@soulcraft/cor 3.0.14 → 3.0.16

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -31,6 +31,8 @@
31
31
  * collation contract (code-point order, matching `compareCodePoints`).
32
32
  */
33
33
  import { existsSync } from 'node:fs';
34
+ import { composeInvariantReport } from './invariantReport.js';
35
+ import { declareDerivedFamilies, withFamilyRetirement, COLUMN_FAMILIES } from './derivedFamilies.js';
34
36
  import { prodLog } from '@soulcraft/brainy/internals';
35
37
  import { ValueType } from './columnStoreTypes.js';
36
38
  import { ColumnManifest } from './ColumnManifest.js';
@@ -145,6 +147,10 @@ export class NativeColumnStore {
145
147
  'Use brainy FileSystemStorage or another adapter implementing BinaryBlobStorage.');
146
148
  }
147
149
  this.storage = storage;
150
+ // ADR-004 §7: declare the column family (feature-detected ≥8.3.0) —
151
+ // the one family whose members are written via saveBinaryBlob, so the
152
+ // storage layer's delete-refusal actively bites here.
153
+ await declareDerivedFamilies(storage, COLUMN_FAMILIES);
148
154
  // Discover fields by listing manifest files under the column-index prefix.
149
155
  let paths;
150
156
  try {
@@ -241,6 +247,43 @@ export class NativeColumnStore {
241
247
  * would have been invisible.
242
248
  * @param field - Field about to be served.
243
249
  */
250
+ /**
251
+ * ADR-004 §6 — the column store's invariant self-report. NEVER throws;
252
+ * bounded (set sizes + per-field segment counts, no IO).
253
+ */
254
+ async validateInvariants() {
255
+ return composeInvariantReport('column', () => this.storage !== null && this.faultedFields.size === 0, [
256
+ () => ({
257
+ name: 'fields-servable',
258
+ holds: this.faultedFields.size === 0,
259
+ detail: this.faultedFields.size === 0
260
+ ? `${this.fieldTypes.size} field(s), none faulted`
261
+ : `${this.faultedFields.size} field(s) UNAVAILABLE after faulted ` +
262
+ `loads (queries throw): ${[...this.faultedFields].join(', ')}`,
263
+ expected: this.fieldTypes.size,
264
+ actual: this.fieldTypes.size - this.faultedFields.size,
265
+ heal: 'rebuild',
266
+ }),
267
+ () => {
268
+ // manifest-residency: every field's manifest-listed segments are
269
+ // attached in the native engine (hasField is the cheap proxy for
270
+ // a field whose segments loaded; per-segment ids are not exposed
271
+ // by the native surface, so the count-level check is the honest
272
+ // bounded version — a faulted segment already trips
273
+ // fields-servable above).
274
+ const unattached = [...this.manifests.keys()].filter((field) => !this.faultedFields.has(field) && !this.native.hasField(field));
275
+ return {
276
+ name: 'manifest-residency',
277
+ holds: unattached.length === 0,
278
+ detail: unattached.length === 0
279
+ ? `${this.manifests.size} field manifest(s), all attached`
280
+ : `${unattached.length} field(s) have persisted manifests but no ` +
281
+ `attached native state: ${unattached.join(', ')}`,
282
+ heal: 'rebuild',
283
+ };
284
+ },
285
+ ]);
286
+ }
244
287
  assertFieldServable(field) {
245
288
  if (this.faultedFields.has(field)) {
246
289
  throw new Error(`NativeColumnStore: column index for field '${field}' is UNAVAILABLE — ` +
@@ -464,18 +507,25 @@ export class NativeColumnStore {
464
507
  // Delete persisted column-index files (manifests + segment/deleted blobs)
465
508
  // before dropping the in-memory manifests we need to locate them.
466
509
  if (this.storage) {
467
- for (const [field, manifest] of this.manifests) {
468
- try {
469
- await this.storage.deleteObjectFromPath(manifest.manifestPath());
470
- }
471
- catch {
472
- /* manifest already gone */
473
- }
474
- for (const seg of manifest.getAllSegments()) {
475
- await this.storage.deleteBinaryBlob(segmentBlobKey(field, seg.level, seg.id));
510
+ // Owner-sanctioned retirement (ADR-004 §7): this reset legitimately
511
+ // deletes the family's own members (a rebuild reassigns entity ints,
512
+ // invalidating old segments), so the protection window opens and
513
+ // closes around exactly this loop.
