@cloudbitmaps/core 0.11.2 → 0.13.0

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.
Files changed (44) hide show
  1. package/dist/{chunk-5OMC5U7U.js → chunk-5W3XS2IU.js} +292 -22
  2. package/dist/chunk-5W3XS2IU.js.map +7 -0
  3. package/dist/core/audit.d.ts +6 -0
  4. package/dist/core/chunk-window.d.ts +24 -0
  5. package/dist/core/codec.d.ts +33 -1
  6. package/dist/core/consistency.d.ts +21 -2
  7. package/dist/core/cost.d.ts +91 -10
  8. package/dist/core/crbm/format.d.ts +42 -3
  9. package/dist/core/crbm/reader.d.ts +45 -3
  10. package/dist/core/crbm/writer.d.ts +32 -3
  11. package/dist/core/crbm-storage-source.d.ts +204 -23
  12. package/dist/core/crypto.d.ts +11 -7
  13. package/dist/core/determinism.d.ts +7 -0
  14. package/dist/core/engine.d.ts +41 -8
  15. package/dist/core/erase-id.d.ts +23 -3
  16. package/dist/core/erasure.d.ts +8 -4
  17. package/dist/core/generation-gc.d.ts +102 -4
  18. package/dist/core/load-input.d.ts +55 -0
  19. package/dist/core/load.d.ts +65 -10
  20. package/dist/core/metadata.d.ts +42 -0
  21. package/dist/core/metrics.d.ts +7 -1
  22. package/dist/core/pinned-storage-source.d.ts +2 -1
  23. package/dist/core/ports.d.ts +119 -13
  24. package/dist/core/registry-scan.d.ts +6 -0
  25. package/dist/core/retention-sweep.d.ts +47 -2
  26. package/dist/core/retry.d.ts +14 -0
  27. package/dist/core/rollback.d.ts +34 -4
  28. package/dist/core/summary.d.ts +61 -0
  29. package/dist/core/token.d.ts +36 -0
  30. package/dist/core/validate.d.ts +2 -0
  31. package/dist/driver-kit.js +39 -20
  32. package/dist/driver-kit.js.map +2 -2
  33. package/dist/drivers/_shared/entropy.d.ts +5 -0
  34. package/dist/drivers/_shared/object-registry.d.ts +36 -6
  35. package/dist/drivers/_shared/registry.d.ts +47 -19
  36. package/dist/drivers/localfs/registry.d.ts +10 -0
  37. package/dist/drivers/memory.d.ts +9 -3
  38. package/dist/drivers/retry/retrying-chunk-source.d.ts +2 -1
  39. package/dist/index.d.ts +4 -3
  40. package/dist/index.js +1878 -501
  41. package/dist/index.js.map +4 -4
  42. package/dist/testing/fuzz-core.d.ts +7 -6
  43. package/package.json +1 -1
  44. package/dist/chunk-5OMC5U7U.js.map +0 -7
package/dist/index.js CHANGED
@@ -5,6 +5,7 @@ import {
5
5
  DEFAULT_NAMESPACE,
6
6
  IntegrityError,
7
7
  KeyUnavailableError,
8
+ MAX_METADATA_BYTES,
8
9
  MAX_NAME_LENGTH,
9
10
  NotFoundError,
10
11
  RESERVED_NAMESPACE_PREFIX,
@@ -13,11 +14,18 @@ import {
13
14
  ValidationError,
14
15
  WriteConflictError,
15
16
  applyRegistryPatch,
17
+ assertRegistryCanWrite,
16
18
  brandAsBackend,
19
+ canonicalMetadataJson,
20
+ copiedMetadata,
17
21
  decodeNameFromPath,
22
+ drawIncarnation,
23
+ drawWrite,
18
24
  encodeNameForKey,
19
25
  encodeNameForPath,
20
26
  encodedNameLength,
27
+ entropyIsAvailable,
28
+ incarnationOf,
21
29
  isCloudRoaringError,
22
30
  isIntegrityError,
23
31
  isNotFoundError,
@@ -27,18 +35,23 @@ import {
27
35
  isValidationError,
28
36
  isWriteConflictError,
29
37
  mapWithConcurrency,
38
+ metadataBytes,
39
+ metadataFromBytes,
30
40
  namespacePathPart,
41
+ newIncarnationToken,
42
+ nextRegistryToken,
31
43
  parseRegistryEnvelope,
32
44
  recordFromNew,
33
- registryCounterOf,
45
+ sameIncarnation,
34
46
  serializeRegistryEnvelope,
35
47
  validateChunkRef,
36
48
  validateNewRegistryRecord,
37
49
  validateRegistryPatch,
38
50
  validateSegmentRef,
39
51
  validateUserNamespace,
40
- validateUserRef
41
- } from "./chunk-5OMC5U7U.js";
52
+ validateUserRef,
53
+ webCryptoEntropy
54
+ } from "./chunk-5W3XS2IU.js";
42
55
 
43
56
  // src/core/bit-route.ts
44
57
  var U32_MAX = 4294967295;
@@ -89,6 +102,110 @@ async function collectWithinBudget(source, budget, op) {
89
102
  return out;
90
103
  }
91
104
 
105
+ // src/core/chunk-window.ts
106
+ var ChunkWindow = class {
107
+ constructor(keys, fetch, max, ramp) {
108
+ this.keys = keys;
109
+ this.fetch = fetch;
110
+ this.max = max;
111
+ this.ramp = ramp;
112
+ }
113
+ keys;
114
+ fetch;
115
+ max;
116
+ ramp;
117
+ open = [];
118
+ launched = 0;
119
+ taken = 0;
120
+ /** The next read's result in key order, or the error that read raised. */
121
+ async take() {
122
+ const width = this.ramp ? Math.min(this.max, 2 ** Math.min(this.taken, 30)) : this.max;
123
+ while (this.launched < this.keys.length && this.launched - this.taken < width) {
124
+ this.open.push(
125
+ this.fetch(this.keys[this.launched++]).then(
126
+ (value) => ({ value }),
127
+ (cause) => ({ value: void 0, error: { cause } })
128
+ )
129
+ );
130
+ }
131
+ const slot = await this.open.shift();
132
+ this.taken += 1;
133
+ if (slot.error) throw slot.error.cause;
134
+ return slot.value;
135
+ }
136
+ };
137
+
138
+ // src/core/crbm/format.ts
139
+ var MAGIC = Uint8Array.of(67, 82, 66, 77);
140
+ var VERSION_MAJOR = 1;
141
+ var VERSION_MINOR = 0;
142
+ var PREAMBLE_BYTES = 8;
143
+ var PAYLOAD_START = PREAMBLE_BYTES;
144
+ var FOOTER_BYTES = 104;
145
+ var FOOTER = {
146
+ indexOffset: 0,
147
+ // u64
148
+ indexLength: 8,
149
+ // u64
150
+ indexCrc32c: 16,
151
+ // u32
152
+ flags: 20,
153
+ // u32
154
+ payloadCodecId: 24,
155
+ // u16 — which codec wrote the chunk payloads
156
+ elementWidth: 26,
157
+ // u8
158
+ containerCodec: 27,
159
+ // u8
160
+ versionMajor: 28,
161
+ // u8
162
+ versionMinor: 29,
163
+ // u8
164
+ reserved2: 30,
165
+ // u16
166
+ generation: 32,
167
+ // u64
168
+ indexNonce: 40,
169
+ // 12 B
170
+ indexTag: 52,
171
+ // 16 B
172
+ keyId: 68,
173
+ // 16 B
174
+ chunkCount: 84,
175
+ // u32
176
+ totalCardinality: 88,
177
+ // u64
178
+ footerCrc32c: 96,
179
+ // u32 — covers footer bytes [0, 96)
180
+ endMagic: 100
181
+ // char[4]
182
+ };
183
+ var FOOTER_CRC_COVERAGE = FOOTER.footerCrc32c;
184
+ var FLAG_ENCRYPTED = 1 << 0;
185
+ var FLAG_INDEX_COMPRESSED = 1 << 1;
186
+ var FLAG_LITTLE_ENDIAN = 1 << 2;
187
+ var FLAG_EXTENSION = 1 << 3;
188
+ var KNOWN_FLAGS = FLAG_ENCRYPTED | FLAG_LITTLE_ENDIAN | FLAG_EXTENSION;
189
+ var ELEMENT_WIDTH_32 = 32;
190
+ var PAYLOAD_CODEC_ROARING_PORTABLE = 1;
191
+ var KNOWN_PAYLOAD_CODEC_IDS = /* @__PURE__ */ new Set([
192
+ PAYLOAD_CODEC_ROARING_PORTABLE
193
+ ]);
194
+ var CONTAINER_CODEC_NONE = 0;
195
+ var DEFAULT_TAIL_BYTES = 256 * 1024;
196
+ var DEFAULT_MAX_INDEX_BYTES = 8 * 1024 * 1024;
197
+ var DEFAULT_MAX_BITMAP_BYTES = 1 << 20;
198
+ var MAX_CHUNK_CARDINALITY = 65536;
199
+ var EXT_MAGIC = Uint8Array.of(67, 82, 66, 88);
200
+ var EXT_TRAILER_BYTES = 12;
201
+ var EXT_SECTION_HEADER_BYTES = 5;
202
+ var MAX_EXT_BYTES = 4 * 1024;
203
+ var EXT_SECTION_METADATA = 1;
204
+ var CRC32C_BYTES = 4;
205
+ var AEAD_NONCE_BYTES = 12;
206
+ var AEAD_TAG_BYTES = 16;
207
+ var DEFAULT_MAX_PAYLOAD_BYTES = DEFAULT_MAX_BITMAP_BYTES + AEAD_NONCE_BYTES + AEAD_TAG_BYTES;
208
+
92
209
  // src/core/keys.ts
93
210
  var FIELD = " ";
94
211
  var NO_NAMESPACE = "/";
@@ -201,8 +318,8 @@ var CountingMetricsSink = class _CountingMetricsSink {
201
318
  };
202
319
 
203
320
  // src/core/engine.ts
204
- var DEFAULT_MAX_BITMAP_BYTES = 1 << 20;
205
- var DEFAULT_INTERSECT_CONCURRENCY = 8;
321
+ var DEFAULT_INTERSECT_CONCURRENCY = 32;
322
+ var COMBINE_WINDOW_START = 8;
206
323
  var ZERO_CLOCK = { now: () => 0 };
207
324
  function windowOf(range) {
208
325
  const after = range?.after;
@@ -255,37 +372,6 @@ function keysWithin(keys, w) {
255
372
  };
256
373
  return keys.slice(firstAtLeast(w.loKey), firstAtLeast(w.hiKey + 1));
257
374
  }
258
- var ChunkWindow = class {
259
- constructor(keys, fetch, max, ramp) {
260
- this.keys = keys;
261
- this.fetch = fetch;
262
- this.max = max;
263
- this.ramp = ramp;
264
- }
265
- keys;
266
- fetch;
267
- max;
268
- ramp;
269
- open = [];
270
- launched = 0;
271
- taken = 0;
272
- /** The next chunk in key order (null if the source holds none), or the error its read raised. */
273
- async take() {
274
- const width = this.ramp ? Math.min(this.max, 2 ** Math.min(this.taken, 30)) : this.max;
275
- while (this.launched < this.keys.length && this.launched - this.taken < width) {
276
- this.open.push(
277
- this.fetch(this.keys[this.launched++]).then(
278
- (chunk) => ({ chunk }),
279
- (cause) => ({ chunk: null, error: { cause } })
280
- )
281
- );
282
- }
283
- const slot = await this.open.shift();
284
- this.taken += 1;
285
- if (slot.error) throw slot.error.cause;
286
- return slot.chunk;
287
- }
288
- };
289
375
  var SegmentEngine = class {
290
376
  storage;
291
377
  cache;
@@ -296,6 +382,12 @@ var SegmentEngine = class {
296
382
  metricsOn;
297
383
  /** Resolved per-op budget (null = disabled); undefined deps ⇒ the generous default. */
298
384
  budget;
385
+ /**
386
+ * The chunk reads open right now, by cache key: a caller that misses the cache and finds its key here awaits that
387
+ * read instead of making its own. An entry lives from the request until it settles, so the map holds at most one
388
+ * promise per distinct key in flight.
389
+ */
390
+ openReads = /* @__PURE__ */ new Map();
299
391
  constructor(deps) {
300
392
  this.storage = deps.storage;
301
393
  this.cache = deps.cache;
@@ -313,15 +405,19 @@ var SegmentEngine = class {
313
405
  return chunk ? chunk.has(remainder) : false;
314
406
  }
315
407
  /**
316
- * Cardinality. When the Storage source can serve per-chunk cardinality from its `.crbm` index (the `.crbm`
317
- * source can), the count is summed from the index with **zero payload reads**. A source without that
318
- * capability (the in-memory source) falls back to fetching every chunk.
408
+ * Cardinality. When the Storage source can describe the current generation (the `.crbm` source can), the count is
409
+ * that description's: the registry row's summary when the row has one it can use, which is one request when cold
410
+ * and no read of the object, and otherwise the sum of the object's index. Either way, **zero payload reads**. A
411
+ * source that can only serve per-chunk cardinality sums those, and one with neither (the in-memory source) falls back
412
+ * to fetching every chunk.
319
413
  *
320
- * The index sum is trusted, not confirmed: the reader checked the index for internal consistency when it opened
414
+ * The row's summary and the index sum are trusted, not confirmed: the source holds the summary against the object
415
+ * whenever it opens it, and the reader checked the index for internal consistency when it opened
321
416
  * the object, and a corrupt index that is still internally consistent yields a wrong count. `iterate()` and the
322
417
  * combines decode the payloads.
323
418
  */
324
419
  async count(seg) {
420
+ if (this.storage.summary) return (await this.storage.summary(seg))?.cardinality ?? 0;
325
421
  const cardinalities = this.storage.cardinalities ? await this.storage.cardinalities(seg) : null;
326
422
  if (cardinalities) {
327
423
  let total2 = 0;
@@ -339,6 +435,17 @@ var SegmentEngine = class {
339
435
  for (let i = 0; i < chunkKeys.length; i++) total += (await window.take())?.size ?? 0;
340
436
  return total;
341
437
  }
438
+ /**
439
+ * What the segment's current generation is: its number, its id count and its metadata, from one resolution, so the
440
+ * three cannot straddle a publish. A source with no summary answers from `count()` and the generation it resolves.
441
+ */
442
+ async stat(seg) {
443
+ if (this.storage.summary) {
444
+ return await this.storage.summary(seg) ?? { generation: null, cardinality: 0 };
445
+ }
446
+ const generation = this.storage.currentGeneration ? await this.storage.currentGeneration(seg) : null;
447
+ return { generation, cardinality: await this.count(seg) };
448
+ }
342
449
  /** Whether the Storage source can measure segment size (for grounded cost); false ⇒ storage isn't grounded. */
343
450
  get supportsStorageSize() {
344
451
  return typeof this.storage.sizeOf === "function";
@@ -361,7 +468,7 @@ var SegmentEngine = class {
361
468
  );
362
469
  }
363
470
  /**
364
- * Every id, ascending, reading ahead through a window of up to 8 chunk fetches that opens 1, 2, 4, 8 wide, so a
471
+ * Every id, ascending, reading ahead through a window of up to 32 chunk fetches that opens 1, 2, 4 … 32 wide, so a
365
472
  * read that stops early has fetched only a handful of chunks past the one it stopped in; with `range`, only the ids in `(after, through]`, fetching only the
366
473
  * chunks the range overlaps (see {@link IdRange}).
367
474
  *
@@ -520,10 +627,17 @@ var SegmentEngine = class {
520
627
  );
521
628
  const inFlight = [];
522
629
  let next = 0;
523
- while (next < common.length && inFlight.length < limit) inFlight.push(startAt(common[next++]));
630
+ let taken = 0;
631
+ const fill = () => {
632
+ const width = Math.min(limit, Math.max(COMBINE_WINDOW_START, 2 ** Math.min(taken, 30)));
633
+ while (next < common.length && inFlight.length < width)
634
+ inFlight.push(startAt(common[next++]));
635
+ };
636
+ fill();
524
637
  while (inFlight.length > 0) {
525
638
  const slot = await inFlight.shift();
526
- if (next < common.length) inFlight.push(startAt(common[next++]));
639
+ taken += 1;
640
+ fill();
527
641
  if (slot.error !== void 0) throw slot.error;
528
642
  if (slot.result && !slot.result.isEmpty) {
529
643
  if (w !== null && isEdge(slot.key, w)) {
@@ -546,9 +660,13 @@ var SegmentEngine = class {
546
660
  async combineChunk(operands, excludes, mode, chunkKey) {
547
661
  const present = mode === "all" ? operands : operands.filter((o) => o.keys.has(chunkKey));
548
662
  if (present.length === 0) return null;
549
- const chunks = await Promise.all(
550
- present.map((op) => this.storageChunk({ ...op.seg, chunkKey }, op.gen))
663
+ const relevant = excludes.filter((e) => e.keys.has(chunkKey));
664
+ const together = relevant.length > 0 && (mode === "any" || present.length === 1);
665
+ const reads = together ? [...present, ...relevant] : present;
666
+ const fetched = await Promise.all(
667
+ reads.map((op) => this.storageChunk({ ...op.seg, chunkKey }, op.gen))
551
668
  );
669
+ const chunks = together ? fetched.slice(0, present.length) : fetched;
552
670
  if (mode === "all" && chunks.some((c) => c === null)) return null;
553
671
  const first = chunks[0];
554
672
  const acc = first ? first.clone() : this.codec.empty();
@@ -563,11 +681,8 @@ var SegmentEngine = class {
563
681
  }
564
682
  }
565
683
  if (acc.isEmpty) return null;
566
- const relevant = excludes.filter((e) => e.keys.has(chunkKey));
567
684
  if (relevant.length > 0) {
568
- const cuts = await Promise.all(
569
- relevant.map((e) => this.storageChunk({ ...e.seg, chunkKey }, e.gen))
570
- );
685
+ const cuts = together ? fetched.slice(present.length) : await Promise.all(relevant.map((e) => this.storageChunk({ ...e.seg, chunkKey }, e.gen)));
571
686
  for (const cut of cuts) {
572
687
  if (cut === null) continue;
573
688
  acc.andNotInPlace(cut);
@@ -671,6 +786,7 @@ var SegmentEngine = class {
671
786
  invalidate(ref) {
672
787
  const prefix = segmentPrefix(ref);
673
788
  this.cache?.deleteWhere((key) => key.startsWith(prefix));
789
+ for (const key of this.openReads.keys()) if (key.startsWith(prefix)) this.openReads.delete(key);
674
790
  this.storage.invalidate?.(ref);
675
791
  }
676
792
  /** The segment's current generation, resolved once per op (`undefined` ⇒ source can't report it). */
@@ -700,7 +816,13 @@ var SegmentEngine = class {
700
816
  * `gen === undefined` ⇒ the source can't report a generation (it only ever serves one) ⇒ the key stays
701
817
  * generation-free. Superseded-generation entries age out under the LRU ceiling — no active purge.
702
818
  *
703
- * The returned instance is **shared** (it may be the cached one): callers read it or clone it, never mutate it.
819
+ * Concurrent misses of one key share **one** request, with or without a cache: the first caller starts the read,
820
+ * later ones await it, and the entry goes when the read settles, so a failure is every waiting caller's and the
821
+ * next call reads again. The key carries the version exactly as the cache's does, so two versions of one chunk
822
+ * are two reads, never one.
823
+ *
824
+ * The returned instance is **shared** (it may be the cached one, and every caller that joined a read gets the same
825
+ * one): callers read it or clone it, never mutate it.
