHUNK));
+        removedThisBatch += this.queries.deleteReferencesByRowIds(resolvedCleanup.rowIds.slice(i, i + PERSIST_CHUNK));
         await maybeYield();
       }
       for (let i = 0; i < resolvedCleanup.legacyKeys.length; i += PERSIST_CHUNK) {
-        this.queries.deleteSpecificResolvedReferences(resolvedCleanup.legacyKeys.slice(i, i + PERSIST_CHUNK));
+        removedThisBatch += this.queries.deleteSpecificResolvedReferences(resolvedCleanup.legacyKeys.slice(i, i + PERSIST_CHUNK));
         await maybeYield();
       }
+      lp('deletes', tLp);
 
       // Park unresolvable refs from this batch as status='failed' so they
       // leave the pending set (the batch reader and non-progress guard below
       // only see pending rows) but stay retryable when a later sync adds a
       // symbol that could satisfy them (#1240).
+      tLp = Date.now();
       const failedCleanup = ReferenceResolver.partitionFailedCleanup(result.unresolved);
       for (let i = 0; i < failedCleanup.byRowId.length; i += PERSIST_CHUNK) {
-        this.queries.markReferencesFailedByRowIds(failedCleanup.byRowId.slice(i, i + PERSIST_CHUNK));
+        removedThisBatch += this.queries.markReferencesFailedByRowIds(failedCleanup.byRowId.slice(i, i + PERSIST_CHUNK));
         await maybeYield();
       }
       for (let i = 0; i < failedCleanup.legacyKeys.length; i += PERSIST_CHUNK) {
-        this.queries.markReferencesFailed(failedCleanup.legacyKeys.slice(i, i + PERSIST_CHUNK));
+        removedThisBatch += this.queries.markReferencesFailed(failedCleanup.legacyKeys.slice(i, i + PERSIST_CHUNK));
         await maybeYield();
       }
+      lp('marks', tLp);
 
       if (process.env.CODEGRAPH_SYNTH_TIMINGS) console.error(`[pool-timing] batch persist: ${Date.now() - tPersist}ms`);
 
@@ -1590,9 +1658,22 @@ export class ReferenceResolver {
       // 1.4 GB before the Go-fallback fix). Stop rather than grow the graph
       // without bound. (An in-flight prefetched batch is abandoned unsettled —
       // fan-out has no side effects until settleBatch appends its results.)
-      const remaining = this.queries.getUnresolvedReferencesCount();
-      if (remaining >= prevRemaining) break;
-      prevRemaining = remaining;
+      // Non-progress signal, now O(1): `changes` summed across this batch's
+      // deletes + failed-parks is the DIRECT evidence the guard's old count
+      // diff inferred — a resolver returning a mismatched name makes the keyed
+      // cleanup no-op, which shows up here as zero removals. The per-batch
+      // COUNT(*) it replaces walked every remaining pending row — O(N²/batch)
+      // over a run, 93.9s of the kernel-scale batch loop (§7a.2). A REAL count
+      // runs only on the suspicious path (claimed-work batch removed nothing —
+      // e.g. every row was a sibling a legacy-key sweep already consumed),
+      // where it arbitrates stop-vs-continue exactly as before.
+      if (removedThisBatch <= 0 && batch.length > 0) {
+        tLp = Date.now();
+        const remaining = this.queries.getUnresolvedReferencesCount();
+        lp('countGuard', tLp);
+        if (remaining >= prevRemaining) break;
+        prevRemaining = remaining;
+      }
 
       // Advance the pipeline: the prefetched batch (already fanned out when
       // the pool is on) becomes the current one.
@@ -1639,6 +1720,12 @@ export class ReferenceResolver {
       if (pool) await pool.destroy().catch(() => undefined);
     }
 
+    if (loopProf) {
+      const parts = Object.entries(loopProf).map(([k, v]) => `${k}=${(v / 1000).toFixed(1)}s`).join(' ');
+      console.error(`[resolve-profile] loop-stages ${parts}`);
+    }
+    this.dumpResolveProfile('main');
+
     return {
       resolved: [],
       unresolved: [],

src/resolution/resolver-worker.ts

@@ -20,7 +20,7 @@ try {
   (require('node:module') as { enableCompileCache?: () => void }).enableCompileCache?.();
 } catch { /* cache is best-effort */ }
 
-import { parentPort } from 'worker_threads';
+import { parentPort, threadId } from 'worker_threads';
 import { createDatabase, SqliteDatabase } from '../db/sqlite-adapter';
 import { QueryBuilder } from '../db/queries';
 import { ReferenceResolver } from './index';
@@ -95,6 +95,9 @@ port.on('message', (msg: InMessage) => {
         break;
       }
       case 'close': {
+        try {
+          resolver?.dumpResolveProfile(`worker#${threadId}`);
+        } catch { /* diagnostics never block shutdown */ }
         try {
           db?.close();
         } catch {