514
+ const storage = this.storage;
515
+ await withFamilyRetirement(storage, COLUMN_FAMILIES[0], async () => {
516
+ for (const [field, manifest] of this.manifests) {
517
+ try {
518
+ await storage.deleteObjectFromPath(manifest.manifestPath());
519
+ }
520
+ catch {
521
+ /* manifest already gone */
522
+ }
523
+ for (const seg of manifest.getAllSegments()) {
524
+ await storage.deleteBinaryBlob(segmentBlobKey(field, seg.level, seg.id));
525
+ }
526
+ await storage.deleteBinaryBlob(deletedBlobKey(field));
476
527
  }
477
- await this.storage.deleteBinaryBlob(deletedBlobKey(field));
478
- }
528
+ });
479
529
  }
480
530
  // Drop all native state by replacing the engine, then clear TS-side state.
481
531
  // Faults clear too: the rebuild that follows repopulates every field
@@ -17,6 +17,7 @@
17
17
  */
18
18
  import type { StorageAdapter } from '@soulcraft/brainy';
19
19
  import { NounType, VerbType } from '@soulcraft/brainy/types/graphTypes';
20
+ import type { ProviderInvariantReport } from '@soulcraft/brainy';
20
21
  import { NativeColumnStore } from './NativeColumnStore.js';
21
22
  import type { EntityIdMapperLike } from './columnStoreTypes.js';
22
23
  import type { MetadataIndexProvider } from '../providerContracts.js';
@@ -115,6 +116,14 @@ export declare class MetadataIndexManager implements MetadataIndexProvider {
115
116
  * everything from canonical. Cleared by a completed {@link rebuild}.
116
117
  */
117
118
  private strandDetected;
119
+ /**
120
+ * SSTable ids actually REGISTERED into the live engine (cold-open
121
+ * registration + post-flush registration). The `manifest-residency`
122
+ * invariant (ADR-004 §6) compares the persisted manifest's claims
123
+ * against this set — the manifest is a claim; registration is the truth.
124
+ * Cleared on the rebuild's fresh re-open.
125
+ */
126
+ private lsmRegisteredSstableIds;
118
127
  /**
119
128
  * **#18** — set true by {@link guardIndexEpoch} outcome 4 (large stale brain,
120
129
  * auto-migrate enabled) and read once at the END of {@link init}, which kicks the
@@ -803,7 +812,28 @@ export declare class MetadataIndexManager implements MetadataIndexProvider {
803
812
  getAllVFSEntityCounts(): Promise<Map<string, number>>;
804
813
  getTotalVFSEntityCount(): Promise<number>;
805
814
  getCountForCriteria(field: string, value: any): Promise<number>;
815
+ /**
816
+ * ADR-004 §6 — the provider's self-report of its cross-layer invariants.
817
+ * NEVER throws; bounded (residency + O(1) count reads, one limit-1
818
+ * canonical probe); safe on a live brain. Brainy's
819
+ * `validateIndexConsistency` feature-detects and aggregates this;
820
+ * `repairIndex` maps `heal: 'rebuild'` failures to {@link rebuild}.
821
+ */
822
+ validateInvariants(): Promise<ProviderInvariantReport>;
823
+ /**
824
+ * Serializes every {@link flush} pass. Two flushes interleaving at the
825
+ * awaits is NOT hypothetical — memory-vm runs a periodic flusher AND
826
+ * brainy's shutdown hook flushes on SIGTERM — and the flushLsm pair
827
+ * both read `nextSstableId(manifest)` before either pushes: SAME id →
828
+ * same blob key → the second write OVERWRITES the first SSTable →
829
+ * postings lost with clean logs and NO replay anomaly (the exact
830
+ * acceptance-#1 signature on CORTEX-RESTART-STRAND, 2026-07-13).
831
+ * Chaining gives each caller its own full pass covering everything
832
+ * written before its call — the semantics a shutdown flush needs.
833
+ */
834
+ private flushChain;
806
835
  flush(): Promise<void>;
836
+ private flushOnce;
807
837
  /**
808
838
  * Persist all state and release the column store's resources. Brainy's
809
839
  * shutdown path calls `metadataIndex.close()` when present; flushing here
@@ -34,6 +34,8 @@ const BUCKETED_INDEX_FIELDS = new Set([
34
34
  ]);
35
35
  import { TypeUtils, NOUN_TYPE_COUNT, VERB_TYPE_COUNT } from '@soulcraft/brainy/types/graphTypes';
36
36
  import { loadNativeModule } from '../native/index.js';
37
+ import { composeInvariantReport } from './invariantReport.js';
38
+ import { declareDerivedFamilies, METADATA_FAMILIES } from './derivedFamilies.js';
37
39
  import { openOrCreateBinaryIdMapper } from './binaryIdMapperFactory.js';
38
40
  import { NativeColumnStore } from './NativeColumnStore.js';
39
41
  import { ResourceManager } from '../resource/ResourceManager.js';
@@ -189,6 +191,14 @@ export class MetadataIndexManager {
189
191
  * everything from canonical. Cleared by a completed {@link rebuild}.