704
826
  */
705
827
  async storageChunk(ref, gen) {
706
828
  if (gen === null) return null;
@@ -713,6 +835,19 @@ var SegmentEngine = class {
713
835
  }
714
836
  if (this.metricsOn) this.metrics.onEvent({ kind: "cache", hit: false });
715
837
  }
838
+ const open4 = this.openReads.get(cacheKey);
839
+ if (open4) return open4.read;
840
+ const token = {};
841
+ const entry = { token, read: this.fetchChunk(ref, cacheKey, token) };
842
+ this.openReads.set(cacheKey, entry);
843
+ const settled = () => {
844
+ if (this.openReads.get(cacheKey) === entry) this.openReads.delete(cacheKey);
845
+ };
846
+ entry.read.then(settled, settled);
847
+ return entry.read;
848
+ }
849
+ /** One request for one chunk, decoded, range-checked and cached. */
850
+ async fetchChunk(ref, cacheKey, token) {
716
851
  const startedAt = this.metricsOn ? this.clock.now() : 0;
717
852
  const bytes = await this.storage.getChunk(ref);
718
853
  if (this.metricsOn) {
@@ -727,7 +862,7 @@ var SegmentEngine = class {
727
862
  if (!bytes) return null;
728
863
  const bitmap = this.codec.safeDeserialize(bytes, this.maxBitmapBytes);
729
864
  this.assertChunkPayloadInRange(bitmap, ref.chunkKey);
730
- this.cache?.set(cacheKey, bitmap);
865
+ if (this.openReads.get(cacheKey)?.token === token) this.cache?.set(cacheKey, bitmap);
731
866
  return bitmap;
732
867
  }
733
868
  };
@@ -871,6 +1006,111 @@ var BoundedLru = class {
871
1006
  }
872
1007
  };
873
1008
 
1009
+ // src/core/generation-gc.ts
1010
+ async function nextGeneration(ref, deps) {
1011
+ return aboveEverything(ref, deps, await deps.registry.get(ref));
1012
+ }
1013
+ async function aboveEverything(ref, deps, record) {
1014
+ let highest = record?.currentGen ?? -1;
1015
+ for await (const key of deps.storage.list(ref)) {
1016
+ if (key.generation > highest) highest = key.generation;
1017
+ }
1018
+ return highest + 1;
1019
+ }
1020
+ async function nextLoadGeneration(ref, deps, row) {
1021
+ const generation = (row?.currentGen ?? -1) + 1;
1022
+ try {
1023
+ await deps.storage.getTail({ namespace: ref.namespace, segment: ref.segment, generation }, 0);
1024
+ } catch (err) {
1025
+ if (isNotFoundError(err)) return { generation, checked: true };
1026
+ }
1027
+ return { generation: await aboveEverything(ref, deps, row), checked: false };
1028
+ }
1029
+ var LIST_COLLECTION_CADENCE = 16;
1030
+ async function gcOrphanGenerations(ref, deps, options = {}) {
1031
+ const keep = options.keep ?? 1;
1032
+ if (!Number.isInteger(keep) || keep < 0) {
1033
+ throw new ValidationError(`keep must be a non-negative integer; got ${String(keep)}`);
1034
+ }
1035
+ const record = await deps.registry.get(ref);
1036
+ if (record === null) return [];
1037
+ const current = record.currentGen;
1038
+ const seen = /* @__PURE__ */ new Set();
1039
+ for await (const key of deps.storage.list(ref)) seen.add(key.generation);
1040
+ const gens = [...seen];
1041
+ const after = await deps.registry.get(ref);
1042
+ if (after === null)
1043
+ throw new WriteConflictError(`registry row for segment ${ref.segment} was purged mid-pass`);
1044
+ if (record.status === "destroyed" && after.token !== record.token) {
1045
+ throw new WriteConflictError(
1046
+ `segment ${ref.segment} changed incarnation while its generations were being listed`
1047
+ );
1048
+ }
1049
+ const cutoff = current === null || after.currentGen === null ? null : Math.min(current, after.currentGen);
1050
+ const stillCollectable = async () => {
1051
+ const still = await deps.registry.get(ref);
1052
+ if (still === null) {
1053
+ throw new WriteConflictError(`registry row for segment ${ref.segment} was purged mid-pass`);
1054
+ }
1055
+ const ok = record.status === "destroyed" ? still.token === after.token : cutoff === null || still.currentGen !== null && still.currentGen >= cutoff;
1056
+ if (!ok) {
1057
+ throw new WriteConflictError(
1058
+ `segment ${ref.segment} changed incarnation while its generations were being collected`
1059
+ );
1060
+ }
1061
+ };
1062
+ const toDelete = record.status === "destroyed" ? gens.sort((a, b) => a - b) : cutoff === null ? (
1063
+ // No Storage pointer yet, so "below current" selects nothing and there is nothing safe to infer: an object
1064
+ // here is either a load about to publish or an orphan we cannot tell apart from it. Deleting would race
1065
+ // that publish into a dangling pointer. It is collected once a pointer exists.
1066
+ []
1067
+ ) : (
1068
+ // Delete generations below the cutoff, except the newest `keep` of them (the grace window).
1069
+ gens.filter((g) => g < cutoff).sort((a, b) => b - a).slice(keep)
1070
+ );
1071
+ for (const generation of toDelete) {
1072
+ await stillCollectable();
1073
+ await deps.storage.delete({ namespace: ref.namespace, segment: ref.segment, generation });
1074
+ }
1075
+ return toDelete;
1076
+ }
1077
+ async function collectAfterLoad(ref, deps, options) {
1078
+ const { generation, keep, byName, proveCurrent } = options;
1079
+ if (keep >= generation) return [];
1080
+ const periodic = generation % LIST_COLLECTION_CADENCE === 0;
1081
+ if (byName && keep <= 1 && !periodic) {
1082
+ if (keep === 1 && proveCurrent !== void 0 && !await objectIsThere(deps.storage, proveCurrent)) {
1083
+ return gcOrphanGenerations(ref, deps, { keep });
1084
+ }
1085
+ return collectByName(ref, deps, { generation, keep });
1086
+ }
1087
+ return gcOrphanGenerations(ref, deps, { keep });
1088
+ }
1089
+ async function objectIsThere(storage, key) {
1090
+ try {
1091
+ await storage.getTail(key, 0);
1092
+ return true;
1093
+ } catch {
1094
+ return false;
1095
+ }
1096
+ }
1097
+ async function collectByName(ref, deps, options) {
1098
+ const { generation, keep } = options;
1099
+ if (!Number.isInteger(keep) || keep < 0 || keep > 1) {
1100
+ throw new ValidationError(`collecting by name needs a keep of 0 or 1; got ${String(keep)}`);
1101
+ }
1102
+ const evicted = generation - keep - 1;
1103
+ if (evicted < 0) return [];
1104
+ const still = await deps.registry.get(ref);
1105
+ if (still === null || still.currentGen === null || still.currentGen < generation) return [];
1106
+ await deps.storage.delete({
1107
+ namespace: ref.namespace,
1108
+ segment: ref.segment,
1109
+ generation: evicted
1110
+ });
1111
+ return [evicted];
1112
+ }
1113
+
874
1114
  // src/core/reader-defaults.ts
875
1115
  var DEFAULT_CURRENT_GEN_TTL_MS = 2e3;
876
1116
  var REFRESH_RETRY_MS = 500;
@@ -918,9 +1158,20 @@ var AWS_US_EAST_1_ONDEMAND = deepFreeze({
918
1158
  var HOURS_PER_MONTH = 730;
919
1159
  var SECONDS_PER_MONTH = HOURS_PER_MONTH * 3600;
920
1160
  var GIB = 1024 ** 3;
921
- var SIZED_READS_PER_COLD_OPERAND = 2;
922
- var STORE_LOAD_PUT_CLASS = 3;
923
- var STORE_LOAD_SIZED_READS = 9;
1161
+ var STORE_LOAD_PUT_CLASS = 1 + 1 / LIST_COLLECTION_CADENCE;
1162
+ var STORE_LOAD_POINTER_READS = 3 + 2 / LIST_COLLECTION_CADENCE;
1163
+ var STORE_LOAD_EXISTENCE_CHECKS = 2 - 1 / LIST_COLLECTION_CADENCE;
1164
+ var RETENTION_SWEEP_REQUESTS = deepFreeze({
1165
+ conditionalDelete: {
1166
+ retirement: { reads: 9, writes: 3, deletes: 1 },
1167
+ purge: { reads: 4, writes: 0, deletes: 2 }
1168
+ },
1169
+ tombstoning: {
1170
+ retirement: { reads: 8, writes: 3, deletes: 0 },
1171
+ purge: { reads: 3, writes: 1, deletes: 0 }
1172
+ }
1173
+ });
1174
+ var shown = (n) => String(Number(n.toFixed(2)));
924
1175
  function requireFiniteNonNeg(n, field) {
925
1176
  if (n === void 0 || !Number.isFinite(n) || n < 0) {
926
1177
  throw new ValidationError(`${field} must be a finite number >= 0; got ${n}`);
@@ -1057,8 +1308,8 @@ function bytesWords(bytes) {
1057
1308
  const units = ["bytes", "KiB", "MiB", "GiB", "TiB", "PiB"];
1058
1309
  let value = bytes;
1059
1310
  let unit = 0;
1060
- const shown = (v, u) => u === 0 ? Math.round(v) : Number(v.toFixed(2));
1061
- while (unit < units.length - 1 && shown(value, unit) >= 1024) {
1311
+ const shown2 = (v, u) => u === 0 ? Math.round(v) : Number(v.toFixed(2));
1312
+ while (unit < units.length - 1 && shown2(value, unit) >= 1024) {
1062
1313
  value /= 1024;
1063
1314
  unit += 1;
1064
1315
  }
@@ -1085,6 +1336,10 @@ function buildReport(input) {
1085
1336
  requireFiniteNonNeg(storage.getPerMillion, "pricing.storage.getPerMillion");
1086
1337
  requireFiniteNonNeg(storage.putPerMillion, "pricing.storage.putPerMillion");
1087
1338
  requireFiniteNonNeg(storage.storagePerGiBMonth, "pricing.storage.storagePerGiBMonth");
1339
+ const pointerRead = requireFiniteNonNeg(
1340
+ storage.requestsPerPointerRead ?? 1,
1341
+ "pricing.storage.requestsPerPointerRead"
1342
+ );
1088
1343
  const sizedRead = requireFiniteNonNeg(
1089
1344
  storage.requestsPerSizedRead ?? 1,
1090
1345
  "pricing.storage.requestsPerSizedRead"
@@ -1131,6 +1386,15 @@ function buildReport(input) {
1131
1386
  input.workload.genTtlMs ?? DEFAULT_CURRENT_GEN_TTL_MS,
1132
1387
  "genTtlMs"
1133
1388
  );
1389
+ const retirementsPerMonth = requireFiniteNonNeg(
1390
+ input.workload.retirementsPerMonth ?? 0,
1391
+ "retirementsPerMonth"
1392
+ );
1393
+ const purgesPerMonth = requireFiniteNonNeg(input.workload.purgesPerMonth ?? 0, "purgesPerMonth");
1394
+ const conditionalDelete = input.workload.conditionalDelete ?? true;
1395
+ if (typeof conditionalDelete !== "boolean") {
1396
+ throw new ValidationError(`conditionalDelete must be a boolean; got ${conditionalDelete}`);
1397
+ }
1134
1398
  const storageGetUSD = storage.getPerMillion / 1e6;
1135
1399
  const putUSD = storage.putPerMillion / 1e6;
1136
1400
  const missFraction = 1 - cacheHitRate;
@@ -1139,16 +1403,26 @@ function buildReport(input) {
1139
1403
  const refreshes = genTtlMs > 0 ? Math.min(readerProcesses * hotSegments * S * 1e3 / genTtlMs, readsPerSec * S) : 0;
1140
1404
  const storageUSD = storedGiB * storage.storagePerGiBMonth;
1141
1405
  const readsUSD = readMisses * storageGetUSD;
1142
- const intersectGets = chunksPerIntersect + SIZED_READS_PER_COLD_OPERAND * operandsPerIntersect * sizedRead;
1406
+ const getsPerColdOperand = pointerRead + sizedRead;
1407
+ const intersectGets = chunksPerIntersect + getsPerColdOperand * operandsPerIntersect;
1143
1408
  const intersectsUSD = intersects * intersectGets * storageGetUSD;
1144
- const loadGets = STORE_LOAD_SIZED_READS * sizedRead;
1409
+ const loadGets = STORE_LOAD_POINTER_READS * pointerRead + STORE_LOAD_EXISTENCE_CHECKS;
1145
1410
  const loadsUSD = loadsPerMonth * ((requestsPerLoad + STORE_LOAD_PUT_CLASS) * putUSD + loadGets * storageGetUSD);
1146
- const refreshUSD = refreshes * sizedRead * storageGetUSD;
1147
- const total = readsUSD + intersectsUSD + storageUSD + loadsUSD + refreshUSD;
1411
+ const refreshUSD = refreshes * pointerRead * storageGetUSD;
1412
+ const sweep = RETENTION_SWEEP_REQUESTS[conditionalDelete ? "conditionalDelete" : "tombstoning"];
1413
+ const retentionUSD = retirementsPerMonth * (sweep.retirement.reads * storageGetUSD + sweep.retirement.writes * putUSD) + purgesPerMonth * (sweep.purge.reads * storageGetUSD + sweep.purge.writes * putUSD);
1414
+ const total = readsUSD + intersectsUSD + storageUSD + loadsUSD + refreshUSD + retentionUSD;
1148
1415
  const headroom = Math.max(0, baselineUSD - storageUSD - refreshUSD);
1149
1416
  const readsCross = storageGetUSD > 0 && missFraction > 0 ? headroom / (S * storageGetUSD * missFraction) : Infinity;
1150
1417
  const verdict = total <= baselineUSD * 0.1 ? "win-big" : total < baselineUSD ? "win" : "lose-zone";
1151
- const dominant = Math.max(readsUSD, intersectsUSD, storageUSD, loadsUSD, refreshUSD);
1418
+ const dominant = Math.max(
1419
+ readsUSD,
1420
+ intersectsUSD,
1421
+ storageUSD,
1422
+ loadsUSD,
1423
+ refreshUSD,
1424
+ retentionUSD
1425
+ );
1152
1426
  let driver = "storage";
1153
1427
  if (dominant === readsUSD && readsUSD > 0) driver = "point reads (object GETs)";
1154
1428
  else if (dominant === intersectsUSD && intersectsUSD > 0) {
@@ -1156,6 +1430,8 @@ function buildReport(input) {
1156
1430
  } else if (dominant === loadsUSD && loadsUSD > 0)
1157
1431
  driver = "loads (object, listing and pointer requests)";
1158
1432
  else if (dominant === refreshUSD && refreshUSD > 0) driver = "the pointer refresh (object GETs)";
1433
+ else if (dominant === retentionUSD && retentionUSD > 0)
1434
+ driver = "the retention sweep (registry reads and writes)";
1159
1435
  const sized = baseline.sized;
1160
1436
  const noBytes = Math.round(storageBytes) === 0;
1161
1437
  const redisWords = sized === void 0 ? void 0 : noBytes ? `$${baselineUSD.toFixed(2)}/mo cheapest Redis cluster in the catalogue, as no bytes are counted (${clusterWords(sized.cluster)})` : `$${baselineUSD.toFixed(2)}/mo Redis that would hold ${bytesWords(storageBytes)} (${clusterWords(sized.cluster)})`;
@@ -1165,12 +1441,15 @@ function buildReport(input) {
1165
1441
  const notes = [
1166
1442
  "Same-region egress treated as free; internet egress is not modeled.",
1167
1443
  "Request cost is from the supplied workload rates (live-metrics-derived request cost is a later phase).",
1168
- loadsPerMonth > 0 ? `Loads modeled: ${loadsPerMonth}/mo, each ${requestsPerLoad} PUT-class request(s) for the object plus ${STORE_LOAD_PUT_CLASS} PUT-class and ${loadGets} GETs that store.load() adds (listings, pointer, index).` : "Loads are NOT modeled \u2014 set workload.loadsPerMonth (+ requestsPerLoad for multipart) to include them.",
1444
+ loadsPerMonth > 0 ? `Loads modeled: ${loadsPerMonth}/mo, each ${requestsPerLoad} PUT-class request(s) for the object plus ${shown(STORE_LOAD_PUT_CLASS)} PUT-class and ${shown(loadGets)} GETs that store.load() adds (the pointer, a check that the next generation number is free, a check that the current generation is there, and a listing every ${LIST_COLLECTION_CADENCE}th load).` : "Loads are NOT modeled \u2014 set workload.loadsPerMonth (+ requestsPerLoad for multipart) to include them.",
1169
1445
  ...intersects > 0 ? [
1170
- `Intersections priced cold: ${intersectGets} GETs each, ${SIZED_READS_PER_COLD_OPERAND * sizedRead} for each of ${operandsPerIntersect} operand(s) plus ${chunksPerIntersect} chunk read(s); cacheHitRate does not apply.`
1446
+ `Intersections priced cold: ${intersectGets} GETs each, ${getsPerColdOperand} for each of ${operandsPerIntersect} operand(s) plus ${chunksPerIntersect} chunk read(s); cacheHitRate does not apply.`
1171
1447
  ] : [],
1172
1448
  hotSegments > 0 ? genTtlMs > 0 ? `Pointer refresh modeled: ${hotSegments} hot segment(s) in each of ${readerProcesses} reader process(es), each re-reading its pointer at most every ${genTtlMs} ms, and at most once a point read.` : "Pointer refresh: none \u2014 genTtlMs 0 turns the timed refresh off. A reader still re-reads a pointer on an eviction, a swept generation or an invalidation, which this does not price." : readsPerSec > 0 ? "The pointer refresh is NOT modeled \u2014 set workload.hotSegments to the segments each long-lived reader keeps reading, and workload.readerProcesses to how many readers there are." : null,
1173
1449
  hotSegments > DEFAULT_MAX_OPEN_SEGMENTS ? `More hot segments in a reader than a store keeps open by default (${DEFAULT_MAX_OPEN_SEGMENTS}): a read of a segment the reader evicted opens it again, a pointer and a tail read, which this does not price. Raise cache.readerMax, and cache.readerMaxBytes, to keep them open.` : null,
1450
+ ...retirementsPerMonth > 0 || purgesPerMonth > 0 ? [
1451
+ `Retention sweep modeled: ${retirementsPerMonth}/mo retirements at ${sweep.retirement.reads} reads, ${sweep.retirement.writes} writes and ${sweep.retirement.deletes} delete(s) each, ${purgesPerMonth}/mo purges at ${sweep.purge.reads} reads, ${sweep.purge.writes} writes and ${sweep.purge.deletes} delete(s) each; reads are priced at the GET rate and writes at the PUT rate, deletes at nothing, as S3 bills none` + (conditionalDelete ? "." : "; a registry that only tombstones is read again, two reads per purged segment, by every later full sweep, which is not priced.")
1452
+ ] : [],
1174
1453
  ...input.extraNotes ?? [],
1175
1454
  // Last, so the notes a report already carried keep their places.
1176
1455
  redisNote
@@ -1182,7 +1461,8 @@ function buildReport(input) {
1182
1461
  intersects: intersectsUSD,
1183
1462
  storage: storageUSD,
1184
1463
  loads: loadsUSD,
1185
- pointerRefresh: refreshUSD
1464
+ pointerRefresh: refreshUSD,
1465
+ retention: retentionUSD
1186
1466
  },
1187
1467
  total
1188
1468
  },
@@ -1291,7 +1571,10 @@ async function drainRegistry(registry, options) {
1291
1571
  const advice = options.narrowable === false ? `${raise[0].toUpperCase()}${raise.slice(1)} if the namespace really is that large and the memory is ` : `Narrow it with \`namespace\`, or ${raise} if the fleet really is that large and the memory is `;
1292
1572
  const rows = [];
1293
1573
  for await (const rec of registry.list(options.namespace)) {
1294
- if (options.namespace === void 0 && isReservedRow(rec)) continue;
1574
+ if (options.namespace === void 0 && isReservedRow(rec)) {
1575
+ options.onReserved?.(rec);
1576
+ continue;
1577
+ }
1295
1578
  if (rows.length >= maxScanSegments) {
1296
1579
  throw new BudgetExceededError(
1297
1580
  `${op} would enumerate more than ${maxScanSegments} segments \u2014 the scan was abandoned there rather than completed. ${advice}available (a record is a few hundred bytes resident).`
@@ -1463,16 +1746,19 @@ var BufferReader = class {
1463
1746
  var MemoryRegistryDriver = class {
1464
1747
  rows = /* @__PURE__ */ new Map();
1465
1748
  now;
1749
+ entropy;
1466
1750
  seq = 0;
1467
1751
  constructor(options = {}) {
1468
1752
  this.now = options.now ?? (() => Date.now());
1753
+ this.entropy = options.entropy ?? webCryptoEntropy;
1469
1754
  }
1470
- nextToken() {
1755
+ nextToken(incarnation) {
1471
1756
  this.seq += 1;
1472
- return String(this.seq);
1757
+ return `${incarnation}.${this.seq}.${drawWrite(this.entropy)}`;
1473
1758
  }
1759
+ /** `conditionalDelete`: a delete checks the token and removes the row in one step, with no await between. */
1474
1760
  capabilities() {
1475
- return { strongRead: true };
1761
+ return entropyIsAvailable(this.entropy) ? { strongRead: true, conditionalDelete: true } : { strongRead: true, canWrite: false, conditionalDelete: true };
1476
1762
  }
1477
1763
  async get(ref) {
1478
1764
  validateSegmentRef(ref);
@@ -1481,25 +1767,25 @@ var MemoryRegistryDriver = class {
1481
1767
  }
1482
1768
  async create(ref, record) {
1483
1769
  validateSegmentRef(ref);
1484
- validateNewRegistryRecord(record);
1770
+ const checked = validateNewRegistryRecord(record);
1485
1771
  const key = segmentKey(ref);
1486
1772
  if (this.rows.has(key)) {
1487
1773
  throw new WriteConflictError(`registry row already exists for segment ${ref.segment}`);
1488
1774
  }
1489
- const token = this.nextToken();
1490
- this.rows.set(key, structuredClone(recordFromNew(ref, record, this.now(), token)));
1775
+ const token = this.nextToken(drawIncarnation(this.entropy));
1776
+ this.rows.set(key, structuredClone(recordFromNew(ref, checked, this.now(), token)));
1491
1777
  return { token };
1492
1778
  }
1493
1779
  async compareAndSwap(ref, expected, patch) {
1494
1780
  validateSegmentRef(ref);
1495
- validateRegistryPatch(patch);
1781
+ const checked = validateRegistryPatch(patch);
1496
1782
  const key = segmentKey(ref);
1497
1783
  const existing = this.rows.get(key);
1498
1784
  if (!existing || existing.token !== expected) {
1499
1785
  throw new WriteConflictError(`OCC token mismatch for registry row ${ref.segment}`);
1500
1786
  }
1501
- const token = this.nextToken();
1502
- this.rows.set(key, structuredClone(applyRegistryPatch(existing, patch, this.now(), token)));
1787
+ const token = this.nextToken(incarnationOf(existing.token) ?? drawIncarnation(this.entropy));
1788
+ this.rows.set(key, structuredClone(applyRegistryPatch(existing, checked, this.now(), token)));
1503
1789
  return { token };
1504
1790
  }
1505
1791
  async *list(namespace) {
@@ -1816,43 +2102,48 @@ var LocalFsRegistryDriver = class {
1816
2102
  constructor(root, options = {}) {
1817
2103
  this.root = root;
1818
2104
  this.now = options.now ?? (() => Date.now());
2105
+ this.entropy = options.entropy ?? webCryptoEntropy;
1819
2106
  }
1820
2107
  root;
1821
2108
  now;
2109
+ entropy;
2110
+ /**
2111
+ * `conditionalDelete`: a delete reads the row, checks it, and unlinks it under the row's lock, which every writer on
2112
+ * the root in this process takes, so no write can land between the check and the unlink.
2113
+ */
1822
2114
  capabilities() {
1823
- return { strongRead: true };
2115
+ return entropyIsAvailable(this.entropy) ? { strongRead: true, conditionalDelete: true } : { strongRead: true, canWrite: false, conditionalDelete: true };
1824
2116
  }
1825
2117
  async get(ref) {
1826
2118
  const env = await this.readRow(registryRowPath(this.root, ref));
1827
2119
  return env && !env.deleted ? env.record : null;
1828
2120
  }
1829
2121
  async create(ref, record) {
1830
- validateNewRegistryRecord(record);
2122
+ const checked = validateNewRegistryRecord(record);
1831
2123
  const path = registryRowPath(this.root, ref);
1832
2124
  return this.withRowLock(path, async () => {
1833
2125
  const current = await this.readRow(path);
1834
2126
  if (current !== null && !current.deleted) {
1835
2127
  throw new WriteConflictError(`registry row already exists for segment ${ref.segment}`);
1836
2128
  }
1837
- const counter = current ? registryCounterOf(current.record) + 1 : 0;
1838
- const token = String(counter);
1839
- await this.writeRow(path, false, recordFromNew(ref, record, this.now(), token));
2129
+ const token = newIncarnationToken(this.entropy, current?.record);
2130
+ await this.writeRow(path, false, recordFromNew(ref, checked, this.now(), token));
1840
2131
  return { token };
1841
2132
  });
1842
2133
  }
1843
2134
  async compareAndSwap(ref, expected, patch) {
1844
- validateRegistryPatch(patch);
2135
+ const checked = validateRegistryPatch(patch);
1845
2136
  const path = registryRowPath(this.root, ref);
1846
2137
  return this.withRowLock(path, async () => {
1847
2138
  const current = await this.readRow(path);
1848
2139
  if (current === null || current.deleted || current.record.token !== expected) {
1849
2140
  throw new WriteConflictError(`OCC token mismatch for registry row ${ref.segment}`);
1850
2141
  }
1851
- const token = String(registryCounterOf(current.record) + 1);
2142
+ const token = nextRegistryToken(current.record, this.entropy);
1852
2143
  await this.writeRow(
1853
2144
  path,
1854
2145
  false,
1855
- applyRegistryPatch(current.record, patch, this.now(), token)
2146
+ applyRegistryPatch(current.record, checked, this.now(), token)
1856
2147
  );
1857
2148
  return { token };
1858
2149
  });
@@ -1886,10 +2177,23 @@ var LocalFsRegistryDriver = class {
1886
2177
  } else if (current === null || current.deleted) {
1887
2178
  return;
1888
2179
  }
1889
- const token = String(registryCounterOf(current.record) + 1);
2180
+ if (incarnationOf(current.record.token) !== void 0) {
2181
+ await this.unlinkRow(path);
2182
+ return;
2183
+ }
2184
+ const token = nextRegistryToken(current.record, this.entropy);
1890
2185
  await this.writeRow(path, true, { ...current.record, token, updatedAt: this.now() });
1891
2186
  });
1892
2187
  }
2188
+ /** Remove a row's file and make the removal durable (fsync the directory). */
2189
+ async unlinkRow(path) {
2190
+ try {
2191
+ await unlink2(path);
2192
+ } catch (err) {
2193
+ if (!isCode(err, "ENOENT")) throw mapFsError(err);
2194
+ }
2195
+ await fsyncDir(dirname2(path));
2196
+ }
1893
2197
  /** Namespaces to scan: just the one requested, or every namespace dir under the root. */
1894
2198
  async namespaceDirs(namespace) {
1895
2199
  if (namespace !== void 0) return [namespace];
@@ -2045,6 +2349,8 @@ function yieldEvery(clock, everyN = YIELD_EVERY) {
2045
2349
  var AAD_VERSION = 1;
2046
2350
  var SCOPE_CHUNK = 0;
2047
2351
  var SCOPE_INDEX = 1;
2352
+ var SCOPE_METADATA = 2;
2353
+ var SCOPE_SUMMARY = 3;
2048
2354
  function aadFor(ref, generation, scope) {
2049
2355
  const enc = new TextEncoder();
2050
2356
  const ns = enc.encode(ref.namespace ?? "");
@@ -2064,9 +2370,9 @@ function aadFor(ref, generation, scope) {
2064
2370
  o += seg.length;
2065
2371
  view.setBigUint64(o, BigInt(generation), true);
2066
2372
  o += 8;
2067
- out[o] = scope === "index" ? SCOPE_INDEX : SCOPE_CHUNK;
2373
+ out[o] = scope === "index" ? SCOPE_INDEX : scope === "metadata" ? SCOPE_METADATA : scope === "summary" ? SCOPE_SUMMARY : SCOPE_CHUNK;
2068
2374
  o += 1;
2069
- view.setUint32(o, scope === "index" ? 0 : scope, true);
2375
+ view.setUint32(o, typeof scope === "number" ? scope : 0, true);
2070
2376
  return out;
2071
2377
  }
2072
2378
 
@@ -2148,71 +2454,9 @@ function readVarint(bytes, offset) {
2148
2454
  throw new IntegrityError("varint exceeds u32 range");
2149
2455
  }
2150
2456
 
2151
- // src/core/crbm/format.ts
2152
- var MAGIC = Uint8Array.of(67, 82, 66, 77);
2153
- var VERSION_MAJOR = 1;
2154
- var VERSION_MINOR = 0;
2155
- var PREAMBLE_BYTES = 8;
2156
- var PAYLOAD_START = PREAMBLE_BYTES;
2157
- var FOOTER_BYTES = 104;
2158
- var FOOTER = {
2159
- indexOffset: 0,
2160
- // u64
2161
- indexLength: 8,
2162
- // u64
2163
- indexCrc32c: 16,
2164
- // u32
2165
- flags: 20,
2166
- // u32
2167
- payloadCodecId: 24,
2168
- // u16 — which codec wrote the chunk payloads
2169
- elementWidth: 26,
2170
- // u8
2171
- containerCodec: 27,
2172
- // u8
2173
- versionMajor: 28,
2174
- // u8
2175
- versionMinor: 29,
2176
- // u8
2177
- reserved2: 30,
2178
- // u16
2179
- generation: 32,
2180
- // u64
2181
- indexNonce: 40,
2182
- // 12 B
2183
- indexTag: 52,
2184
- // 16 B
2185
- keyId: 68,
2186
- // 16 B
2187
- chunkCount: 84,
2188
- // u32
2189
- totalCardinality: 88,
2190
- // u64
2191
- footerCrc32c: 96,
2192
- // u32 — covers footer bytes [0, 96)
2193
- endMagic: 100
2194
- // char[4]
2195
- };
2196
- var FOOTER_CRC_COVERAGE = FOOTER.footerCrc32c;
2197
- var FLAG_ENCRYPTED = 1 << 0;
2198
- var FLAG_INDEX_COMPRESSED = 1 << 1;
2199
- var FLAG_LITTLE_ENDIAN = 1 << 2;
2200
- var KNOWN_FLAGS = FLAG_ENCRYPTED | FLAG_LITTLE_ENDIAN;
2201
- var ELEMENT_WIDTH_32 = 32;
2202
- var PAYLOAD_CODEC_ROARING_PORTABLE = 1;
2203
- var KNOWN_PAYLOAD_CODEC_IDS = /* @__PURE__ */ new Set([
2204
- PAYLOAD_CODEC_ROARING_PORTABLE
2205
- ]);
2206
- var CONTAINER_CODEC_NONE = 0;
2207
- var DEFAULT_TAIL_BYTES = 256 * 1024;
2208
- var DEFAULT_MAX_INDEX_BYTES = 8 * 1024 * 1024;
2209
- var DEFAULT_MAX_PAYLOAD_BYTES = 16 * 1024 * 1024;
2210
- var MAX_CHUNK_CARDINALITY = 65536;
2211
- var CRC32C_BYTES = 4;
2212
- var AEAD_NONCE_BYTES = 12;
2213
- var AEAD_TAG_BYTES = 16;
2214
-
2215
2457
  // src/core/crbm/reader.ts
2458
+ var RETAINED_BYTES_PER_METADATA_BYTE = 2;
2459
+ var RETAINED_BYTES_PER_METADATA_KEY = 160;
2216
2460
  var MIN_INDEX_RECORD_BYTES = 4 + CRC32C_BYTES;
2217
2461
  function indexCapacity(indexLength) {
2218
2462
  return Math.min(65536, Math.floor(indexLength / MIN_INDEX_RECORD_BYTES));
@@ -2243,6 +2487,11 @@ function checkedFooter(tail, size) {
2243
2487
  }
2244
2488
  return { footer, fview, storedFooterCrc };
2245
2489
  }
2490
+ async function footerSaysEncrypted(blob) {
2491
+ const { bytes: tail, size } = await blob.getTail(FOOTER_BYTES);
2492
+ const { fview } = checkedFooter(tail, size);
2493
+ return (fview.getUint32(FOOTER.flags, true) & FLAG_ENCRYPTED) !== 0;
2494
+ }
2246
2495
  function readU64(view, offset, field) {
2247
2496
  const big = view.getBigUint64(offset, true);
2248
2497
  if (big > BigInt(Number.MAX_SAFE_INTEGER)) {
@@ -2251,7 +2500,7 @@ function readU64(view, offset, field) {
2251
2500
  return Number(big);
2252
2501
  }
2253
2502
  var CrbmReader = class _CrbmReader {
2254
- constructor(blob, objectSize, generation, lineage, totalCardinality, index, servedFromTail, crypto, footerCrc) {
2503
+ constructor(blob, objectSize, generation, lineage, totalCardinality, index, servedFromTail, crypto, footerCrc, metadata, metadataWeight) {
2255
2504
  this.blob = blob;
2256
2505
  this.objectSize = objectSize;
2257
2506
  this.generation = generation;
@@ -2261,6 +2510,8 @@ var CrbmReader = class _CrbmReader {
2261
2510
  this.servedFromTail = servedFromTail;
2262
2511
  this.crypto = crypto;
2263
2512
  this.footerCrc = footerCrc;
2513
+ this.metadata = metadata;
2514
+ this.metadataWeight = metadataWeight;
2264
2515
  }
2265
2516
  blob;
2266
2517
  objectSize;
@@ -2271,6 +2522,8 @@ var CrbmReader = class _CrbmReader {
2271
2522
  servedFromTail;
2272
2523
  crypto;
2273
2524
  footerCrc;
2525
+ metadata;
2526
+ metadataWeight;
2274
2527
  /**
2275
2528
  * Which object this reader opened, as its size and its footer's CRC. The footer covers the index's CRC, so two
2276
2529
  * objects with the same fingerprint hold the same chunks, as far as a CRC can tell; a pin compares it to know that
@@ -2292,6 +2545,15 @@ var CrbmReader = class _CrbmReader {
2292
2545
  const { keys, cardinalityMinusOne, lengths, crcs, offsets } = this.index;
2293
2546
  return keys.byteLength + cardinalityMinusOne.byteLength + lengths.byteLength + crcs.byteLength + offsets.byteLength;
2294
2547
  }
2548
+ /**
2549
+ * What this reader holds, the weight the storage reader cache bounds on: its parsed index
2550
+ * ({@link retainedIndexBytes}), plus a weight for the metadata it decoded, from its canonical length and its number
2551
+ * of keys. Equal to {@link retainedIndexBytes} for a generation with no metadata. A cache heuristic, not a contract:
2552
+ * how the metadata is weighed may change in any release.