190
192
  */
191
193
  strandDetected = false;
194
+ /**
195
+ * SSTable ids actually REGISTERED into the live engine (cold-open
196
+ * registration + post-flush registration). The `manifest-residency`
197
+ * invariant (ADR-004 §6) compares the persisted manifest's claims
198
+ * against this set — the manifest is a claim; registration is the truth.
199
+ * Cleared on the rebuild's fresh re-open.
200
+ */
201
+ lsmRegisteredSstableIds = new Set();
192
202
  /**
193
203
  * **#18** — set true by {@link guardIndexEpoch} outcome 4 (large stale brain,
194
204
  * auto-migrate enabled) and read once at the END of {@link init}, which kicks the
@@ -485,6 +495,7 @@ export class MetadataIndexManager {
485
495
  throw new Error(`openLsmEngine: getBinaryBlobPath returned null for persisted SSTable '${key}'`);
486
496
  }
487
497
  engine.registerSstablePath(path);
498
+ this.lsmRegisteredSstableIds.add(entry.id);
488
499
  }
489
500
  }
490
501
  // Cold-open replay forensics (CORTEX-RESTART-STRAND). The engine's
@@ -699,6 +710,7 @@ export class MetadataIndexManager {
699
710
  });
700
711
  this.lsmEngine = engine;
701
712
  this.lsmManifest = newEmptyLsmManifest();
713
+ this.lsmRegisteredSstableIds.clear();
702
714
  this.native.setLsmEngine(engine);
703
715
  // Re-apply the current admission ceiling immediately (the subscriber
704
716
  // callback only refreshes it on the next rebalance/pressure tick). The
@@ -783,6 +795,7 @@ export class MetadataIndexManager {
783
795
  throw new Error(`flushLsm: getBinaryBlobPath returned null for '${key}'`);
784
796
  }
785
797
  engine.registerSstablePath(path);
798
+ this.lsmRegisteredSstableIds.add(id);
786
799
  // (4) Retire the memtable log below the flushed epoch. The `.sst` bytes
787
800
  // AND the manifest referencing them are already durable here, so a failure
788
801
  // at THIS step is non-fatal per the ordering invariant above: the records
@@ -1030,15 +1043,29 @@ export class MetadataIndexManager {
1030
1043
  getUuid: (intId) => native.intToUuid(intId) ?? undefined,
1031
1044
  getInt: (uuid) => native.uuidToInt(uuid) ?? undefined,
1032
1045
  getOrAssign: (uuid) => {
1033
- // The native index assigns ids inside addToIndex. If a caller resolves an
1034
- // id before the entity is indexed, surface the existing mapping; an
1035
- // unmapped UUID has no int id yet and is reported as such rather than
1036
- // fabricating one (which would desync the canonical native mapper).
1046
+ // The native index assigns ids inside addToIndex; a mapped UUID
1047
+ // resolves to its existing int. An UNMAPPED uuid here is a
1048
+ // canonical-committed entity whose index ops were lost (the
1049
+ // restart-strand class memory's boot heal hit exactly this,
1050
+ // CORTEX-RESTART-STRAND 2026-07-13): the cold vector rebuild
1051
+ // walks canonical and must be able to MINT the id through the
1052
+ // SAME allocator addToIndex uses (durable via the mapper's delta
1053
+ // log; a later addToIndex reuses it) instead of aborting the
1054
+ // whole heal. The entity's missing metadata postings stay
1055
+ // visible to posted-count-floor — this masks nothing.
1037
1056
  const existing = native.uuidToInt(uuid);
1038
1057
  if (existing !== null)
1039
1058
  return existing;
1040
- throw new Error(`EntityIdMapper.getOrAssign: '${uuid}' has no native id yet — ids are ` +
1041
- `assigned by addToIndex. Index the entity before resolving its int id.`);
1059
+ if (typeof native.getOrAssignInt === 'function') {
1060
+ prodLog.warn(`EntityIdMapper.getOrAssign: '${uuid}' had no native id at resolve ` +
1061
+ `time (canonical-committed entity with lost index ops) — assigned ` +
1062
+ `through the canonical allocator; its metadata postings heal via ` +
1063
+ `the posted-count invariant / rebuild.`);
1064
+ return Number(native.getOrAssignInt(uuid));
1065
+ }
1066
+ throw new Error(`EntityIdMapper.getOrAssign: '${uuid}' has no native id yet and the ` +
1067
+ `loaded native binary predates 3.0.16's getOrAssignInt — upgrade ` +
1068
+ `the paired .node build.`);
1042
1069
  },
1043
1070
  intsIterableToUuids: (ints) => native.intsToUuids([...ints]),
1044
1071
  // Brainy 8.0 `restore()` calls this BEFORE the metadata/vector/graph
@@ -1073,6 +1100,9 @@ export class MetadataIndexManager {
1073
1100
  // UUID ↔ entity-int allocator. Must run before columnStore.init (which
1074
1101
  // resolves ids through getIdMapper()) and before any mutation.
1075
1102
  this.wireIdMapper();
1103
+ // ADR-004 §7: declare this provider's protected families (feature-
1104
+ // detected on brainy ≥8.3.0; idempotent upsert; loud-but-non-blocking).
1105
+ await declareDerivedFamilies(this.storage, METADATA_FAMILIES);
1076
1106
  // LSM durability (steps 3-4): open the native LSM engine here (not in the
1077
1107
  // constructor) so it can be seeded from the on-disk manifest and cold-load
1078
1108
  // every persisted SSTable into its live read set — the entire restore,
@@ -1114,6 +1144,7 @@ export class MetadataIndexManager {
1114
1144
  prodLog.info('[NativeMetadataIndex] Field registry missing but the LSM engine ' +
1115
1145
  'recovered durable state (SSTables/memtable) — serving from it, ' +
1116
1146
  'NO rebuild.');
1147
+ this.readinessInitialized = true;
1117
1148
  return;
1118
1149
  }
1119
1150
  // **Tripwire — a rebuild here on a PREVIOUSLY-HEALTHY brain is a bug,
@@ -1142,11 +1173,21 @@ export class MetadataIndexManager {
1142
1173
  }
1143
1174
  console.warn(`[NativeMetadataIndex] Field registry missing but ${probe.totalCount ?? 'unknown'} entities exist — rebuilding index`);
1144
1175
  await this.rebuild();
1176
+ // A completed rebuild IS the readiness determination for this
1177
+ // open (sweep finding: these early exits previously left
1178
+ // readinessInitialized false — a fresh-rebuilt brain reported
1179
+ // not-ready forever, and brainy's gate re-rebuilt every boot).
1180
+ this.readinessInitialized = true;
1145
1181
  return;
1146
1182
  }
1183
+ // Probe says genuinely empty — an empty brain is READY by contract.
1184
+ this.readinessInitialized = true;
1147
1185
  }
1148
1186
  catch {
1149
- // Storage probe failed — genuinely empty
1187
+ // Storage probe failed — treated as empty per current semantics;
1188
+ // readiness stays UNINITIALIZED (conservative: a faulted probe must
1189
+ // not certify an unverified brain as ready — the fault-vs-absent
1190
+ // discrimination for this probe is tracked ADR-004 §4 M-work).
1150
1191
  }
1151
1192
  return;
1152
1193
  }
@@ -1946,6 +1987,14 @@ export class MetadataIndexManager {
1946
1987
  const resultJson = this.native.removeFromIndex(id);
1947
1988
  const result = JSON.parse(resultJson);
1948
1989
  await this.persistMutationResult(result);
1990
+ // Count symmetry (sweep finding; ADR-004 anti-pattern C): this path
1991
+ // has no metadata to know WHICH type to decrement, and skipping the
1992
+ // decrement permanently inflated totalEntitiesByType /
1993
+ // entityCountsByTypeFixed until restart. Instead of guessing,
1994
+ // RECOMPUTE the counters from native ground truth — bounded
1995
+ // (per-type O(1) bitmap-cardinality reads, no id materialization).
1996
+ await this.lazyLoadCounts();
1997
+ this.syncTypeCountsToFixed();
1949
1998
  }
1950
1999
  // Tombstone in the column store so the entity drops out of sort/filter/range.
1951
2000
  if (intId !== null) {
@@ -2278,7 +2327,195 @@ export class MetadataIndexManager {
2278
2327
  // ==========================================================================
2279
2328
  // Flush
2280
2329
  // ==========================================================================
2330
+ /**
2331
+ * ADR-004 §6 — the provider's self-report of its cross-layer invariants.
2332
+ * NEVER throws; bounded (residency + O(1) count reads, one limit-1
2333
+ * canonical probe); safe on a live brain. Brainy's
2334
+ * `validateIndexConsistency` feature-detects and aggregates this;
2335
+ * `repairIndex` maps `heal: 'rebuild'` failures to {@link rebuild}.
2336
+ */
2337
+ async validateInvariants() {
2338
+ return composeInvariantReport('metadata', () => this.isReady(), [
2339
+ // manifest-residency: every persisted-manifest-named SSTable is
2340
+ // actually registered in the live engine.