2553
+ */
2554
+ get retainedBytes() {
2555
+ return this.retainedIndexBytes + this.metadataWeight;
2556
+ }
2295
2557
  /** Per-chunk cardinality (`chunkKey → count`) from the parsed index — no payload reads. */
2296
2558
  cardinalities() {
2297
2559
  const { keys, cardinalityMinusOne } = this.index;
@@ -2315,7 +2577,6 @@ var CrbmReader = class _CrbmReader {
2315
2577
  }
2316
2578
  static async open(blob, options = {}) {
2317
2579
  const tailBytes = Math.max(options.tailBytes ?? DEFAULT_TAIL_BYTES, FOOTER_BYTES);
2318
- const maxPayloadBytes = options.maxPayloadBytes ?? DEFAULT_MAX_PAYLOAD_BYTES;
2319
2580
  const maxIndexBytes = options.maxIndexBytes ?? DEFAULT_MAX_INDEX_BYTES;
2320
2581
  const { bytes: tail, size } = await blob.getTail(tailBytes);
2321
2582
  const { footer, fview, storedFooterCrc } = checkedFooter(tail, size);
@@ -2358,6 +2619,13 @@ var CrbmReader = class _CrbmReader {
2358
2619
  const chunkCount = fview.getUint32(FOOTER.chunkCount, true);
2359
2620
  const totalCardinality = readU64(fview, FOOTER.totalCardinality, "total_cardinality");
2360
2621
  const generation = readU64(fview, FOOTER.generation, "generation");
2622
+ if (!encrypted && options.crypto !== void 0) {
2623
+ throw new IntegrityError(
2624
+ `.crbm generation ${generation} is not encrypted, but it was opened with a key: a cleartext object where an encrypted one belongs is forged, corrupt, or a write that never published (a crash, before the segment's key was made); a load's collection, dropSegment or deleting the object removes it`
2625
+ );
2626
+ }
2627
+ const versionMinor = footer[FOOTER.versionMinor];
2628
+ const hasExtension = (flags & FLAG_EXTENSION) !== 0;
2361
2629
  if (indexOffset < PAYLOAD_START || indexOffset + indexLength > size - FOOTER_BYTES) {
2362
2630
  throw new IntegrityError(
2363
2631
  `.crbm index region [${indexOffset}, +${indexLength}) out of bounds`
@@ -2368,20 +2636,27 @@ var CrbmReader = class _CrbmReader {
2368
2636
  }
2369
2637
  const tailCoversFrom = size - tail.length;
2370
2638
  if (tailCoversFrom === 0) {
2371
- if (!magicMatches(tail, 0) || tail[4] !== versionMajor) {
2639
+ if (!magicMatches(tail, 0) || tail[4] !== versionMajor || tail[5] !== versionMinor) {
2372
2640
  throw new IntegrityError(".crbm preamble magic/version mismatch with footer");
2373
2641
  }
2374
2642
  }
2375
- let indexBytes;
2643
+ const lowest = hasExtension ? Math.max(PAYLOAD_START, indexOffset - EXT_TRAILER_BYTES - MAX_EXT_BYTES) : indexOffset;
2644
+ let region;
2645
+ let regionStart;
2376
2646
  let servedFromTail;
2377
2647
  if (indexOffset >= tailCoversFrom) {
2378
- const start = indexOffset - tailCoversFrom;
2379
- indexBytes = tail.subarray(start, start + indexLength);
2648
+ region = tail;
2649
+ regionStart = tailCoversFrom;
2380
2650
  servedFromTail = true;
2381
2651
  } else {
2382
- indexBytes = await blob.getRange(indexOffset, indexLength);
2652
+ region = await blob.getRange(lowest, indexOffset + indexLength - lowest);
2653
+ regionStart = lowest;
2383
2654
  servedFromTail = false;
2384
2655
  }
2656
+ const indexBytes = region.subarray(
2657
+ indexOffset - regionStart,
2658
+ indexOffset - regionStart + indexLength
2659
+ );
2385
2660
  if (crc32c(indexBytes) !== indexCrc) {
2386
2661
  throw new IntegrityError(".crbm index CRC mismatch");
2387
2662
  }
@@ -2393,10 +2668,57 @@ var CrbmReader = class _CrbmReader {
2393
2668
  },
2394
2669
  options.crypto.aadFor("index")
2395
2670
  ) : indexBytes;
2671
+ let payloadEnd = indexOffset;
2672
+ let metadata;
2673
+ let metadataLength = 0;
2674
+ if (hasExtension) {
2675
+ const trailerStart = indexOffset - EXT_TRAILER_BYTES;
2676
+ if (trailerStart < PAYLOAD_START) {
2677
+ throw new IntegrityError(".crbm flags an extension block, but has no room for one");
2678
+ }
2679
+ let window = { bytes: region, start: regionStart };
2680
+ if (trailerStart < regionStart) {
2681
+ window = { bytes: await blob.getRange(lowest, indexOffset - lowest), start: lowest };
2682
+ servedFromTail = false;
2683
+ }
2684
+ const trailer = within(window.bytes, window.start, trailerStart, indexOffset);
2685
+ const tview = new DataView(trailer.buffer, trailer.byteOffset, trailer.byteLength);
2686
+ if (!EXT_MAGIC.every((b, i) => trailer[8 + i] === b)) {
2687
+ throw new IntegrityError(".crbm extension block trailer magic mismatch");
2688
+ }
2689
+ const sectionsLength = tview.getUint32(0, true);
2690
+ if (sectionsLength > MAX_EXT_BYTES) {
2691
+ throw new IntegrityError(
2692
+ `.crbm extension block ${sectionsLength}B exceeds cap ${MAX_EXT_BYTES}B`
2693
+ );
2694
+ }
2695
+ const extStart = trailerStart - sectionsLength;
2696
+ if (extStart < PAYLOAD_START) {
2697
+ throw new IntegrityError(`.crbm extension block of ${sectionsLength}B out of bounds`);
2698
+ }
2699
+ let covered;
2700
+ if (extStart >= window.start) {
2701
+ covered = within(window.bytes, window.start, extStart, trailerStart + 4);
2702
+ } else {
2703
+ covered = await blob.getRange(extStart, sectionsLength + 4);
2704
+ servedFromTail = false;
2705
+ if (covered.length !== sectionsLength + 4) {
2706
+ throw new IntegrityError(".crbm extension block read short");
2707
+ }
2708
+ }
2709
+ if (crc32c(covered) !== tview.getUint32(4, true)) {
2710
+ throw new IntegrityError(".crbm extension block CRC mismatch");
2711
+ }
2712
+ ({ metadata, metadataLength } = extensionContents(
2713
+ covered.subarray(0, sectionsLength),
2714
+ options.crypto
2715
+ ));
2716
+ payloadEnd = extStart;
2717
+ }
2396
2718
  const index = parseIndex(
2397
2719
  indexForParse,
2398
- indexOffset,
2399
- maxPayloadBytes,
2720
+ payloadEnd,
2721
+ options.maxPayloadBytes ?? (encrypted ? DEFAULT_MAX_PAYLOAD_BYTES : DEFAULT_MAX_BITMAP_BYTES),
2400
2722
  encrypted ? AEAD_NONCE_BYTES + AEAD_TAG_BYTES : 1
2401
2723
  );
2402
2724
  if (encrypted) {
@@ -2426,7 +2748,9 @@ var CrbmReader = class _CrbmReader {
2426
2748
  index,
2427
2749
  servedFromTail,
2428
2750
  options.crypto,
2429
- storedFooterCrc
2751
+ storedFooterCrc,
2752
+ metadata,
2753
+ metadata === void 0 ? 0 : RETAINED_BYTES_PER_METADATA_BYTE * metadataLength + RETAINED_BYTES_PER_METADATA_KEY * Object.keys(metadata).length
2430
2754
  );
2431
2755
  }
2432
2756
  /** Chunk keys present in this generation, ascending. */
@@ -2486,6 +2810,67 @@ var CrbmReader = class _CrbmReader {
2486
2810
  );
2487
2811
  }
2488
2812
  };
2813
+ function within(bytes, start, from, to) {
2814
+ if (from < start || to > start + bytes.length) {
2815
+ throw new IntegrityError(".crbm extension block read short");
2816
+ }
2817
+ return bytes.subarray(from - start, to - start);
2818
+ }
2819
+ function extensionCorrupt(message) {
2820
+ throw new IntegrityError(`.crbm extension block: ${message}`);
2821
+ }
2822
+ function extensionContents(sections, crypto) {
2823
+ const view = new DataView(sections.buffer, sections.byteOffset, sections.byteLength);
2824
+ let metadata;
2825
+ let metadataLength = 0;
2826
+ let lastType = 0;
2827
+ let pos = 0;
2828
+ while (pos < sections.length) {
2829
+ if (sections.length - pos < EXT_SECTION_HEADER_BYTES)
2830
+ extensionCorrupt("a section header is cut off");
2831
+ const type = sections[pos];
2832
+ const length = view.getUint32(pos + 1, true);
2833
+ pos += EXT_SECTION_HEADER_BYTES;
2834
+ if (type <= lastType) {
2835
+ extensionCorrupt(
2836
+ type === 0 ? "section type 0 is not a section type" : `section type ${type} after ${lastType}`
2837
+ );
2838
+ }
2839
+ if (length > sections.length - pos) extensionCorrupt(`section ${type} runs past the block`);
2840
+ const body = sections.subarray(pos, pos + length);
2841
+ pos += length;
2842
+ lastType = type;
2843
+ if (type === EXT_SECTION_METADATA)
2844
+ ({ metadata, metadataLength } = metadataSection(body, crypto));
2845
+ }
2846
+ return { metadata, metadataLength };
2847
+ }
2848
+ function metadataSection(body, crypto) {
2849
+ if (crypto === void 0) {
2850
+ return { metadata: metadataFromBytes(body, extensionCorrupt), metadataLength: body.length };
2851
+ }
2852
+ const framing = AEAD_NONCE_BYTES + AEAD_TAG_BYTES;
2853
+ if (body.length <= framing || body.length > framing + MAX_METADATA_BYTES) {
2854
+ extensionCorrupt(`sealed metadata of ${body.length}B is not a nonce, up to 1 KiB and a tag`);
2855
+ }
2856
+ let plain;
2857
+ try {
2858
+ plain = crypto.aead.open(
2859
+ {
2860
+ nonce: body.subarray(0, AEAD_NONCE_BYTES),
2861
+ ciphertext: body.subarray(AEAD_NONCE_BYTES, body.length - AEAD_TAG_BYTES),
2862
+ tag: body.subarray(body.length - AEAD_TAG_BYTES)
2863
+ },
2864
+ crypto.aadFor("metadata")
2865
+ );
2866
+ } catch (err) {
2867
+ if (!isIntegrityError(err)) throw err;
2868
+ extensionCorrupt(
2869
+ "the sealed metadata does not open: tampered bytes, a block moved from another object, or a CrbmCrypto whose aadFor does not map the 'metadata' scope"
2870
+ );
2871
+ }
2872
+ return { metadata: metadataFromBytes(plain, extensionCorrupt), metadataLength: plain.length };
2873
+ }
2489
2874
  function parseIndex(indexBytes, payloadEnd, maxPayloadBytes, minPayloadBytes = 1) {
2490
2875
  const capacity = indexCapacity(indexBytes.length);
2491
2876
  const keys = new Uint16Array(capacity);
@@ -2520,7 +2905,9 @@ function parseIndex(indexBytes, payloadEnd, maxPayloadBytes, minPayloadBytes = 1
2520
2905
  }
2521
2906
  if (chunkKey > 65535) throw new IntegrityError(`.crbm chunkKey ${chunkKey} out of range`);
2522
2907
  if (len < minPayloadBytes || len > maxPayloadBytes) {
2523
- throw new IntegrityError(`.crbm chunk ${chunkKey} length ${len} invalid`);
2908
+ throw new IntegrityError(
2909
+ `.crbm chunk ${chunkKey} length ${len} invalid (most is ${maxPayloadBytes})`
2910
+ );
2524
2911
  }
2525
2912
  if (card < 1 || card > MAX_CHUNK_CARDINALITY) {
2526
2913
  throw new IntegrityError(`.crbm chunk ${chunkKey} cardinality ${card} invalid`);
@@ -2561,6 +2948,9 @@ var CrbmWriter = class {
2561
2948
  `generation must be a non-negative integer; got ${options.generation}`
2562
2949
  );
2563
2950
  }
2951
+ this.metadata = metadataBytes(options.metadata, (message) => {
2952
+ throw new ValidationError(message);
2953
+ });
2564
2954
  }
2565
2955
  sink;
2566
2956
  options;
@@ -2570,6 +2960,11 @@ var CrbmWriter = class {
2570
2960
  totalCardinality = 0;
2571
2961
  preambleWritten = false;
2572
2962
  finished = false;
2963
+ /** The metadata's canonical JSON, copied at construction; `undefined` leaves the footer's extension flag clear. */
2964
+ metadata;
2965
+ /** Bytes written so far, the footer included once {@link finish} has run. */
2966
+ written = 0;
2967
+ footerCrc32c;
2573
2968
  /**
2574
2969
  * Append one chunk's payload (the roaring-serialized bytes) with its cardinality. `chunkKey` must be
2575
2970
  * strictly greater than the previous one (ascending, no duplicates); `cardinality` is `[1, 65536]`.
@@ -2591,7 +2986,8 @@ var CrbmWriter = class {
2591
2986
  }
2592
2987
  if (payload.length === 0) throw new ValidationError("chunk payload must be non-empty");
2593
2988
  await this.ensurePreamble();
2594
- const stored = this.options.crypto ? this.sealChunk(chunkKey, payload) : payload;
2989
+ const crypto = this.options.crypto;
2990
+ const stored = crypto ? framed(crypto.aead.seal(payload, crypto.aadFor(chunkKey))) : payload;
2595
2991
  await this.sink.write(stored);
2596
2992
  this.entries.push({
2597
2993
  chunkKey,
@@ -2604,20 +3000,18 @@ var CrbmWriter = class {
2604
3000
  this.totalCardinality += cardinality;
2605
3001
  this.lastKey = chunkKey;
2606
3002
  }
2607
- /** Seal one chunk payload into the on-disk frame `nonce ‖ ciphertext ‖ tag`, bound to its chunkKey. */
2608
- sealChunk(chunkKey, payload) {
2609
- const crypto = this.options.crypto;
2610
- const sealed = crypto.aead.seal(payload, crypto.aadFor(chunkKey));
2611
- const out = new Uint8Array(sealed.nonce.length + sealed.ciphertext.length + sealed.tag.length);
2612
- out.set(sealed.nonce, 0);
2613
- out.set(sealed.ciphertext, sealed.nonce.length);
2614
- out.set(sealed.tag, sealed.nonce.length + sealed.ciphertext.length);
2615
- return out;
2616
- }
2617
- /** Write the index region + the fixed footer. After this the object is complete and immutable. */
3003
+ /**
3004
+ * Write the extension block (when there is metadata), the index region and the fixed footer. After this the object
3005
+ * is complete and immutable.
3006
+ */
2618
3007
  async finish() {
2619
3008
  if (this.finished) throw new ValidationError("CrbmWriter already finished");
2620
3009
  await this.ensurePreamble();
3010
+ if (this.metadata !== void 0) {
3011
+ const block = this.extensionBlock(this.metadata);
3012
+ await this.sink.write(block);
3013
+ this.offset += block.length;
3014
+ }
2621
3015
  const indexPlain = this.encodeIndex();
2622
3016
  let indexRegion = indexPlain;
2623
3017
  let indexNonce;
@@ -2633,8 +3027,22 @@ var CrbmWriter = class {
2633
3027
  this.offset += indexRegion.length;
2634
3028
  const footer = this.buildFooter(indexOffset, indexRegion, indexNonce, indexTag);
2635
3029
  await this.sink.write(footer);
3030
+ this.written = this.offset + footer.length;
3031
+ this.footerCrc32c = new DataView(footer.buffer, footer.byteOffset).getUint32(
3032
+ FOOTER.footerCrc32c,
3033
+ true
3034
+ );
2636
3035
  this.finished = true;
2637
3036
  }
3037
+ /**
3038
+ * The finished object's size and its footer's CRC: what names it, as a reader's fingerprint does. A writer that
3039
+ * has to tell its own object from another stored later under the same key compares these with the object's
3040
+ * footer. Throws before {@link finish}.
3041
+ */
3042
+ get identity() {
3043
+ if (this.footerCrc32c === void 0) throw new ValidationError("CrbmWriter not finished");
3044
+ return { size: this.written, footerCrc: this.footerCrc32c };
3045
+ }
2638
3046
  async ensurePreamble() {
2639
3047
  if (this.preambleWritten) return;
2640
3048
  const preamble = new Uint8Array(PREAMBLE_BYTES);
@@ -2644,6 +3052,25 @@ var CrbmWriter = class {
2644
3052
  await this.sink.write(preamble);
2645
3053
  this.preambleWritten = true;
2646
3054
  }
3055
+ /**
3056
+ * The extension block: one metadata section (`u8 type ‖ u32 length ‖ bytes`), then the trailer
3057
+ * (`u32 sectionsLength ‖ u32 crc32c(sections ‖ sectionsLength) ‖ "CRBX"`). Encrypted, the section holds
3058
+ * `nonce ‖ ciphertext ‖ tag`, sealed with the metadata scope of the associated data.
3059
+ */
3060
+ extensionBlock(metadata) {
3061
+ const crypto = this.options.crypto;
3062
+ const body = crypto ? framed(crypto.aead.seal(metadata, crypto.aadFor("metadata"))) : metadata;
3063
+ const sectionsLength = EXT_SECTION_HEADER_BYTES + body.length;
3064
+ const block = new Uint8Array(sectionsLength + EXT_TRAILER_BYTES);
3065
+ const view = new DataView(block.buffer);
3066
+ block[0] = EXT_SECTION_METADATA;
3067
+ view.setUint32(1, body.length, true);
3068
+ block.set(body, EXT_SECTION_HEADER_BYTES);
3069
+ view.setUint32(sectionsLength, sectionsLength, true);
3070
+ view.setUint32(sectionsLength + 4, crc32c(block.subarray(0, sectionsLength + 4)), true);
3071
+ block.set(EXT_MAGIC, sectionsLength + 8);
3072
+ return block;
3073
+ }
2647
3074
  encodeIndex() {
2648
3075
  const out = [];
2649
3076
  let prevKey = 0;
@@ -2671,7 +3098,11 @@ var CrbmWriter = class {
2671
3098
  view.setBigUint64(FOOTER.indexOffset, BigInt(indexOffset), true);
2672
3099
  view.setBigUint64(FOOTER.indexLength, BigInt(indexRegion.length), true);
2673
3100
  view.setUint32(FOOTER.indexCrc32c, crc32c(indexRegion), true);
2674
- view.setUint32(FOOTER.flags, FLAG_LITTLE_ENDIAN | (encrypted ? FLAG_ENCRYPTED : 0), true);
3101
+ view.setUint32(
3102
+ FOOTER.flags,
3103
+ FLAG_LITTLE_ENDIAN | (encrypted ? FLAG_ENCRYPTED : 0) | (this.metadata === void 0 ? 0 : FLAG_EXTENSION),
3104
+ true
3105
+ );
2675
3106
  view.setUint16(
2676
3107
  FOOTER.payloadCodecId,
2677
3108
  this.options.payloadCodecId ?? PAYLOAD_CODEC_ROARING_PORTABLE,
@@ -2691,6 +3122,226 @@ var CrbmWriter = class {
2691
3122
  return footer;
2692
3123
  }
2693
3124
  };
3125
+ function framed(sealed) {
3126
+ const out = new Uint8Array(sealed.nonce.length + sealed.ciphertext.length + sealed.tag.length);
3127
+ out.set(sealed.nonce, 0);
3128
+ out.set(sealed.ciphertext, sealed.nonce.length);
3129
+ out.set(sealed.tag, sealed.nonce.length + sealed.ciphertext.length);
3130
+ return out;
3131
+ }
3132
+
3133
+ // src/core/load-input.ts
3134
+ var MAX_SERIALIZED_LOAD_BYTES = 4 + 8192 + 8 * 65536 + 65536 * 8192;
3135
+ var DecodedLoadInput = class {
3136
+ constructor(bitmap) {
3137
+ this.bitmap = bitmap;
3138
+ }
3139
+ bitmap;
3140
+ };
3141
+ function tagOf(value) {
3142
+ return Object.prototype.toString.call(value).slice(8, -1);
3143
+ }
3144
+ var VIEW = Object.getPrototypeOf(Uint8Array.prototype);
3145
+ var viewGetter = (name) => Object.getOwnPropertyDescriptor(VIEW, name)?.get;
3146
+ var bufferOf = viewGetter("buffer");
3147
+ var byteOffsetOf = viewGetter("byteOffset");
3148
+ var byteLengthOf = viewGetter("byteLength");
3149
+ function plainView(bytes, what) {
3150
+ let buffer;
3151
+ let offset;
3152
+ let length;
3153
+ try {
3154
+ buffer = bufferOf.call(bytes);
3155
+ offset = byteOffsetOf.call(bytes);
3156
+ length = byteLengthOf.call(bytes);
3157
+ } catch {
3158
+ throw new ValidationError(`${what} must be a Uint8Array of portable Roaring bytes`);
3159
+ }
3160
+ return length === 0 ? new Uint8Array(0) : new Uint8Array(buffer, offset, length);
3161
+ }
3162
+ var sharedByteLength = typeof SharedArrayBuffer === "function" ? Object.getOwnPropertyDescriptor(SharedArrayBuffer.prototype, "byteLength")?.get : void 0;
3163
+ function isShared(buffer) {
3164
+ if (sharedByteLength === void 0) return false;
3165
+ try {
3166
+ sharedByteLength.call(buffer);
3167
+ return true;
3168
+ } catch {
3169
+ return false;
3170
+ }
3171
+ }
3172
+ function overCap(what, length) {
3173
+ return new ValidationError(
3174
+ `${what} is ${length} bytes, more than any canonical 32-bit bitmap serializes to (${MAX_SERIALIZED_LOAD_BYTES}). Call runOptimize() on the bitmap before serializing it.`
3175
+ );
3176
+ }
3177
+ function prepareLoadInput(input, codec) {
3178
+ if (typeof input !== "object" || input === null) {
3179
+ throw new ValidationError(
3180
+ `a load takes ids, { serialized } or { bitmap }; got ${String(input)}`
3181
+ );
3182
+ }
3183
+ const tag = tagOf(input);
3184
+ if (tag === "Uint8Array" || tag === "Uint8ClampedArray") {
3185
+ throw new ValidationError(
3186
+ `a load's ids cannot be a ${tag} or Buffer: each byte would be loaded as an id. Pass portable Roaring bytes as { serialized }, and ids as a Uint32Array or an array of numbers.`
3187
+ );
3188
+ }
3189
+ if (Symbol.iterator in input || Symbol.asyncIterator in input) return input;
3190
+ const keys = Object.keys(input);
3191
+ if (keys.length === 1 && keys[0] === "serialized") {