2341
+ () => {
2342
+ const manifest = this.lsmManifest;
2343
+ if (manifest === null || this.lsmMetaDir === null) {
2344
+ return {
2345
+ name: 'manifest-residency',
2346
+ holds: true,
2347
+ detail: 'in-RAM engine or fresh brain — no persisted SSTables to verify',
2348
+ heal: 'none',
2349
+ };
2350
+ }
2351
+ const missing = manifest.sstables
2352
+ .map((e) => e.id)
2353
+ .filter((id) => !this.lsmRegisteredSstableIds.has(id));
2354
+ return {
2355
+ name: 'manifest-residency',
2356
+ holds: missing.length === 0,
2357
+ detail: missing.length === 0
2358
+ ? `${manifest.sstables.length} manifest-named SSTable(s), all registered`
2359
+ : `manifest names ${manifest.sstables.length} SSTable(s) but ` +
2360
+ `${missing.length} never registered (ids: ${missing.join(', ')}) — ` +
2361
+ `the index serves a subset of its durable postings`,
2362
+ expected: manifest.sstables.length,
2363
+ actual: manifest.sstables.length - missing.length,
2364
+ heal: 'rebuild',
2365
+ };
2366
+ },
2367
+ // posted-count-floor: Σ posted noun counts ≥ canonical noun count.
2368
+ async () => {
2369
+ const getNouns = this.storage.getNouns;
2370
+ if (typeof getNouns !== 'function') {
2371
+ return {
2372
+ name: 'posted-count-floor',
2373
+ holds: true,
2374
+ detail: 'adapter exposes no getNouns — floor unverifiable, skipped',
2375
+ heal: 'none',
2376
+ };
2377
+ }
2378
+ let canonical;
2379
+ try {
2380
+ const probe = await getNouns.call(this.storage, { pagination: { limit: 1 } });
2381
+ canonical = probe?.totalCount;
2382
+ }
2383
+ catch (error) {
2384
+ return {
2385
+ name: 'posted-count-floor',
2386
+ holds: false,
2387
+ detail: `canonical count probe FAULTED — divergence unverifiable: ${String(error)}`,
2388
+ heal: 'none',
2389
+ };
2390
+ }
2391
+ if (typeof canonical !== 'number' || !Number.isFinite(canonical) || canonical >= 0xffff_ffff) {
2392
+ return {
2393
+ name: 'posted-count-floor',
2394
+ holds: true,
2395
+ detail: `canonical count not exactly countable (${String(canonical)}) — floor skipped by design`,
2396
+ heal: 'none',
2397
+ };
2398
+ }
2399
+ const { NounType } = await import('@soulcraft/brainy');
2400
+ let posted = 0;
2401
+ for (const t of Object.values(NounType)) {
2402
+ posted += this.native.getEntityCountByType(t);
2403
+ }
2404
+ return {
2405
+ name: 'posted-count-floor',
2406
+ holds: posted >= canonical,
2407
+ detail: posted >= canonical
2408
+ ? `posted ${posted} ≥ canonical ${canonical}`
2409
+ : `canonical holds ${canonical} entities but only ${posted} are ` +
2410
+ `posted — ${canonical - posted} durable-but-unqueryable`,
2411
+ expected: canonical,
2412
+ actual: posted,
2413
+ heal: 'rebuild',
2414
+ };
2415
+ },
2416
+ // replay-clean: the cold open recovered nothing anomalous.
2417
+ () => {
2418
+ const engine = this.lsmEngine;
2419
+ if (!engine || typeof engine.replayStatsJson !== 'function') {
2420
+ return {
2421
+ name: 'replay-clean',
2422
+ holds: true,
2423
+ detail: 'replay stats unavailable (in-RAM engine or pre-3.0.12 binary)',
2424
+ heal: 'none',
2425
+ };
2426
+ }
2427
+ const s = JSON.parse(engine.replayStatsJson());
2428
+ const unreadable = s.archivesUnreadable ?? 0;
2429
+ const recreated = s.liveLogsRecreated ?? 0;
2430
+ const clean = unreadable === 0 && recreated === 0;
2431
+ return {
2432
+ name: 'replay-clean',
2433
+ holds: clean,
2434
+ detail: clean
2435
+ ? 'cold-open replay clean (no unreadable archives, no recreated live logs)'
2436
+ : `cold-open replay anomalies: ${recreated} live log(s) were missing ` +
2437
+ `(recreated), ${unreadable} archive(s) unreadable`,
2438
+ heal: 'rebuild',
2439
+ };
2440
+ },
2441
+ // registry-discoverable: every field this process knows is in the
2442
+ // persisted registry a cold open would discover.