3192
+ return decode(input.serialized, "{ serialized }", codec);
3193
+ }
3194
+ if (keys.length === 1 && keys[0] === "bitmap") {
3195
+ const bitmap = input.bitmap;
3196
+ if (typeof bitmap !== "object" || bitmap === null || typeof bitmap.serialize !== "function") {
3197
+ throw new ValidationError(`{ bitmap } needs an object with serialize('portable')`);
3198
+ }
3199
+ const what = "{ bitmap }'s serialize('portable')";
3200
+ if (typeof bitmap.getSerializationSizeInBytes === "function") {
3201
+ const size = bitmap.getSerializationSizeInBytes("portable");
3202
+ if (typeof size === "number" && size > MAX_SERIALIZED_LOAD_BYTES) throw overCap(what, size);
3203
+ }
3204
+ return decode(bitmap.serialize("portable"), what, codec);
3205
+ }
3206
+ throw new ValidationError(
3207
+ `a load takes ids (an iterable of integers), { serialized } or { bitmap }, as the only key; got an object with keys [${keys.join(", ")}]. Load options go in the next argument.`
3208
+ );
3209
+ }
3210
+ function decode(bytes, what, codec) {
3211
+ if (typeof bytes !== "object" || bytes === null || tagOf(bytes) !== "Uint8Array") {
3212
+ throw new ValidationError(`${what} must be a Uint8Array of portable Roaring bytes`);
3213
+ }
3214
+ const view = plainView(bytes, what);
3215
+ if (view.byteLength > MAX_SERIALIZED_LOAD_BYTES) throw overCap(what, view.byteLength);
3216
+ const own = isShared(view.buffer) ? new Uint8Array(view) : view;
3217
+ try {
3218
+ return new DecodedLoadInput(
3219
+ codec.safeDeserialize(own, MAX_SERIALIZED_LOAD_BYTES, { whole: true })
3220
+ );
3221
+ } catch (err) {
3222
+ if (!isIntegrityError(err)) throw err;
3223
+ throw new ValidationError(`${what} is not a valid bitmap: ${err.message}`);
3224
+ }
3225
+ }
3226
+
3227
+ // src/core/summary.ts
3228
+ var NONCE_BYTES = 12;
3229
+ var TAG_BYTES = 16;
3230
+ var COUNT_BYTES = 8;
3231
+ var MAX_CARDINALITY = 2 ** 32;
3232
+ var BASE64 = "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789+/";
3233
+ function toBase64(bytes) {
3234
+ let out = "";
3235
+ for (let i = 0; i < bytes.length; i += 3) {
3236
+ const a = bytes[i];
3237
+ const b = bytes[i + 1];
3238
+ const c = bytes[i + 2];
3239
+ out += BASE64[a >> 2] + BASE64[(a & 3) << 4 | (b ?? 0) >> 4];
3240
+ out += b === void 0 ? "=" : BASE64[(b & 15) << 2 | (c ?? 0) >> 6];
3241
+ out += c === void 0 ? "=" : BASE64[c & 63];
3242
+ }
3243
+ return out;
3244
+ }
3245
+ function fromBase64(text) {
3246
+ if (text.length % 4 !== 0 || !/^[A-Za-z0-9+/]*={0,2}$/.test(text)) return void 0;
3247
+ const pad = text.endsWith("==") ? 2 : text.endsWith("=") ? 1 : 0;
3248
+ const out = new Uint8Array(text.length / 4 * 3 - pad);
3249
+ let o = 0;
3250
+ for (let i = 0; i < text.length; i += 4) {
3251
+ const n = BASE64.indexOf(text[i]) << 18 | BASE64.indexOf(text[i + 1]) << 12 | Math.max(BASE64.indexOf(text[i + 2]), 0) << 6 | Math.max(BASE64.indexOf(text[i + 3]), 0);
3252
+ if (o < out.length) out[o++] = n >> 16 & 255;
3253
+ if (o < out.length) out[o++] = n >> 8 & 255;
3254
+ if (o < out.length) out[o++] = n & 255;
3255
+ }
3256
+ return out;
3257
+ }
3258
+ function clearSummary(generation, cardinality, metadata) {
3259
+ return metadata === void 0 || Object.keys(metadata).length === 0 ? { generation, cardinality } : { generation, cardinality, metadata };
3260
+ }
3261
+ function sealSummary(aead, ref, generation, cardinality, metadata) {
3262
+ const meta = metadataBytes(metadata, (message) => {
3263
+ throw new IntegrityError(`summary metadata: ${message}`);
3264
+ });
3265
+ const plain = new Uint8Array(COUNT_BYTES + (meta?.length ?? 0));
3266
+ new DataView(plain.buffer).setBigUint64(0, BigInt(cardinality), true);
3267
+ if (meta !== void 0) plain.set(meta, COUNT_BYTES);
3268
+ const { nonce, ciphertext, tag } = aead.seal(plain, aadFor(ref, generation, "summary"));
3269
+ const framed2 = new Uint8Array(nonce.length + ciphertext.length + tag.length);
3270
+ framed2.set(nonce, 0);
3271
+ framed2.set(ciphertext, nonce.length);
3272
+ framed2.set(tag, nonce.length + ciphertext.length);
3273
+ return { generation, sealed: toBase64(framed2) };
3274
+ }
3275
+ function openSummary(aead, ref, summary) {
3276
+ const fail = (message) => {
3277
+ throw new IntegrityError(`registry record summary: ${message}`);
3278
+ };
3279
+ const framed2 = fromBase64(summary.sealed);
3280
+ const framing = NONCE_BYTES + TAG_BYTES;
3281
+ if (framed2 === void 0 || framed2.length < framing + COUNT_BYTES || framed2.length > framing + COUNT_BYTES + MAX_METADATA_BYTES) {
3282
+ fail("the sealed summary is not a nonce, a count, up to 1 KiB of metadata and a tag");
3283
+ }
3284
+ const bytes = framed2;
3285
+ let plain;
3286
+ try {
3287
+ plain = aead.open(
3288
+ {
3289
+ nonce: bytes.subarray(0, NONCE_BYTES),
3290
+ ciphertext: bytes.subarray(NONCE_BYTES, bytes.length - TAG_BYTES),
3291
+ tag: bytes.subarray(bytes.length - TAG_BYTES)
3292
+ },
3293
+ aadFor(ref, summary.generation, "summary")
3294
+ );
3295
+ } catch (err) {
3296
+ if (!isIntegrityError(err)) throw err;
3297
+ return fail(
3298
+ "the sealed summary does not open: tampered bytes, or a summary moved from another generation"
3299
+ );
3300
+ }
3301
+ if (plain.length < COUNT_BYTES) fail("the sealed summary is shorter than its count");
3302
+ const count = new DataView(plain.buffer, plain.byteOffset, plain.byteLength).getBigUint64(
3303
+ 0,
3304
+ true
3305
+ );
3306
+ if (count > BigInt(MAX_CARDINALITY)) fail(`cardinality ${count} is outside 0 to 2^32`);
3307
+ const metadata = plain.length === COUNT_BYTES ? void 0 : metadataFromBytes(plain.subarray(COUNT_BYTES), fail);
3308
+ return { cardinality: Number(count), metadata };
3309
+ }
3310
+ function usableSummary(ref, row, aead) {
3311
+ const summary = row.summary;
3312
+ if (summary === void 0 || row.status !== "active" || summary.generation !== row.currentGen) {
3313
+ return void 0;
3314
+ }
3315
+ const keyed = row.wrappedDeks !== void 0 && row.wrappedDeks.length > 0;
3316
+ if ("sealed" in summary) {
3317
+ if (!keyed || aead === void 0) return void 0;
3318
+ try {
3319
+ return openSummary(aead, ref, summary);
3320
+ } catch (err) {
3321
+ if (isIntegrityError(err)) return void 0;
3322
+ throw err;
3323
+ }
3324
+ }
3325
+ if (keyed) return void 0;
3326
+ return { cardinality: summary.cardinality, metadata: summary.metadata };
3327
+ }
3328
+ function canonicalOrNone(metadata) {
3329
+ if (metadata === void 0) return void 0;
3330
+ const json = canonicalMetadataJson(metadata, (message) => {
3331
+ throw new IntegrityError(`metadata: ${message}`);
3332
+ });
3333
+ return json === "{}" ? void 0 : json;
3334
+ }
3335
+ function summaryAgrees(summary, object) {
3336
+ return summary.cardinality === object.cardinality && canonicalOrNone(summary.metadata) === canonicalOrNone(object.metadata);
3337
+ }
3338
+ function metadataToCarry(object, authenticated) {
3339
+ const empty = (metadata) => metadata === void 0 || Object.keys(metadata).length === 0;
3340
+ return empty(object) && authenticated !== void 0 && !empty(authenticated.metadata) ? authenticated.metadata : object;
3341
+ }
3342
+ function summaryOf(ref, generation, description, aead) {
3343
+ return aead === void 0 ? clearSummary(generation, description.cardinality, description.metadata) : sealSummary(aead, ref, generation, description.cardinality, description.metadata);
3344
+ }
2694
3345
 
2695
3346
  // src/core/crbm-storage-source.ts
2696
3347
  function storageBlobReader(driver, key) {
@@ -2709,6 +3360,88 @@ var notThePinned = (ref, generation) => new NotFoundError(
2709
3360
  function versionOf(generation, lineage) {
2710
3361
  return lineage === void 0 ? String(generation) : `${generation}:${String(lineage)}`;
2711
3362
  }
3363
+ function summaryOfGeneration(generation, description) {
3364
+ const { cardinality, metadata } = description;
3365
+ return metadata === void 0 || Object.keys(metadata).length === 0 ? { generation, cardinality } : { generation, cardinality, metadata: Object.freeze({ ...metadata }) };
3366
+ }
3367
+ var sameKeys = (a, b) => a !== void 0 && JSON.stringify(a) === JSON.stringify(b);
3368
+ function once(fn) {
3369
+ let pending;
3370
+ return () => {
3371
+ if (pending === void 0) {
3372
+ const mine = fn();
3373
+ pending = mine;
3374
+ mine.catch(() => {
3375
+ if (pending === mine) pending = void 0;
3376
+ });
3377
+ }
3378
+ return pending;
3379
+ };
3380
+ }
3381
+ var readerIsOf = (reader, live) => reader.generation === live.target.generation && reader.lineage === live.target.lineage;
3382
+ function describe(reader) {
3383
+ let cardinality = 0;
3384
+ for (const n of reader.cardinalities().values()) cardinality += n;
3385
+ return { cardinality, metadata: reader.metadata };
3386
+ }
3387
+ var Snapshot = class _Snapshot {
3388
+ constructor(target, open4, prior) {
3389
+ this.target = target;
3390
+ this.open = open4;
3391
+ this.prior = prior;
3392
+ }
3393
+ target;
3394
+ open;
3395
+ prior;
3396
+ /** When the snapshot's refresh clock started. A refresh that failed transiently moves it back, so it lapses sooner. */
3397
+ installedAtMs = 0;
3398
+ /** Set by the cache, to hear of a reader as it is opened. */
3399
+ onReader;
3400
+ opened;
3401
+ static live(target, open4, prior) {
3402
+ return new _Snapshot(target, open4, prior);
3403
+ }
3404
+ static eager(reader) {
3405
+ const snap = new _Snapshot(void 0, void 0, void 0);
3406
+ snap.opened = reader;
3407
+ return snap;
3408
+ }
3409
+ /** The reader, opened by the first call. */
3410
+ get reader() {
3411
+ if (this.opened === void 0) {
3412
+ const target = this.target;
3413
+ this.opened = target.then(async (live) => {
3414
+ if (live === null) return null;
3415
+ const kept = await this.reuse(live);
3416
+ this.prior = void 0;
3417
+ return kept ?? this.open(live);
3418
+ });
3419
+ this.onReader?.(this.opened);
3420
+ }
3421
+ return this.opened;
3422
+ }
3423
+ /** The reader if one has been asked for, without asking for one. */
3424
+ get openedReader() {
3425
+ return this.opened;
3426
+ }
3427
+ /** The prior snapshot's reader, if it is of what this snapshot resolved. */
3428
+ async reuse(live) {
3429
+ const prior = this.prior;
3430
+ const reader = prior === void 0 ? null : await prior.catch(() => null);
3431
+ return reader !== null && readerIsOf(reader, live) ? reader : null;
3432
+ }
3433
+ /** Keep the prior snapshot's reader when the refresh found its generation, so a count does not drop it. */
3434
+ async adopt(live) {
3435
+ if (this.opened !== void 0 || this.prior === void 0) return;
3436
+ const reader = await this.reuse(live);
3437
+ this.prior = void 0;
3438
+ if (reader !== null && this.opened === void 0) {
3439
+ this.opened = Promise.resolve(reader);
3440
+ this.onReader?.(this.opened);
3441
+ }
3442
+ }
3443
+ };
3444
+ var SNAPSHOT_BASE_BYTES = 256;
2712
3445
  var BULK_FLUSH_IDS = 1 << 20;
2713
3446
  var YIELD_EVERY_IDS = 1 << 14;
2714
3447
  var CrbmStorageChunkSource = class {
@@ -2746,6 +3479,10 @@ var CrbmStorageChunkSource = class {
2746
3479
  maxBytes: maxOpenIndexBytes ?? DEFAULT_MAX_OPEN_INDEX_BYTES,
2747
3480
  clock: clock ?? { now: () => 0 }
2748
3481
  });
3482
+ this.distrusted = new BoundedLru({
3483
+ maxEntries: maxOpenSegments ?? DEFAULT_MAX_OPEN_SEGMENTS,
3484
+ clock: { now: () => 0 }
3485
+ });
2749
3486
  this.replacedPins = new BoundedLru({
2750
3487
  maxEntries: maxOpenSegments ?? DEFAULT_MAX_OPEN_SEGMENTS,
2751
3488
  clock: { now: () => 0 }
@@ -2787,23 +3524,46 @@ var CrbmStorageChunkSource = class {
2787
3524
  * check again.
2788
3525
  */
2789
3526
  invalidations = 0;
3527
+ /**
3528
+ * Generations whose row summary this process found to disagree with their object, by segment and generation. A
3529
+ * summary here is not used again: a count of that generation reads the object. Bounded as the readers are, so a
3530
+ * long-lived process cannot grow it without limit; one that falls out is trusted again until its next open finds
3531
+ * the disagreement again.
3532
+ */
3533
+ distrusted;
2790
3534
  registry;
2791
3535
  keystore;
2792
3536
  requireEncryption;
2793
3537
  readerOptions;
2794
3538
  clock;
2795
3539
  currentGenTtlMs;
2796
- /** The current resolved reader for a segment, refreshing on the TTL. Cheap within the TTL window. */
2797
- resolvedReader(ref) {
3540
+ /**
3541
+ * The segment's live snapshot, refreshing on the TTL. Cheap within the TTL window. The snapshot holds the resolved
3542
+ * target, and its reader once a read has asked for one.
3543
+ */
3544
+ liveSnapshot(ref) {
2798
3545
  const key = segmentKey(ref);
2799
3546
  const existing = this.snapshots.get(key);
2800
- if (existing === void 0) return this.install(key, this.resolveLatest(ref)).reader;
2801
- if (!this.expired(existing.installedAtMs)) return existing.reader;
3547
+ if (existing === void 0) {
3548
+ return this.install(
3549
+ key,
3550
+ Snapshot.live(this.resolveLive(ref), (live) => this.openLive(ref, live))
3551
+ );
3552
+ }
3553
+ if (!this.expired(existing.installedAtMs)) return existing;
2802
3554
  const snap = this.install(
2803
3555
  key,
2804
- this.computeRefreshedReader(ref, existing, () => this.retrySoon(snap))
3556
+ Snapshot.live(
3557
+ this.refreshedTarget(ref, existing, () => this.retrySoon(snap)),
3558
+ (live) => this.openLive(ref, live),
3559
+ existing.openedReader
3560
+ )
2805
3561
  );
2806
- return snap.reader;
3562
+ return snap;
3563
+ }
3564
+ /** The current resolved reader for a segment, refreshing on the TTL. */
3565
+ resolvedReader(ref) {
3566
+ return this.liveSnapshot(ref).reader;
2807
3567
  }
2808
3568
  /**
2809
3569
  * A refresh failed transiently and the prior reader keeps serving: make this snapshot lapse after
@@ -2813,45 +3573,56 @@ var CrbmStorageChunkSource = class {
2813
3573
  retrySoon(snap) {
2814
3574
  snap.installedAtMs = this.now() - (this.currentGenTtlMs - Math.min(this.currentGenTtlMs, REFRESH_RETRY_MS));
2815
3575
  }
2816
- /** TTL elapsed: cheaply re-resolve `currentGen`; reopen only if it actually changed, else reuse the prior reader. */
2817
- async computeRefreshedReader(ref, existing, onTransientFault) {
2818
- const current = await existing.reader.catch(() => null);
2819
- let target;
3576
+ /**
3577
+ * TTL elapsed: cheaply re-resolve the pointer. The row read is always fresh, so what the new snapshot says of the
3578
+ * generation (its summary included) is what the row says now, never the prior snapshot's. The reader is reopened
3579
+ * only if the generation or the row changed, and only when a read asks for it.
3580
+ */
3581
+ async refreshedTarget(ref, existing, onTransientFault) {
3582
+ const prior = await existing.target.catch(() => null);
2820
3583
  try {
2821
- target = await this.resolveTarget(ref);
3584
+ return await this.resolveLive(ref, prior);
2822
3585
  } catch (err) {
2823
3586
  if (!isTransientError(err)) throw err;
2824
- target = void 0;
2825
- }
2826
- if (target === void 0) {
2827
- if (current === null) return this.resolveLatest(ref);
3587
+ const alive = prior !== null;
3588
+ if (!alive) return this.resolveLive(ref);
2828
3589
  onTransientFault();
2829
- return current;
2830
- }
2831
- if (target === null) return null;
2832
- if (current !== null && target.generation === current.generation && target.lineage === current.lineage) {
2833
- return current;
3590
+ return prior;
2834
3591
  }
2835
- return this.openForTarget(ref, target);
2836
3592
  }
2837
3593
  /**
2838
- * Memoize a reader-promise as the segment's snapshot at `now()`, forgetting it later if it resolves to
2839
- * null / throws (identity-guarded, so a stale cleanup never clobbers a fresher snapshot). Installing the
2840
- * (pending) promise **synchronously** is what lets concurrent callers coalesce onto one in-flight resolve.
3594
+ * Memoize a snapshot as the segment's at `now()`, forgetting it later if it resolves to null / throws
3595
+ * (identity-guarded, so a stale cleanup never clobbers a fresher snapshot). Installing the (pending) snapshot
3596
+ * **synchronously** is what lets concurrent callers coalesce onto one in-flight resolve.
2841
3597
  */
2842
- install(key, reader) {
2843
- const snap = { reader, installedAtMs: this.now() };
3598
+ install(key, snap) {
3599
+ snap.installedAtMs = this.now();
2844
3600
  this.snapshots.set(key, snap);
2845
3601
  const forgetIfStale = () => {
2846
3602
  if (this.snapshots.get(key) === snap) this.snapshots.delete(key);
2847
3603
  };
2848
- reader.then((r) => {
2849
- if (r === null) {
2850
- forgetIfStale();
2851
- return;
2852
- }
2853
- if (this.snapshots.peek(key) === snap) this.snapshots.setWeight(key, r.retainedIndexBytes);
2854
- }, forgetIfStale);
3604
+ snap.onReader = (reader) => {
3605
+ reader.then((r) => {
3606
+ if (r === null) {
3607
+ forgetIfStale();
3608
+ return;
3609
+ }
3610
+ if (this.snapshots.peek(key) === snap) this.snapshots.setWeight(key, r.retainedBytes);
3611
+ }, forgetIfStale);
3612
+ };
3613
+ if (snap.target !== void 0) {
3614
+ snap.target.then(async (live) => {
3615
+ if (live === null) {
3616
+ forgetIfStale();
3617
+ return;
3618
+ }
3619
+ if (snap.openedReader === void 0 && this.snapshots.peek(key) === snap) {
3620
+ this.snapshots.setWeight(key, live.bytes);
3621
+ }
3622
+ await snap.adopt(live);
3623
+ }, forgetIfStale);
3624
+ }
3625
+ if (snap.openedReader !== void 0) snap.onReader(snap.openedReader);
2855
3626
  return snap;
2856
3627
  }
2857
3628
  now() {
@@ -2883,12 +3654,12 @@ var CrbmStorageChunkSource = class {
2883
3654
  * concurrent sweep. One retry, then propagate — same contract, same eviction rule.
2884
3655
  */
2885
3656
  async currentGeneration(ref) {
2886
- const pending = this.resolvedReader(ref);
3657
+ const snap = this.liveSnapshot(ref);
2887
3658
  try {
2888
- return (await pending)?.generation ?? null;
3659
+ return (await snap.reader)?.generation ?? null;
2889
3660
  } catch (err) {
2890
3661
  if (!isNotFoundError(err)) throw err;
2891
- this.dropStale(segmentKey(ref), pending);
3662
+ this.dropStale(segmentKey(ref), snap);
2892
3663
  return (await this.resolvedReader(ref))?.generation ?? null;
2893
3664
  }
2894
3665
  }
@@ -2908,6 +3679,7 @@ var CrbmStorageChunkSource = class {
2908
3679
  const key = segmentKey(ref);
2909
3680
  this.snapshots.delete(key);
2910
3681
  const pinnedPrefix = `${key}@`;
3682
+ this.distrusted.deleteWhere((k) => k.startsWith(pinnedPrefix));
2911
3683
  this.snapshots.deleteWhere((k) => k.startsWith(pinnedPrefix));
2912
3684
  this.replacedPins.deleteWhere((k) => k.startsWith(pinnedPrefix));
2913
3685
  for (const k of this.checking.keys()) if (k.startsWith(pinnedPrefix)) this.checking.delete(k);
@@ -2928,13 +3700,13 @@ var CrbmStorageChunkSource = class {
2928
3700
  * almost always served from the cached snapshot with no backend call at all. One retry, then propagate.
2929
3701
  */
2930
3702
  async currentVersion(ref) {
2931
- const pending = this.resolvedReader(ref);
3703
+ const snap = this.liveSnapshot(ref);
2932
3704
  let reader;
2933
3705
  try {
2934
- reader = await pending;
3706
+ reader = await snap.reader;
2935
3707
  } catch (err) {
2936
3708
  if (!isNotFoundError(err)) throw err;
2937
- this.dropStale(segmentKey(ref), pending);
3709
+ this.dropStale(segmentKey(ref), snap);
2938
3710
  reader = await this.resolvedReader(ref);
2939
3711
  }
2940
3712
  return reader === null ? null : versionOf(reader.generation, reader.lineage);
@@ -2957,34 +3729,38 @@ var CrbmStorageChunkSource = class {
2957
3729
  }
2958
3730
  }
2959
3731
  async pinOnce(ref) {
2960
- const target = await this.resolveTarget(ref);
2961
- if (target === null) return null;
3732
+ const live = await this.resolveLive(ref);
3733
+ if (live === null) return null;
3734
+ const target = live.target;
2962
3735
  const version = versionOf(target.generation, target.lineage);
2963
3736
  let reader;
2964
3737
  let opened = true;
2965
3738
  if (target.lineage === void 0) {
2966
- reader = await this.openForTarget(ref, target);
2967
- this.install(this.pinnedKey(ref, version, reader.fingerprint), Promise.resolve(reader));
3739
+ reader = await this.openLive(ref, live);
3740
+ this.install(
3741
+ this.pinnedKey(ref, version, reader.fingerprint),
3742
+ Snapshot.eager(Promise.resolve(reader))
3743
+ );
2968
3744
  } else {
2969
3745
  const epoch = this.invalidations;
2970
3746
  const key = this.pinnedKey(ref, version);
2971
3747
  let entry = this.snapshots.get(key);
2972
3748
  opened = entry === void 0;
2973
- entry ??= this.install(key, this.openForTarget(ref, target));
3749
+ entry ??= this.install(key, Snapshot.eager(this.openLive(ref, live)));
2974
3750
  reader = await entry.reader;
2975
3751
  if (reader === null) return null;
2976
3752
  const replaced = (r) => this.replacedPins.get(this.heldKey(ref, version, r.fingerprint)) !== void 0;
2977
3753
  const known = !opened && replaced(reader);
2978
3754
  if (!known && this.invalidations === epoch && this.snapshots.peek(key) === void 0) {
2979
- this.install(key, Promise.resolve(reader));
3755
+ this.install(key, Snapshot.eager(Promise.resolve(reader)));
2980
3756
  }
2981
3757
  if (known) {
2982
3758
  const current = this.snapshots.peek(key);
2983
- let fresh = current !== void 0 && current !== entry ? current : this.install(key, this.openForTarget(ref, target));
3759
+ let fresh = current !== void 0 && current !== entry ? current : this.install(key, Snapshot.eager(this.openLive(ref, live)));
2984
3760
  reader = await fresh.reader;
2985
3761
  if (reader === null) return null;
2986
3762
  if (fresh === current && replaced(reader)) {
2987
- fresh = this.install(key, this.openForTarget(ref, target));
3763
+ fresh = this.install(key, Snapshot.eager(this.openLive(ref, live)));
2988
3764
  reader = await fresh.reader;
2989
3765
  if (reader === null) return null;
2990
3766
  }
@@ -3050,7 +3826,7 @@ var CrbmStorageChunkSource = class {
3050
3826
  if (entry === void 0) {
3051
3827
  if (fingerprint !== void 0 && this.replacedPins.get(this.heldKey(ref, at, fingerprint)) !== void 0)
3052
3828
  throw notThePinned(ref, generation);
3053
- entry = this.install(key, this.openAt(ref, generation));
3829
+ entry = this.install(key, Snapshot.eager(this.openAt(ref, generation)));
3054
3830
  }
3055
3831
  let reader;
3056
3832
  try {
@@ -3185,58 +3961,94 @@ var CrbmStorageChunkSource = class {
3185
3961
  return false;
3186
3962
  }
3187
3963
  /**
3188
- * Evict a snapshot we just failed to read from — matched by its reader *promise*, so a concurrent call's
3964
+ * Evict a snapshot we just failed to read from — matched by the snapshot itself, so a concurrent call's
3189
3965
  * fresher snapshot is never clobbered. The identity guard is the whole point: `install` replaces the entry
3190
3966
  * wholesale, so comparing anything else would drop a good snapshot on the floor.
3191
3967
  */
3192
- dropStale(key, pending) {
3193
- const cur = this.snapshots.get(key);
3194
- if (cur !== void 0 && cur.reader === pending) this.snapshots.delete(key);
3195
- }
3196
- async resolveLatest(ref) {
3197
- const target = await this.resolveTarget(ref);
3198
- if (target === null) return null;
3199
- return this.openForTarget(ref, target);
3968
+ dropStale(key, snap) {
3969
+ if (this.snapshots.get(key) === snap) this.snapshots.delete(key);
3200
3970
  }
3201
3971
  /** Open a {@link CrbmReader} for an already-resolved generation target (decrypting if the segment is encrypted). */
3202
3972
  async openForTarget(ref, target) {
3973
+ const crypto = await this.cryptoForRead(ref, target.generation, target.wrappedDeks);
3974
+ return this.openGeneration(ref, target, crypto);
3975
+ }
3976
+ openGeneration(ref, target, crypto) {
3203
3977
  const genKey = {
3204
3978
  namespace: ref.namespace,
3205
3979
  segment: ref.segment,
3206
3980
  generation: target.generation
3207
3981
  };
3208
- const crypto = await this.cryptoForRead(ref, target.generation, target.wrappedDeks);
3209
3982
  return openChecked(this.driver, genKey, {
3210
3983
  ...this.readerOptions,
3211
3984
  crypto,
3212
3985
  lineage: target.lineage
3213
3986
  });
3214
3987
  }
3988
+ /**
3989
+ * Open the generation a resolution found, and hold the row's summary against the object while it is open anyway:
3990
+ * the same count and metadata, or the summary is not believed again in this process ({@link distrusted}). This
3991
+ * costs no request; a mismatch is not an error, since the object is the truth and this read has it.
3992
+ */
3993
+ async openLive(ref, live) {
3994
+ const reader = await this.openGeneration(ref, live.target, await live.crypto());
3995
+ const summary = await live.summary();
3996
+ if (summary !== void 0 && !summaryAgrees(summary, describe(reader))) {
3997
+ this.distrusted.set(this.distrustKey(ref, live.target.generation, live.target.lineage), true);
3998
+ }
3999
+ return reader;
4000
+ }
4001
+ distrustKey(ref, generation, lineage) {
4002
+ return `${segmentKey(ref)}@${generation}:${String(lineage)}`;
4003
+ }
4004
+ /** What the row says of the generation, unless it is a summary this process has found to disagree with its object. */
4005
+ async trustedSummary(ref, live) {
4006
+ if (this.distrusted.peek(this.distrustKey(ref, live.target.generation, live.target.lineage)) !== void 0) {
4007
+ return void 0;
4008
+ }
4009
+ return live.summary();
4010
+ }
3215
4011
  /**
3216
4012
  * The current generation + its DEK wrappings: the registry's authoritative record (one strong read), or a
3217
4013
  * `list` scan for the max generation when there's no registry (a registry-less source can only read cleartext
3218
- * — there's nowhere a wrapped DEK could live).
4014
+ * — there's nowhere a wrapped DEK could live). The row's summary rides along, to be used or not by whoever
4015
+ * asks.
3219
4016
  */
3220
- async resolveTarget(ref) {
4017
+ async resolveLive(ref, prior) {
3221
4018
  if (this.registry !== void 0) {
3222
4019
  const record = await this.registry.get(ref);
3223
4020
  if (record === null) return null;
3224
4021
  if (record.status === "destroyed") return null;
3225
4022
  if (record.currentGen === null) return null;
4023
+ const generation = record.currentGen;
4024
+ const keys = record.wrappedDeks;
4025
+ const unwrap = keys === void 0 || keys.length === 0 || this.keystore === void 0 ? void 0 : prior?.unwrap !== void 0 && sameKeys(prior.target.wrappedDeks, keys) ? prior.unwrap : once(() => this.keystore.openDek(keys));
4026
+ const crypto = once(() => this.cryptoForRead(ref, generation, keys, unwrap));
3226
4027
  return {
3227
- generation: record.currentGen,
3228
- lineage: record.token,
3229
- wrappedDeks: record.wrappedDeks
4028
+ target: { generation, lineage: record.token, wrappedDeks: record.wrappedDeks },
4029
+ crypto,
4030
+ unwrap,
4031
+ // The same checks as a read of the object makes (`requireEncryption`, a keystore for an encrypted row)
4032
+ // come first, then the row's summary, used only if it names this generation in the shape the keys call for.
4033
+ summary: once(async () => usableSummary(ref, record, (await crypto())?.aead)),
4034
+ bytes: SNAPSHOT_BASE_BYTES + (record.summary === void 0 ? 0 : JSON.stringify(record.summary).length)
3230
4035
  };
3231
4036
  }
3232
4037
  let maxGen = -1;
3233
4038
  for await (const key of this.driver.list(ref)) {
3234
4039
  if (key.generation > maxGen) maxGen = key.generation;
3235
4040
  }
3236
- return maxGen < 0 ? null : { generation: maxGen };
4041
+ if (maxGen < 0) return null;
4042
+ return {
4043
+ target: { generation: maxGen },
4044
+ crypto: () => Promise.resolve(void 0),
4045
+ unwrap: void 0,
4046
+ summary: () => Promise.resolve(void 0),
4047
+ bytes: SNAPSHOT_BASE_BYTES
4048
+ };
3237
4049
  }
3238
4050
  /** Build the per-generation decryption context for an encrypted segment (undefined for cleartext). */
3239
- async cryptoForRead(ref, generation, wrappedDeks) {
4051
+ async cryptoForRead(ref, generation, wrappedDeks, unwrap) {
3240
4052
  if (wrappedDeks === void 0 || wrappedDeks.length === 0) {
3241
4053
  if (this.requireEncryption) {
3242
4054
  throw new KeyUnavailableError(
@@ -3245,12 +4057,13 @@ var CrbmStorageChunkSource = class {
3245
4057
  }
3246
4058
  return void 0;
3247
4059
  }
3248
- if (this.keystore === void 0) {
4060
+ const keystore = this.keystore;
4061
+ if (keystore === void 0) {
3249
4062
  throw new KeyUnavailableError(
3250
4063
  `segment "${ref.segment}" is encrypted but this CrbmStorageChunkSource has no keystore`
3251
4064
  );
3252
4065
  }
3253
- const aead = await this.keystore.openDek(wrappedDeks);
4066
+ const aead = await (unwrap ?? (() => keystore.openDek(wrappedDeks)))();
3254
4067
  return { aead, aadFor: (scope) => aadFor(ref, generation, scope) };
3255
4068
  }
3256
4069
  async getChunk(ref) {
@@ -3277,6 +4090,46 @@ var CrbmStorageChunkSource = class {
3277
4090
  null
3278
4091
  );
3279
4092
  }
4093
+ /**
4094
+ * The current generation's number, id count and metadata. The answer comes from the row the snapshot resolved when
4095
+ * it can: the row's summary of the generation it names, on an active row, in the shape the row's keys call for
4096
+ * (a sealed one is opened with the segment's key), and one this process has not found to disagree with its
4097
+ * object. That is one registry read when the snapshot is cold, none when it is warm, and no read of the object.
4098
+ * With no summary it can use (an older row, one for another generation, one that does not open), it opens the
4099
+ * generation and answers from the object. A row that names an object that is gone still answers, while a read
4100
+ * of the object throws: the number is true of the generation the row names.
4101
+ *
4102
+ * What the row says is not confirmed here: a summary edited to agree with itself by whoever can write the row is
4103
+ * believed until the next time the object is opened, which holds it against the summary.
4104
+ */
4105
+ async summary(ref) {
4106
+ validateUserRef(ref);
4107
+ for (let attempt = 0; attempt < 2; attempt++) {
4108
+ const snap = this.liveSnapshot(ref);
4109
+ try {
4110
+ const live = await snap.target;
4111
+ if (live === null) return null;
4112
+ const row = await this.trustedSummary(ref, live);
4113
+ if (row !== void 0) return summaryOfGeneration(live.target.generation, row);
4114
+ const reader = await snap.reader;
4115
+ return reader === null ? null : summaryOfGeneration(reader.generation, describe(reader));
4116
+ } catch (err) {
4117
+ if (attempt === 1) throw err;
4118
+ const checks = isNotFoundError(err) || isIntegrityError(err) || isValidationError(err);
4119
+ if (!checks || !isNotFoundError(err) && !await this.replacedUnder(ref, snap.reader, err)) {
4120
+ throw err;
4121
+ }
4122
+ this.dropStale(segmentKey(ref), snap);
4123
+ }
4124
+ }
4125
+ return null;
4126
+ }
4127
+ /** {@link summary} of a specific generation, from the object a pin holds. `held` as for {@link getChunkAt}. */
4128
+ async summaryAt(ref, generation, held) {
4129
+ validateUserRef(ref);
4130
+ const reader = await this.readerAt(ref, generation, held?.version, held?.fingerprint);
4131
+ return reader === null ? null : summaryOfGeneration(reader.generation, describe(reader));
4132
+ }
3280
4133
  /**
3281
4134
  * Run `read` against the pinned snapshot, healing the torn-read window generation GC can open: if the
3282
4135
  * generation we resolved was superseded *and* swept (the grace window elapsed), the Storage driver throws
@@ -3305,40 +4158,143 @@ var CrbmStorageChunkSource = class {
3305
4158
  */
3306
4159
  async withFreshSnapshot(ref, read, ifGone) {
3307
4160
  for (let attempt = 0; attempt < 2; attempt++) {
3308
- const pending = this.resolvedReader(ref);
4161
+ const snap = this.liveSnapshot(ref);
4162
+ const pending = snap.reader;
3309
4163
  try {
3310
4164
  const reader = await pending;
3311
4165
  if (reader === null) return ifGone;
3312
4166
  return await read(reader);
3313
4167
  } catch (err) {
3314
- if (!isNotFoundError(err) || attempt === 1) throw err;
3315
- this.dropStale(segmentKey(ref), pending);
4168
+ if (attempt === 1) throw err;
4169
+ if (!isNotFoundError(err) && !await this.replacedUnder(ref, pending, err)) throw err;
4170
+ this.dropStale(segmentKey(ref), snap);
3316
4171
  }
3317
4172
  }
3318
4173
  return ifGone;
3319
4174
  }
4175
+ /**
4176
+ * Whether `err`, raised by a read through the snapshot's reader, came from an object that is no longer the one the
4177
+ * reader opened. A generation number can be taken again once its object is deleted (an erasure removes the
4178
+ * generations above a rolled-back pointer that held the id, a collection pass can stop part-way, and a load then
4179
+ * numbers `currentGen + 1` again), so a reader that still holds the old object's index reads the new object's
4180
+ * bytes at the old offsets: a chunk checksum fails, or a range runs past the end. The object's footer says which it
4181
+ * was, as it does for a pin ({@link CrbmReader.sameObject}): replaced, or gone, re-resolves; the same object means
4182
+ * the error is real, and it stands. Only these two errors pay for the footer read, never a read that succeeded. A
4183
+ * transient fault in the check is rethrown; a footer that cannot be read says nothing, and the error stands.