2443
+ async () => {
2444
+ if (this.knownFields.size === 0) {
2445
+ return {
2446
+ name: 'registry-discoverable',
2447
+ holds: true,
2448
+ detail: 'no fields registered yet',
2449
+ heal: 'none',
2450
+ };
2451
+ }
2452
+ let persisted = [];
2453
+ try {
2454
+ const raw = await this.storage.getMetadata('__metadata_field_registry__');
2455
+ persisted = Array.isArray(raw?.fields)
2456
+ ? (raw.fields)
2457
+ : Array.isArray(raw)
2458
+ ? raw
2459
+ : [];
2460
+ }
2461
+ catch (error) {
2462
+ return {
2463
+ name: 'registry-discoverable',
2464
+ holds: false,
2465
+ detail: `persisted registry read FAULTED — discoverability unverifiable: ${String(error)}`,
2466
+ heal: 'none',
2467
+ };
2468
+ }
2469
+ const persistedSet = new Set(persisted);
2470
+ const undiscoverable = [...this.knownFields].filter((f) => !persistedSet.has(f));
2471
+ return {
2472
+ name: 'registry-discoverable',
2473
+ holds: undiscoverable.length === 0,
2474
+ detail: undiscoverable.length === 0
2475
+ ? `${this.knownFields.size} known field(s), all in the persisted registry`
2476
+ : `${undiscoverable.length} known field(s) missing from the persisted ` +
2477
+ `registry (would be undiscoverable after restart): ${undiscoverable.join(', ')}`,
2478
+ expected: this.knownFields.size,
2479
+ actual: this.knownFields.size - undiscoverable.length,
2480
+ heal: 'repair',
2481
+ };
2482
+ },
2483
+ ]);
2484
+ }
2485
+ /**
2486
+ * Serializes every {@link flush} pass. Two flushes interleaving at the
2487
+ * awaits is NOT hypothetical — memory-vm runs a periodic flusher AND
2488
+ * brainy's shutdown hook flushes on SIGTERM — and the flushLsm pair
2489
+ * both read `nextSstableId(manifest)` before either pushes: SAME id →
2490
+ * same blob key → the second write OVERWRITES the first SSTable →
2491
+ * postings lost with clean logs and NO replay anomaly (the exact
2492
+ * acceptance-#1 signature on CORTEX-RESTART-STRAND, 2026-07-13).
2493
+ * Chaining gives each caller its own full pass covering everything
2494
+ * written before its call — the semantics a shutdown flush needs.
2495
+ */
2496
+ flushChain = Promise.resolve();
2281
2497
  async flush() {
2498
+ // Flush-during-rebuild guard (CORTEX-RESTART-STRAND, memory-vm's
2499
+ // 80-minute heal): a consumer's periodic flush cycle running DURING a
2500
+ // rebuild drains the growing memtable + rewrites field indexes on
2501
+ // every pass — O(current-index-size) IO per pass, interleaved with the
2502
+ // walk (measured live: 572 msync/s + escalating 29s→45s flush passes).
2503
+ // The rebuild's own completion flush owns durability (canonical is the
2504
+ // source of truth throughout), so a concurrent flush only multiplies
2505
+ // IO. Loud no-op, never silent.
2506
+ if (this.isRebuilding) {
2507
+ prodLog.warn('[NativeMetadataIndex] flush() during an active rebuild — deferred ' +
2508
+ '(the rebuild\'s completion flush owns durability; a concurrent ' +
2509
+ 'flush would multiply heal IO).');
2510
+ return;
2511
+ }
2512
+ const run = this.flushChain.then(() => this.flushOnce());
2513
+ // Keep the chain alive on failure — the NEXT flush must still run; the
2514
+ // failure itself propagates to THIS caller below.
2515
+ this.flushChain = run.catch(() => { });
2516
+ return run;
2517
+ }
2518
+ async flushOnce() {
2282
2519
  // Always save field registry + msync the entity ID mapper — even with
2283
2520
  // no dirty fields. The field registry is the critical file init() needs
2284
2521
  // to discover persisted indices. Without it, the metadata index appears
@@ -2679,7 +2916,10 @@ export class MetadataIndexManager {
2679
2916
  if (bulkCapable)
2680
2917
  this.lsmEngine.setBulkLoad(false);
2681
2918
  }
2682
- // Final flush
2919
+ // Final flush. isRebuilding drops FIRST so the completion flush
2920
+ // passes the flush-during-rebuild guard above; readiness is still
2921
+ // gated by strandDetected until the flush lands and clears it below.
2922
+ this.isRebuilding = false;
2683
2923
  await this.flushRebuildDirty();
2684
2924
  await this.flush();
2685
2925
  prodLog.info(`Metadata index rebuild completed! Processed ${totalNounsProcessed} nouns and ${totalVerbsProcessed} verbs`);
@@ -2710,13 +2950,27 @@ export class MetadataIndexManager {
2710
2950
  let processed = 0;
2711
2951
  let sinceFlush = 0;
2712
2952
  let noMetadata = 0;
2953
+ // Heal observability (CORTEX-RESTART-STRAND, memory's ask): a rebuild
2954
+ // must never be silent for minutes — the operator cannot distinguish
2955
+ // converging from stuck. Every 500 posted entities logs progress with
2956
+ // the sustained rate, so a pathological per-item cost names itself in
2957
+ // the journal in real time.
2958
+ const startedAt = Date.now();
2959
+ let lastProgressAt = 0;
2960
+ const fetchPage = (c) => kind === 'noun'
2961
+ ? this.storage.getNouns({ pagination: { limit: PAGE, cursor: c } })
2962
+ : this.storage.getVerbs({ pagination: { limit: PAGE, cursor: c } });
2963
+ // Prefetch pipeline: the NEXT canonical page is fetched while the
2964
+ // CURRENT page indexes, so adapter enumeration latency overlaps the
2965
+ // RAM-speed posting work instead of adding to it.
2966
+ let pending = fetchPage(cursor);
2713
2967
  while (hasMore) {
2714
- const page = kind === 'noun'
2715
- ? await this.storage.getNouns({ pagination: { limit: PAGE, cursor } })
2716
- : await this.storage.getVerbs({ pagination: { limit: PAGE, cursor } });
2968
+ const page = await pending;
2717
2969
  const items = page.items;
2718
2970
  if (items.length === 0)
2719
2971
  break; // empty page = done (defensive backstop)
2972
+ if (page.hasMore)
2973
+ pending = fetchPage(page.nextCursor);
2720
2974
  const ids = items.map((it) => it.id);
2721
2975
  const md = await this.loadRebuildMetadataBatch(kind, ids);
2722
2976
  for (const it of items) {
@@ -2733,6 +2987,12 @@ export class MetadataIndexManager {
2733
2987
  if (metadata) {
2734
2988
  await this.addToIndex(it.id, metadata, true, true);
2735
2989
  processed++;
2990
+ if (processed - lastProgressAt >= 500) {
2991
+ lastProgressAt = processed;
2992
+ const elapsedS = (Date.now() - startedAt) / 1000;
2993
+ prodLog.info(`[NativeMetadataIndex] rebuild(${kind}): ${processed} posted ` +
2994
+ `in ${elapsedS.toFixed(1)}s (${(processed / Math.max(elapsedS, 0.001)).toFixed(0)}/s)`);
2995
+ }
2736
2996
  if (++sinceFlush >= 5000) {
2737
2997
  await this.flushRebuildDirty();
2738
2998
  sinceFlush = 0;
@@ -2743,9 +3003,12 @@ export class MetadataIndexManager {
2743
3003
  }
2744
3004
  }
2745
3005
  hasMore = page.hasMore;
2746
- cursor = page.nextCursor;
2747
3006
  await this.yieldToEventLoop();
2748
3007
  }
3008
+ const totalS = (Date.now() - startedAt) / 1000;
3009
+ prodLog.info(`[NativeMetadataIndex] rebuild(${kind}): COMPLETE — ${processed} posted ` +
3010
+ `in ${totalS.toFixed(1)}s (${(processed / Math.max(totalS, 0.001)).toFixed(0)}/s)` +
3011
+ (noMetadata > 0 ? `, ${noMetadata} without metadata` : ''));
2749
3012
  if (noMetadata > 0) {
2750
3013
  prodLog.warn(`[NativeMetadataIndex] rebuild(${kind}): ${noMetadata} enumerated ` +
2751
3014
  `entit${noMetadata === 1 ? 'y' : 'ies'} carried no canonical ` +
@@ -2793,16 +3056,29 @@ export class MetadataIndexManager {
2793
3056
  else if (this.storage.getVerbMetadataBatch) {
2794
3057
  return this.storage.getVerbMetadataBatch(ids);
2795
3058
  }
3059
+ // Bounded-concurrency reads (16-way). The reads are independent, and a
3060
+ // SERIAL per-id walk multiplies the adapter's per-op latency by N — on
3061
+ // a slow adapter (memory-vm measured ~1s/op shapes) that turned a
3062
+ // seconds-long heal into ~1s/entity (CORTEX-RESTART-STRAND, the
3063
+ // 52-minute heal). Concurrency divides heal time by ~16 on ANY
3064
+ // adapter; fail-loud is preserved (a faulted read rejects the whole
3065
+ // batch via Promise.all — readRebuildMetadataOrThrow retries once then
3066
+ // throws, never silently excludes).