4184
+ */
4185
+ async replacedUnder(ref, pending, err) {
4186
+ if (!(isIntegrityError(err) || isValidationError(err))) return false;
4187
+ let reader;
4188
+ try {
4189
+ reader = await pending;
4190
+ } catch {
4191
+ return false;
4192
+ }
4193
+ if (reader === null) return false;
4194
+ const at = {
4195
+ namespace: ref.namespace,
4196
+ segment: ref.segment,
4197
+ generation: reader.generation
4198
+ };
4199
+ try {
4200
+ return !await CrbmReader.sameObject(storageBlobReader(this.driver, at), reader.fingerprint);
4201
+ } catch (checkErr) {
4202
+ if (isNotFoundError(checkErr)) return true;
4203
+ if (isTransientError(checkErr)) throw checkErr;
4204
+ return false;
4205
+ }
4206
+ }
3320
4207
  };
3321
- function writeCrbmGeneration(driver, key, chunks, options = {}) {
3322
- const sorted = [...chunks].sort((a, b) => a.chunkKey - b.chunkKey);
4208
+ function* encodeEach(chunks) {
4209
+ for (const { chunkKey, bitmap } of chunks) {
4210
+ if (bitmap.isEmpty) continue;
4211
+ bitmap.optimize?.();
4212
+ yield { chunkKey, payload: bitmap.serialize(), cardinality: bitmap.size };
4213
+ }
4214
+ }
4215
+ async function writeEncodedChunks(driver, key, chunks, options) {
3323
4216
  const tick = yieldEvery(options.clock);
3324
- return driver.putImmutable(key, async (sink) => {
3325
- const writer = new CrbmWriter(sink, { generation: key.generation, crypto: options.crypto });
3326
- for (const { chunkKey, bitmap } of sorted) {
3327
- if (bitmap.isEmpty) continue;
3328
- bitmap.optimize?.();
3329
- await writer.addChunk(chunkKey, bitmap.serialize(), bitmap.size);
4217
+ let chunkCount = 0;
4218
+ let cardinality = 0;
4219
+ let identity;
4220
+ const { size, sha256 } = await driver.putImmutable(key, async (sink) => {
4221
+ const writer = new CrbmWriter(sink, {
4222
+ generation: key.generation,
4223
+ crypto: options.crypto,
4224
+ metadata: options.metadata
4225
+ });
4226
+ for (const chunk of chunks) {
4227
+ await writer.addChunk(chunk.chunkKey, chunk.payload, chunk.cardinality);
4228
+ chunkCount++;
4229
+ cardinality += chunk.cardinality;
3330
4230
  const pause = tick();
3331
4231
  if (pause !== null) await pause;
3332
4232
  }
3333
4233
  await writer.finish();
4234
+ identity = writer.identity;
3334
4235
  });
4236
+ if (identity === void 0) {
4237
+ throw new IntegrityError(
4238
+ `the driver committed ${key.segment}.${key.generation} without writing it`
4239
+ );
4240
+ }
4241
+ return {
4242
+ size,
4243
+ sha256,
4244
+ chunkCount,
4245
+ cardinality,
4246
+ fingerprint: fingerprintFor(identity.size, identity.footerCrc)
4247
+ };
4248
+ }
4249
+ async function holdsObject(storage, key, fingerprint) {
4250
+ try {
4251
+ return await CrbmReader.sameObject(storageBlobReader(storage, key), fingerprint);
4252
+ } catch {
4253
+ return false;
4254
+ }
4255
+ }
4256
+ function objectIsEncrypted(storage, key) {
4257
+ return footerSaysEncrypted(storageBlobReader(storage, key));
4258
+ }
4259
+ function rememberingTail(blob) {
4260
+ let first;
4261
+ return {
4262
+ getRange: (offset, length) => blob.getRange(offset, length),
4263
+ async getTail(maxBytes) {
4264
+ if (first !== void 0 && (maxBytes <= first.bytes.length || first.bytes.length === first.size)) {
4265
+ const take = Math.min(maxBytes, first.bytes.length);
4266
+ return { bytes: first.bytes.subarray(first.bytes.length - take), size: first.size };
4267
+ }
4268
+ const read = await blob.getTail(maxBytes);
4269
+ first ??= read;
4270
+ return read;
4271
+ }
4272
+ };
4273
+ }
4274
+ async function openRollbackTarget(storage, key, crypto) {
4275
+ const blob = rememberingTail(storageBlobReader(storage, key));
4276
+ await blob.getTail(DEFAULT_TAIL_BYTES);
4277
+ const encrypted = await footerSaysEncrypted(blob);
4278
+ if (encrypted !== (crypto !== void 0)) return { encrypted, reader: void 0 };
4279
+ const reader = await CrbmReader.open(blob, { crypto });
4280
+ if (reader.generation !== key.generation) {
4281
+ throw new IntegrityError(
4282
+ `segment "${key.segment}" generation ${key.generation}: its footer says generation ${reader.generation}`
4283
+ );
4284
+ }
4285
+ return { encrypted, reader };
3335
4286
  }
3336
4287
  async function writeCrbmGenerationStream(driver, key, chunks, options = {}) {
3337
4288
  const chunkKeys = [];
3338
4289
  let cardinality = 0;
3339
4290
  const tick = yieldEvery(options.clock);
4291
+ let identity;
3340
4292
  const { size, sha256 } = await driver.putImmutable(key, async (sink) => {
3341
- const writer = new CrbmWriter(sink, { generation: key.generation, crypto: options.crypto });
4293
+ const writer = new CrbmWriter(sink, {
4294
+ generation: key.generation,
4295
+ crypto: options.crypto,
4296
+ metadata: options.metadata
4297
+ });
3342
4298
  for await (const { chunkKey, bitmap } of chunks) {
3343
4299
  if (bitmap.isEmpty) continue;
3344
4300
  bitmap.optimize?.();
@@ -3349,12 +4305,57 @@ async function writeCrbmGenerationStream(driver, key, chunks, options = {}) {
3349
4305
  if (pause !== null) await pause;
3350
4306
  }
3351
4307
  await writer.finish();
4308
+ identity = writer.identity;
3352
4309
  });
3353
- return { size, sha256, chunkKeys, cardinality };
4310
+ if (identity === void 0) {
4311
+ throw new IntegrityError(
4312
+ `the driver committed ${key.segment}.${key.generation} without writing it`
4313
+ );
4314
+ }
4315
+ return {
4316
+ size,
4317
+ sha256,
4318
+ chunkKeys,
4319
+ cardinality,
4320
+ fingerprint: fingerprintFor(identity.size, identity.footerCrc)
4321
+ };
4322
+ }
4323
+ function refuseCleartextOntoKey(key, record, cleartext) {
4324
+ if (cleartext !== true || record.wrappedDeks === void 0 || record.wrappedDeks.length === 0) {
4325
+ return;
4326
+ }
4327
+ throw new KeyUnavailableError(
4328
+ `segment "${key.segment}" is encrypted, and generation ${key.generation} was written in cleartext: another writer created the segment with a key while this one was writing. Nothing was published. Re-run the write with the keystore that holds the segment's key.`
4329
+ );
3354
4330
  }
4331
+ var UNANSWERED_RESENDS = 3;
4332
+ var UNANSWERED_RESEND_BASE_MS = 500;
3355
4333
  async function publishGeneration(registry, key, options = {}) {
4334
+ let fresh;
4335
+ let failedOn;
4336
+ let unanswered;
4337
+ let resends = 0;
4338
+ let sawUnanswered = false;
4339
+ let lastRecord;
3356
4340
  for (let attempt = 0; attempt < 5; attempt++) {
3357
- const record = await registry.get(key);
4341
+ const reused = attempt === 0 && options.row !== void 0;
4342
+ const record = fresh !== void 0 ? fresh : reused ? options.row ?? null : await registry.get(key);
4343
+ fresh = void 0;
4344
+ unanswered = void 0;
4345
+ lastRecord = record;
4346
+ if (failedOn !== void 0) {
4347
+ let landed;
4348
+ try {
4349
+ landed = await landedHere(record, failedOn, key, options.holdsOwnObject);
4350
+ } catch (proofErr) {
4351
+ throw outcomeUnknown(key, proofErr);
4352
+ }
4353
+ if (landed) return true;
4354
+ if (record !== null && record.status === "active" && record.currentGen === key.generation) {
4355
+ return false;
4356
+ }
4357
+ failedOn = void 0;
4358
+ }
3358
4359
  try {
3359
4360
  if (options.expectFrom !== void 0 && record?.currentGen !== options.expectFrom) {
3360
4361
  return false;
@@ -3365,103 +4366,138 @@ async function publishGeneration(registry, key, options = {}) {
3365
4366
  if (options.expectToken !== void 0 && record?.token !== options.expectToken) {
3366
4367
  return false;
3367
4368
  }
4369
+ if (record === null && options.summary !== void 0 && "sealed" in options.summary && options.wrappedDeks === void 0) {
4370
+ if (reused) continue;
4371
+ return false;
4372
+ }
3368
4373
  if (record === null) {
3369
4374
  await registry.create(key, {
3370
4375
  currentGen: key.generation,
3371
- wrappedDeks: options.wrappedDeks
4376
+ wrappedDeks: options.wrappedDeks,
4377
+ ...options.summary === void 0 ? {} : { summary: options.summary }
3372
4378
  });
3373
4379
  } else if (record.status === "destroyed") {
3374
4380
  throw new ValidationError(
3375
4381
  `segment "${key.segment}" was destroyed (crypto-shredded) while generation ${key.generation} was being written \u2014 refusing to publish it; the written object is unreadable. Use a new segment.`
3376
4382
  );
3377
4383
  } else if (record.currentGen === null) {
4384
+ refuseCleartextOntoKey(key, record, options.cleartext);
3378
4385
  await registry.compareAndSwap(key, record.token, {
3379
4386
  currentGen: key.generation,
3380
- ...options.wrappedDeks === void 0 ? {} : { wrappedDeks: options.wrappedDeks }
4387
+ ...options.wrappedDeks === void 0 ? {} : { wrappedDeks: options.wrappedDeks },
4388
+ ...options.summary === void 0 ? {} : { summary: options.summary }
3381
4389
  });
3382
4390
  } else if (record.currentGen > key.generation) {
3383
4391
  return false;
3384
4392
  } else if (record.currentGen === key.generation) {
4393
+ if (reused) continue;
3385
4394
  return true;
3386
4395
  } else {
3387
4396
  if (options.wrappedDeks !== void 0 && options.wrappedDeks.length > 0) {
3388
4397
  throw new ValidationError(
3389
- `publishGeneration: refusing to publish generation ${key.generation} of "${key.segment}" with new key material onto a segment that already has generation ${record.currentGen} \u2014 its existing generations are not encrypted under this key. Encrypt a new segment and load into that instead.`
4398
+ `publishGeneration: refusing to publish generation ${key.generation} of "${key.segment}" with new key material onto a segment that already has generation ${record.currentGen}: its generations are not encrypted under this key, and nothing was published. If another writer published to the segment while this one was writing, re-run the write, which then uses the segment's own key; if the segment's generations are cleartext, encrypt a new segment and load into that instead.`
3390
4399
  );
3391
4400
  }
3392
- await registry.compareAndSwap(key, record.token, { currentGen: key.generation });
4401
+ refuseCleartextOntoKey(key, record, options.cleartext);
4402
+ await registry.compareAndSwap(key, record.token, {
4403
+ currentGen: key.generation,
4404
+ ...options.summary === void 0 ? {} : { summary: options.summary }
4405
+ });
3393
4406
  }
3394
4407
  return true;
3395
4408
  } catch (err) {
3396
- if (isWriteConflictError(err)) continue;
3397
- throw err;
4409
+ if (isWriteConflictError(err)) {
4410
+ failedOn = record;
4411
+ continue;
4412
+ }
4413
+ if (!isTransientError(err)) throw err;
4414
+ options.onUnanswered?.();
4415
+ sawUnanswered = true;
4416
+ let now;
4417
+ try {
4418
+ now = await registry.get(key);
4419
+ } catch (readErr) {
4420
+ if (isIntegrityError(readErr)) throw readErr;
4421
+ throw outcomeUnknown(key, err);
4422
+ }
4423
+ if (sameRow(now, record)) {
4424
+ const sleep = options.clock?.sleep;
4425
+ if (sleep === void 0 || resends >= UNANSWERED_RESENDS) throw outcomeUnknown(key, err);
4426
+ const bound = UNANSWERED_RESEND_BASE_MS * 2 ** resends;
4427
+ resends += 1;
4428
+ await sleep.call(options.clock, Math.floor((options.rng?.next() ?? 1) * bound));
4429
+ unanswered = outcomeUnknown(key, err);
4430
+ fresh = now;
4431
+ continue;
4432
+ }
4433
+ failedOn = record;
4434
+ fresh = now;
3398
4435
  }
3399
4436
  }
3400
4437
  const final = await registry.get(key);
3401
- if (final !== null && final.currentGen !== null && final.currentGen >= key.generation)
3402
- return final.currentGen === key.generation;
4438
+ if (final !== null && final.currentGen !== null && final.currentGen >= key.generation) {
4439
+ if (final.currentGen !== key.generation) return false;
4440
+ if (!sawUnanswered) return true;
4441
+ try {
4442
+ return await landedHere(final, lastRecord ?? null, key, options.holdsOwnObject);
4443
+ } catch (proofErr) {
4444
+ throw outcomeUnknown(key, proofErr);
4445
+ }
4446
+ }
4447
+ if (unanswered !== void 0) throw unanswered;
3403
4448
  throw new WriteConflictError(
3404
4449
  `publishGeneration: contention setting currentGen for ${key.segment}`
3405
4450
  );
3406
4451
  }
4452
+ async function landedHere(now, before, key, holdsOwnObject) {
4453
+ if (now === null || now.status === "destroyed" || now.currentGen !== key.generation) return false;
4454
+ if (before !== null && !sameIncarnation(before, now)) return false;
4455
+ if (holdsOwnObject === void 0) {
4456
+ throw new TransientError(
4457
+ `publish of "${key.segment}" generation ${key.generation}: whether its registry write landed could not be settled, since the caller gave no way to tell its own object from another's. Its object is kept; re-run the write, which numbers past it.`
4458
+ );
4459
+ }
4460
+ return holdsOwnObject();
4461
+ }
4462
+ function sameRow(a, b) {
4463
+ return a === null ? b === null : b !== null && a.token === b.token;
4464
+ }
4465
+ function outcomeUnknown(key, cause) {
4466
+ if (isTransientError(cause)) return cause;
4467
+ return new TransientError(
4468
+ `publish of "${key.segment}" generation ${key.generation}: whether its registry write landed could not be settled (${cause?.name ?? "error"} while checking its object). Its object is kept, since the write may still land; re-run the write, which numbers past it.`,
4469
+ { cause }
4470
+ );
4471
+ }
4472
+ async function provesOwnObject(storage, key, fingerprint) {
4473
+ try {
4474
+ return await CrbmReader.sameObject(storageBlobReader(storage, key), fingerprint);
4475
+ } catch (err) {
4476
+ if (isNotFoundError(err)) return false;
4477
+ throw err;
4478
+ }
4479
+ }
3407
4480
  async function bulkLoadCrbmGeneration(driver, key, ids, options = {}) {
3408
4481
  if (options.keystore === void 0 && options.requireEncryption === true) {
3409
4482
  throw new ValidationError("requireEncryption: a load needs a keystore to write encrypted");
3410
4483
  }
3411
4484
  const codec = requireCodec(options.codec, "bulkLoadCrbmGeneration");
3412
- const byChunk = /* @__PURE__ */ new Map();
3413
- const pendingByChunk = /* @__PURE__ */ new Map();
3414
- let pending = 0;
3415
- const tickChunk = yieldEvery(options.clock);
3416
- const tickId = yieldEvery(options.clock, YIELD_EVERY_IDS);
3417
- const flushPending = async () => {
3418
- for (const [chunkKey, rems] of pendingByChunk) {
3419
- if (rems.length === 0) continue;
3420
- const existingBitmap = byChunk.get(chunkKey);
3421
- if (existingBitmap === void 0) byChunk.set(chunkKey, codec.fromValues(rems));
3422
- else existingBitmap.addMany(rems);
3423
- rems.length = 0;
3424
- const pause = tickChunk();
3425
- if (pause !== null) await pause;
3426
- }
3427
- pending = 0;
3428
- };
3429
- const ingest = (id) => {
3430
- const { chunkKey, remainder } = splitId(id);
3431
- let bucket = pendingByChunk.get(chunkKey);
3432
- if (bucket === void 0) {
3433
- bucket = [];
3434
- pendingByChunk.set(chunkKey, bucket);
3435
- }
3436
- bucket.push(remainder);
3437
- return ++pending >= BULK_FLUSH_IDS;
3438
- };
3439
- if (Symbol.asyncIterator in ids) {
3440
- for await (const id of ids) {
3441
- if (ingest(id)) await flushPending();
3442
- const pause = tickId();
3443
- if (pause !== null) await pause;
3444
- }
4485
+ let chunks;
4486
+ if (ids instanceof DecodedLoadInput && ids.bitmap.encodeChunks !== void 0) {
4487
+ const pause = yieldEvery(options.clock, 1);
4488
+ await pause();
4489
+ chunks = ids.bitmap.encodeChunks();
4490
+ await pause();
3445
4491
  } else {
3446
- for (const id of ids) {
3447
- if (ingest(id)) await flushPending();
3448
- const pause = tickId();
3449
- if (pause !== null) await pause;
3450
- }
3451
- }
3452
- await flushPending();
3453
- let cardinality = 0;
3454
- const chunks = [];
3455
- for (const [chunkKey, bitmap] of byChunk) {
3456
- cardinality += bitmap.size;
3457
- chunks.push({ chunkKey, bitmap });
3458
- const pause = tickChunk();
3459
- if (pause !== null) await pause;
4492
+ chunks = encodeEach(
4493
+ await bucketIds(ids instanceof DecodedLoadInput ? ids.bitmap : ids, codec, options.clock)
4494
+ );
3460
4495
  }
3461
4496
  if (options.keystore !== void 0 && options.registry === void 0) {
3462
4497
  throw new ValidationError("an encrypted load requires a registry to store the wrapped DEK");
3463
4498
  }
3464
- const existing = options.registry !== void 0 ? await options.registry.get(key) : null;
4499
+ const passed = options.row;
4500
+ const existing = passed !== void 0 && passed !== null && (passed.wrappedDeks === void 0 || passed.wrappedDeks.length === 0) ? passed : options.registry !== void 0 ? await options.registry.get(key) : null;
3465
4501
  if (existing?.status === "destroyed") {
3466
4502
  throw new ValidationError(
3467
4503
  `segment "${key.segment}" is destroyed (crypto-shredded) \u2014 refusing to write; use a new segment`
@@ -3490,13 +4526,25 @@ async function bulkLoadCrbmGeneration(driver, key, ids, options = {}) {
3490
4526
  crypto = { aead: minted.aead, aadFor: (scope) => aadFor(key, key.generation, scope) };
3491
4527
  }
3492
4528
  }
3493
- const { size, sha256 } = await writeCrbmGeneration(driver, key, chunks, {
3494
- crypto,
3495
- clock: options.clock
3496
- });
4529
+ const { size, sha256, chunkCount, cardinality, fingerprint } = await writeEncodedChunks(
4530
+ driver,
4531
+ key,
4532
+ chunks,
4533
+ { crypto, clock: options.clock, metadata: options.metadata }
4534
+ );
4535
+ const summary = summaryOf(
4536
+ key,
4537
+ key.generation,
4538
+ { cardinality, metadata: options.metadata },
4539
+ crypto?.aead
4540
+ );
3497
4541
  if (options.registry !== void 0 && options.publish !== false) {
3498
4542
  const becameCurrent = await publishGeneration(options.registry, key, {
3499
- wrappedDeks: newWrapped
4543
+ wrappedDeks: newWrapped,
4544
+ summary,
4545
+ cleartext: crypto === void 0,
4546
+ holdsOwnObject: () => provesOwnObject(driver, key, fingerprint),
4547
+ clock: options.clock
3500
4548
  });
3501
4549
  if (becameCurrent) {
3502
4550
  safeAudit(options.audit ?? NOOP_AUDIT).onEvent({
@@ -3506,16 +4554,72 @@ async function bulkLoadCrbmGeneration(driver, key, ids, options = {}) {
3506
4554
  generation: key.generation
3507
4555
  });
3508
4556
  }
3509
- return {
3510
- size,
3511
- sha256,
3512
- chunkCount: chunks.length,
3513
- cardinality,
3514
- becameCurrent,
3515
- wrappedDeks: newWrapped
3516
- };
4557
+ return {
4558
+ size,
4559
+ sha256,
4560
+ chunkCount,
4561
+ cardinality,
4562
+ summary,
4563
+ becameCurrent,
4564
+ wrappedDeks: newWrapped,
4565
+ encrypted: crypto !== void 0,
4566
+ fingerprint
4567
+ };
4568
+ }
4569
+ return {
4570
+ size,
4571
+ sha256,
4572
+ chunkCount,
4573
+ cardinality,
4574
+ summary,
4575
+ wrappedDeks: newWrapped,
4576
+ encrypted: crypto !== void 0,
4577
+ fingerprint
4578
+ };
4579
+ }
4580
+ async function bucketIds(ids, codec, clock) {
4581
+ const byChunk = /* @__PURE__ */ new Map();
4582
+ const pendingByChunk = /* @__PURE__ */ new Map();
4583
+ let pending = 0;
4584
+ const tickChunk = yieldEvery(clock);
4585
+ const tickId = yieldEvery(clock, YIELD_EVERY_IDS);
4586
+ const flushPending = async () => {
4587
+ for (const [chunkKey, rems] of pendingByChunk) {
4588
+ if (rems.length === 0) continue;
4589
+ const existingBitmap = byChunk.get(chunkKey);
4590
+ if (existingBitmap === void 0) byChunk.set(chunkKey, codec.fromValues(rems));
4591
+ else existingBitmap.addMany(rems);
4592
+ rems.length = 0;
4593
+ const pause = tickChunk();
4594
+ if (pause !== null) await pause;
4595
+ }
4596
+ pending = 0;
4597
+ };
4598
+ const ingest = (id) => {
4599
+ const { chunkKey, remainder } = splitId(id);
4600
+ let bucket = pendingByChunk.get(chunkKey);
4601
+ if (bucket === void 0) {
4602
+ bucket = [];
4603
+ pendingByChunk.set(chunkKey, bucket);
4604
+ }
4605
+ bucket.push(remainder);
4606
+ return ++pending >= BULK_FLUSH_IDS;
4607
+ };
4608
+ if (Symbol.asyncIterator in ids) {
4609
+ for await (const id of ids) {
4610
+ if (ingest(id)) await flushPending();
4611
+ const pause = tickId();
4612
+ if (pause !== null) await pause;
4613
+ }
4614
+ } else {
4615
+ for (const id of ids) {
4616
+ if (ingest(id)) await flushPending();
4617
+ const pause = tickId();
4618
+ if (pause !== null) await pause;
4619
+ }
3517
4620
  }
3518
- return { size, sha256, chunkCount: chunks.length, cardinality, wrappedDeks: newWrapped };
4621
+ await flushPending();
4622
+ return [...byChunk].map(([chunkKey, bitmap]) => ({ chunkKey, bitmap })).sort((a, b) => a.chunkKey - b.chunkKey);
3519
4623
  }
3520
4624
  function openGenerationReader(storage, key, crypto, options = {}) {
3521
4625
  return openChecked(storage, key, { ...options, crypto });
@@ -3544,6 +4648,11 @@ async function verifyGeneration(storage, key, expected, crypto) {
3544
4648
  `verify failed for ${key.segment}.${key.generation}: cardinality ${reader.count()} != ${expected.cardinality}`
3545
4649
  );
3546
4650
  }
4651
+ if (!summaryAgrees(expected, { cardinality: reader.count(), metadata: reader.metadata })) {
4652
+ throw new IntegrityError(
4653
+ `verify failed for ${key.segment}.${key.generation}: the object's metadata is not the metadata it was written with`
4654
+ );
4655
+ }
3547
4656
  }
3548
4657
 
3549
4658
  // src/core/pinned-storage-source.ts
@@ -3592,6 +4701,11 @@ var PinnedStorageChunkSource = class {
3592
4701
  if (pin === void 0) return this.inner.cardinalities(ref);
3593
4702
  return pin.generation === null ? Promise.resolve(null) : this.inner.cardinalitiesAt(ref, pin.generation, heldBy(pin));
3594
4703
  }
4704
+ summary(ref) {
4705
+ const pin = this.pinFor(ref);
4706
+ if (pin === void 0) return this.inner.summary(ref);
4707
+ return pin.generation === null ? Promise.resolve(null) : this.inner.summaryAt(ref, pin.generation, heldBy(pin));
4708
+ }
3595
4709
  currentGeneration(ref) {
3596
4710
  const pin = this.pinFor(ref);
3597
4711
  return pin === void 0 ? this.inner.currentGeneration(ref) : Promise.resolve(pin.generation);
@@ -3628,67 +4742,58 @@ var PinnedStorageChunkSource = class {
3628
4742
  }
3629
4743
  };
3630
4744
 
3631
- // src/core/generation-gc.ts
3632
- async function nextGeneration(ref, deps) {
3633
- const record = await deps.registry.get(ref);
3634
- let highest = record?.currentGen ?? -1;
3635
- for await (const key of deps.storage.list(ref)) {
3636
- if (key.generation > highest) highest = key.generation;
3637
- }
3638
- return highest + 1;
4745
+ // src/core/retry.ts
4746
+ var DEFAULT_RETRY_POLICY = {
4747
+ maxAttempts: 4,
4748
+ baseDelayMs: 50,
4749
+ maxDelayMs: 2e3,
4750
+ backoffFactor: 2,
4751
+ jitter: "full"
4752
+ };
4753
+ function backoffDelayMs(policy, attempt) {
4754
+ const raw = policy.baseDelayMs * policy.backoffFactor ** (attempt - 1);
4755
+ return Math.min(policy.maxDelayMs, raw);
3639
4756
  }
3640
- async function gcOrphanGenerations(ref, deps, options = {}) {
3641
- const keep = options.keep ?? 1;
3642
- if (!Number.isInteger(keep) || keep < 0) {
3643
- throw new ValidationError(`keep must be a non-negative integer; got ${String(keep)}`);
3644
- }
3645
- const record = await deps.registry.get(ref);
3646
- if (record === null) return [];
3647
- const current = record.currentGen;
3648
- const seen = /* @__PURE__ */ new Set();
3649
- for await (const key of deps.storage.list(ref)) seen.add(key.generation);
3650
- const gens = [...seen];
3651
- const after = await deps.registry.get(ref);
3652
- if (after === null)
3653
- throw new WriteConflictError(`registry row for segment ${ref.segment} was purged mid-pass`);
3654
- if (record.status === "destroyed" && after.token !== record.token) {
3655
- throw new WriteConflictError(
3656
- `segment ${ref.segment} changed incarnation while its generations were being listed`
4757
+ function applyJitter(policy, delayMs, rng) {
4758
+ if (policy.jitter === "none") return delayMs;
4759
+ return rng.next() * delayMs;
4760
+ }
4761
+ function retryRead(op, retry) {
4762
+ return retry === void 0 ? op() : withRetry(op, retry.policy ?? DEFAULT_RETRY_POLICY, retry);
4763
+ }
4764
+ async function withRetry(op, policy, deps) {
4765
+ const retryable = deps.isRetryable ?? isTransientError;
4766
+ if (!Number.isFinite(policy.maxAttempts) || policy.maxAttempts < 1) {
4767
+ throw new ValidationError(
4768
+ `retry.maxAttempts must be a finite number >= 1 (got ${String(policy.maxAttempts)})`
3657
4769
  );
3658
4770
  }
3659
- const cutoff = current === null || after.currentGen === null ? null : Math.min(current, after.currentGen);
3660
- const stillCollectable = async () => {
3661
- const still = await deps.registry.get(ref);
3662
- if (still === null) {
3663
- throw new WriteConflictError(`registry row for segment ${ref.segment} was purged mid-pass`);
3664
- }
3665
- const ok = record.status === "destroyed" ? still.token === after.token : cutoff === null || still.currentGen !== null && still.currentGen >= cutoff;
3666
- if (!ok) {
3667
- throw new WriteConflictError(
3668
- `segment ${ref.segment} changed incarnation while its generations were being collected`
3669
- );
4771
+ const attempts = Math.floor(policy.maxAttempts);
4772
+ let lastErr;
4773
+ for (let attempt = 1; attempt <= attempts; attempt++) {
4774
+ try {
4775
+ return await op();
4776
+ } catch (err) {
4777
+ lastErr = err;
4778
+ if (attempt >= attempts || !retryable(err)) throw err;
4779
+ const delayMs = applyJitter(policy, backoffDelayMs(policy, attempt), deps.rng);
4780
+ try {
4781
+ deps.onRetry?.({ attempt, delayMs, err });
4782
+ } catch {
4783
+ }
4784
+ await deps.clock.sleep(delayMs);
3670
4785
  }
3671
- };
3672
- const toDelete = record.status === "destroyed" ? gens.sort((a, b) => a - b) : cutoff === null ? (
3673
- // No Storage pointer yet, so "below current" selects nothing and there is nothing safe to infer: an object
3674
- // here is either a load about to publish or an orphan we cannot tell apart from it. Deleting would race
3675
- // that publish into a dangling pointer. It is collected once a pointer exists.