2796
3067
  const m = new Map();
2797
- for (const id of ids) {
2798
- // Retry-once-then-throw (readRebuildMetadataOrThrow): a faulted
2799
- // read must fail the rebuild loudly, never silently exclude the
2800
- // entity — a swallowed fault here completed the rebuild
2801
- // under-posted (the strand bug re-created by its own healer).
2802
- const md = await this.readRebuildMetadataOrThrow(kind, id);
2803
- if (md)
2804
- m.set(id, md);
2805
- }
3068
+ const CONCURRENCY = 16;
3069
+ let next = 0;
3070
+ const workers = Array.from({ length: Math.min(CONCURRENCY, ids.length) }, async () => {
3071
+ for (;;) {
3072
+ const i = next++;
3073
+ if (i >= ids.length)
3074
+ break;
3075
+ const id = ids[i];
3076
+ const md = await this.readRebuildMetadataOrThrow(kind, id);
3077
+ if (md)
3078
+ m.set(id, md);
3079
+ }
3080
+ });
3081
+ await Promise.all(workers);
2806
3082
  return m;
2807
3083
  }
2808
3084
  /**
@@ -15,7 +15,7 @@
15
15
  * brain pays only for what it uses. This is the constant-RAM,
16
16
  * billion-scale store that replaces the resident JSON id-maps (#72).
17
17
  */
18
- import { existsSync } from 'node:fs';
18
+ import { statSync } from 'node:fs';
19
19
  import { loadNativeModule } from '../native/index.js';
20
20
  /** Default storage key for the extendible-hash UUID → int file. */
21
21
  export const DEFAULT_UUID_TO_INT_KEY = '_id_mapper/uuid_to_int.mkv';
@@ -65,10 +65,17 @@ export function openOrCreateBinaryIdMapper(storage, opts = {}) {
65
65
  // Both files must exist together (paired write semantics). A
66
66
  // half-present state is crash corruption surfaced as an error,
67
67
  // never silently recreated.
68
+ //
69
+ // Fault-vs-absent (spine ADR-004 §4): the probes below THROW on any stat
70
+ // fault other than a clean ENOENT. `existsSync` returned false on
71
+ // EACCES/EIO/EMFILE too — and a double false routed to `create()`, which
72
+ // TRUNCATES both stores: a transient permissions blip (backup tool,
73
+ // scanner, remount) could wipe the brain's entire UUID↔int allocator.
74
+ // Deciding open-vs-create is only legal on certain knowledge.
68
75
  let mapper;
69
76
  let created;
70
- const uuidFileExists = existsSync(uuidToIntPath);
71
- const intFileExists = existsSync(intToUuidPath);
77
+ const uuidFileExists = durableFilePresent(uuidToIntPath, uuidToIntKey);
78
+ const intFileExists = durableFilePresent(intToUuidPath, intToUuidKey);
72
79
  if (uuidFileExists && intFileExists) {
73
80
  mapper = NativeBinaryIdMapper.openExisting(config);
74
81
  created = false;
@@ -91,4 +98,27 @@ export function openOrCreateBinaryIdMapper(storage, opts = {}) {
91
98
  created,
92
99
  };
93
100
  }
101
+ /**
102
+ * Fault-discriminating existence probe for the id-mapper stores (spine
103
+ * ADR-004 §4). Returns `true`/`false` only on CERTAIN knowledge; any stat
104
+ * fault other than a clean ENOENT throws — because the caller's "neither
105
+ * file exists" branch invokes `create()`, which TRUNCATES, and a transient
106
+ * EACCES/EIO must never be allowed to masquerade as a fresh brain.
107
+ */
108
+ function durableFilePresent(path, key) {
109
+ try {
110
+ statSync(path);
111
+ return true;
112
+ }
113
+ catch (error) {
114
+ const code = error.code;
115
+ if (code === 'ENOENT')
116
+ return false;
117
+ throw new Error(`openOrCreateBinaryIdMapper: existence probe FAULTED for '${key}' ` +
118
+ `(${code ?? 'unknown'}) — refusing to decide open-vs-create on a ` +
119
+ `fault (create() truncates the brain's id allocator; a transient ` +
120
+ `IO/permissions fault must never wipe it). Restore storage health ` +
121
+ `and reopen. Cause: ${String(error)}`);
122
+ }
123
+ }
94
124
  //# sourceMappingURL=binaryIdMapperFactory.js.map