3676
- []
3677
- ) : (
3678
- // Delete generations below the cutoff, except the newest `keep` of them (the grace window).
3679
- gens.filter((g) => g < cutoff).sort((a, b) => b - a).slice(keep)
3680
- );
3681
- for (const generation of toDelete) {
3682
- await stillCollectable();
3683
- await deps.storage.delete({ namespace: ref.namespace, segment: ref.segment, generation });
3684
4786
  }
3685
- return toDelete;
4787
+ throw lastErr;
3686
4788
  }
3687
4789
 
3688
4790
  // src/core/load.ts
3689
4791
  async function currentCardinality(ref, deps, record) {
3690
- if (record === null || record.currentGen === null || record.status === "destroyed") return null;
4792
+ if (record === null || record.currentGen === null || record.status === "destroyed") {
4793
+ return { cardinality: null, fromSummary: false };
4794
+ }
3691
4795
  const generation = record.currentGen;
4796
+ let aead;
3692
4797
  let crypto;
3693
4798
  const wrapped = record.wrappedDeks;
3694
4799
  if (wrapped !== void 0 && wrapped.length > 0) {
@@ -3697,23 +4802,29 @@ async function currentCardinality(ref, deps, record) {
3697
4802
  `segment "${ref.segment}" is encrypted but load was given no keystore`
3698
4803
  );
3699
4804
  }
3700
- const aead = await deps.keystore.openDek(wrapped);
3701
- crypto = { aead, aadFor: (scope) => aadFor(ref, generation, scope) };
4805
+ aead = await deps.keystore.openDek(wrapped);
4806
+ const opened = aead;
4807
+ crypto = { aead: opened, aadFor: (scope) => aadFor(ref, generation, scope) };
3702
4808
  }
4809
+ const described = usableSummary(ref, record, aead);
4810
+ if (described !== void 0) return { cardinality: described.cardinality, fromSummary: true };
3703
4811
  let reader;
3704
4812
  try {
3705
- reader = await openGenerationReader(deps.storage, { ...ref, generation }, crypto);
4813
+ reader = await retryRead(
4814
+ () => openGenerationReader(deps.storage, { ...ref, generation }, crypto),
4815
+ deps.readRetry
4816
+ );
3706
4817
  } catch (err) {
3707
4818
  if (!isNotFoundError(err)) throw err;
3708
- return null;
4819
+ return { cardinality: null, fromSummary: false };
3709
4820
  }
3710
4821
  let total = 0;
3711
4822
  for (const n of reader.cardinalities().values()) total += n;
3712
- return total;
4823
+ return { cardinality: total, fromSummary: false };
3713
4824
  }
3714
- async function loadSegment(ref, ids, deps, options = {}) {
4825
+ async function loadSegment(ref, input, deps, options = {}) {
3715
4826
  validateUserRef(ref);
3716
- requireCodec(deps.codec, "loadSegment");
4827
+ const codec = requireCodec(deps.codec, "loadSegment");
3717
4828
  const keep = options.keep ?? 1;
3718
4829
  if (!Number.isInteger(keep) || keep < 0) {
3719
4830
  throw new ValidationError(`keep must be a non-negative integer; got ${String(keep)}`);
@@ -3733,13 +4844,19 @@ async function loadSegment(ref, ids, deps, options = {}) {
3733
4844
  throw new ValidationError(`guard.minRetained must be a fraction in 0..1; got ${String(s)}`);
3734
4845
  }
3735
4846
  }
4847
+ const metadata = copiedMetadata(options.metadata, (message) => {
4848
+ throw new ValidationError(message);
4849
+ });
3736
4850
  const audit = safeAudit(options.audit ?? NOOP_AUDIT);
4851
+ const ids = prepareLoadInput(input, codec);
4852
+ assertRegistryCanWrite(deps.registry, "load");
3737
4853
  const row = await deps.registry.get(ref);
3738
4854
  const fromToken = row?.token;
3739
4855
  const fromGeneration = row?.currentGen ?? void 0;
3740
4856
  const needsBefore = guard?.minRetained !== void 0 || options.allowEmpty !== true;
3741
- const before = needsBefore ? await currentCardinality(ref, deps, row) : null;
3742
- const generation = await nextGeneration(ref, deps);
4857
+ const current = needsBefore ? await currentCardinality(ref, deps, row) : { cardinality: null, fromSummary: false };
4858
+ const before = current.cardinality;
4859
+ const { generation, checked } = await nextLoadGeneration(ref, deps, row);
3743
4860
  const key = { namespace: ref.namespace, segment: ref.segment, generation };
3744
4861
  let written;
3745
4862
  try {
@@ -3749,7 +4866,9 @@ async function loadSegment(ref, ids, deps, options = {}) {
3749
4866
  keystore: deps.keystore,
3750
4867
  requireEncryption: deps.requireEncryption,
3751
4868
  codec: deps.codec,
3752
- clock: deps.clock
4869
+ clock: deps.clock,
4870
+ metadata,
4871
+ row
3753
4872
  });
3754
4873
  } catch (err) {
3755
4874
  if (!isWriteConflictError(err)) throw err;
@@ -3773,16 +4892,28 @@ async function loadSegment(ref, ids, deps, options = {}) {
3773
4892
  collected: []
3774
4893
  };
3775
4894
  }
3776
- const refuse = async (reason) => {
4895
+ const reclaim = async () => {
3777
4896
  const now = await deps.registry.get(ref);
3778
- if (now === null || now.token === fromToken) await deps.storage.delete(key);
4897
+ if (now !== null && (now.token === fromToken || now.status === "destroyed")) {
4898
+ await deps.storage.delete(key);
4899
+ return;
4900
+ }
4901
+ const keyed = now !== null && now.wrappedDeks !== void 0 && now.wrappedDeks.length > 0;
4902
+ if ((now === null || keyed && !written.encrypted) && await holdsObject(deps.storage, key, written.fingerprint)) {
4903
+ await deps.storage.delete(key);
4904
+ }
4905
+ };
4906
+ let unanswered = false;
4907
+ const refuse = async (reason) => {
4908
+ await reclaim();
3779
4909
  audit.onEvent({
3780
4910
  kind: "segment.load-refused",
3781
4911
  segment: ref.segment,
3782
4912
  namespace: ref.namespace,
3783
4913
  generation,
3784
4914
  reason,
3785
- cardinality: written.cardinality
4915
+ cardinality: written.cardinality,
4916
+ ...unanswered ? { unanswered: true } : {}
3786
4917
  });
3787
4918
  return {
3788
4919
  generation,
@@ -3805,19 +4936,37 @@ async function loadSegment(ref, ids, deps, options = {}) {
3805
4936
  if (guard?.minRetained !== void 0 && before !== null) {
3806
4937
  if (written.cardinality < before * guard.minRetained) return refuse("min-retained");
3807
4938
  }
3808
- const published = await publishGeneration(deps.registry, key, {
3809
- wrappedDeks: written.wrappedDeks,
3810
- ...fromToken === void 0 ? {} : { expectToken: fromToken },
3811
- ...needsBefore && fromGeneration !== void 0 ? { expectFrom: fromGeneration } : {},
3812
- // The third case, and the one the two fences above structurally cannot cover: the guard judged a segment
3813
- // that had NO ROW. Both `expectFrom` and `expectToken` compare against a value read from a row, so with no
3814
- // row both are omitted and the publish becomes a bare forward-only advance — which lands over anything
3815
- // that appeared in between. `before` was `null`, so the empty and `minRetained` bounds had nothing to
3816
- // judge and passed vacuously. Verified: an empty generation published over a thousand ids that a
3817
- // concurrent writer had created meanwhile, reporting success. A guarded write therefore has to fence on
3818
- // the ABSENCE it relied on, exactly as it fences on the pointer it relied on.
3819
- ...needsBefore && row === null ? { expectAbsent: true } : {}
3820
- });
4939
+ let published;
4940
+ try {
4941
+ published = await publishGeneration(deps.registry, key, {
4942
+ row,
4943
+ wrappedDeks: written.wrappedDeks,
4944
+ summary: written.summary,
4945
+ cleartext: !written.encrypted,
4946
+ ...fromToken === void 0 ? {} : { expectToken: fromToken },
4947
+ ...needsBefore && fromGeneration !== void 0 ? { expectFrom: fromGeneration } : {},
4948
+ // The third case, and the one the two fences above structurally cannot cover: the guard judged a segment
4949
+ // that had NO ROW. Both `expectFrom` and `expectToken` compare against a value read from a row, so with no
4950
+ // row both are omitted and the publish becomes a bare forward-only advance — which lands over anything
4951
+ // that appeared in between. `before` was `null`, so the empty and `minRetained` bounds had nothing to
4952
+ // judge and passed vacuously. Verified: an empty generation published over a thousand ids that a
4953
+ // concurrent writer had created meanwhile, reporting success. A guarded write therefore has to fence on
4954
+ // the ABSENCE it relied on, exactly as it fences on the pointer it relied on.
4955
+ ...needsBefore && row === null ? { expectAbsent: true } : {},
4956
+ // A registry write that fails without an answer is reconciled by reading the row: the pointer at this number
4957
+ // is this load's publish only over the object this load wrote, which one footer read proves.
4958
+ holdsOwnObject: () => provesOwnObject(deps.storage, key, written.fingerprint),
4959
+ // And the wait before a fresh write, when the first left the row as it was.
4960
+ clock: deps.clock,
4961
+ rng: deps.rng ?? deps.readRetry?.rng,
4962
+ onUnanswered: () => {
4963
+ unanswered = true;
4964
+ }
4965
+ });
4966
+ } catch (err) {
4967
+ if (isValidationError(err) || err instanceof KeyUnavailableError) await reclaim();
4968
+ throw err;
4969
+ }
3821
4970
  if (!published) return refuse("superseded");
3822
4971
  audit.onEvent({
3823
4972
  kind: "segment.publish",
@@ -3825,7 +4974,13 @@ async function loadSegment(ref, ids, deps, options = {}) {
3825
4974
  namespace: ref.namespace,
3826
4975
  generation
3827
4976
  });
3828
- const collected = await gcOrphanGenerations(ref, deps, { keep });
4977
+ const currentObjectGone = needsBefore && before === null;
4978
+ const collected = await collectAfterLoad(ref, deps, {
4979
+ generation,
4980
+ keep,
4981
+ byName: checked && !currentObjectGone && deps.collectByListing !== true,
4982
+ ...current.fromSummary && fromGeneration !== void 0 ? { proveCurrent: { ...ref, generation: fromGeneration } } : {}
4983
+ });
3829
4984
  return {
3830
4985
  generation,
3831
4986
  published: true,
@@ -3845,7 +5000,25 @@ async function listGenerations(ref, deps) {
3845
5000
  const current = record?.currentGen ?? null;
3846
5001
  const seen = /* @__PURE__ */ new Set();
3847
5002
  for await (const key of deps.storage.list(ref)) seen.add(key.generation);
3848
- return [...seen].sort((a, b) => a - b).map((generation) => ({ generation, current: generation === current }));
5003
+ const described = record === null ? void 0 : await currentDescription(ref, record, deps);
5004
+ return [...seen].sort((a, b) => a - b).map(
5005
+ (generation) => generation === current ? { generation, current: true, ...described } : { generation, current: false }
5006
+ );
5007
+ }
5008
+ async function currentDescription(ref, record, deps) {
5009
+ if (record.summary === void 0) return void 0;
5010
+ let aead;
5011
+ if ("sealed" in record.summary && record.wrappedDeks !== void 0 && deps.keystore !== void 0) {
5012
+ try {
5013
+ aead = await deps.keystore.openDek(record.wrappedDeks);
5014
+ } catch {
5015
+ return void 0;
5016
+ }
5017
+ }
5018
+ const usable = usableSummary(ref, record, aead);
5019
+ if (usable === void 0) return void 0;
5020
+ const metadata = usable.metadata;
5021
+ return metadata === void 0 || Object.keys(metadata).length === 0 ? { cardinality: usable.cardinality } : { cardinality: usable.cardinality, metadata: Object.freeze({ ...metadata }) };
3849
5022
  }
3850
5023
  async function rollbackSegment(ref, toGeneration, deps, options = {}) {
3851
5024
  validateUserRef(ref);
@@ -3879,8 +5052,34 @@ async function rollbackSegment(ref, toGeneration, deps, options = {}) {
3879
5052
  `rollback: generation ${toGeneration} of "${ref.segment}" is above the current pointer (${record.currentGen}) \u2014 it may never have been published. Pass { allowForward: true } if you are undoing an earlier rollback.`
3880
5053
  );
3881
5054
  }
5055
+ const keyed = record.wrappedDeks !== void 0 && record.wrappedDeks.length > 0;
5056
+ let aead;
5057
+ if (keyed && deps.keystore !== void 0) {
5058
+ try {
5059
+ aead = await deps.keystore.openDek(record.wrappedDeks);
5060
+ } catch {
5061
+ }
5062
+ }
5063
+ const crypto = aead === void 0 ? void 0 : { aead, aadFor: (scope) => aadFor(ref, toGeneration, scope) };
5064
+ const { encrypted, reader } = await openRollbackTarget(
5065
+ deps.storage,
5066
+ { ...ref, generation: toGeneration },
5067
+ crypto
5068
+ );
5069
+ if (encrypted !== keyed) {
5070
+ throw new IntegrityError(
5071
+ `rollback: generation ${toGeneration} of "${ref.segment}" is ${encrypted ? "encrypted" : "cleartext"}, but the segment is ${keyed ? "encrypted" : "cleartext"}, so no read of the segment could use it; it was never one of its generations. Roll back to another generation.`
5072
+ );
5073
+ }
5074
+ const summary = reader === void 0 ? void 0 : summaryOf(
5075
+ ref,
5076
+ toGeneration,
5077
+ { cardinality: reader.count(), metadata: reader.metadata },
5078
+ aead
5079
+ );
3882
5080
  const { token } = await deps.registry.compareAndSwap(ref, record.token, {
3883
- currentGen: toGeneration
5081
+ currentGen: toGeneration,
5082
+ summary
3884
5083
  });
3885
5084
  let stillThere = false;
3886
5085
  for await (const key of deps.storage.list(ref)) {
@@ -3889,15 +5088,31 @@ async function rollbackSegment(ref, toGeneration, deps, options = {}) {
3889
5088
  break;
3890
5089
  }
3891
5090
  }
3892
- if (!stillThere) {
5091
+ let replaced = false;
5092
+ if (stillThere && reader !== void 0 && toGeneration > (record.currentGen ?? -1)) {
5093
+ try {
5094
+ replaced = !await provesOwnObject(
5095
+ deps.storage,
5096
+ { ...ref, generation: toGeneration },
5097
+ reader.fingerprint
5098
+ );
5099
+ } catch (err) {
5100
+ if (!isIntegrityError(err)) throw err;
5101
+ replaced = true;
5102
+ }
5103
+ }
5104
+ if (!stillThere || replaced) {
3893
5105
  let undone = false;
3894
5106
  try {
3895
- await deps.registry.compareAndSwap(ref, token, { currentGen: record.currentGen });
5107
+ await deps.registry.compareAndSwap(ref, token, {
5108
+ currentGen: record.currentGen,
5109
+ summary: describingSummary(record)
5110
+ });
3896
5111
  undone = true;
3897
5112
  } catch {
3898
5113
  }
3899
5114
  throw new NotFoundError(
3900
- `rollback: generation ${toGeneration} of "${ref.segment}" was collected while the pointer was moving` + (undone ? " \u2014 the pointer was put back" : `, and the move back failed, so the pointer may still name ${toGeneration} (a write that landed and lost its response reads as a failure here). Check which generation is current, then re-run a rollback to a generation that exists, or restore the object.`)
5115
+ `rollback: generation ${toGeneration} of "${ref.segment}" was ${replaced ? "replaced" : "collected"} while the pointer was moving` + (undone ? " \u2014 the pointer was put back" : `, and the move back failed, so the pointer may still name ${toGeneration} (a write that landed and lost its response reads as a failure here). Check which generation is current, then re-run a rollback to a generation that exists, or restore the object.`)
3901
5116
  );
3902
5117
  }
3903
5118
  safeAudit(options.audit ?? NOOP_AUDIT).onEvent({
@@ -3909,15 +5124,21 @@ async function rollbackSegment(ref, toGeneration, deps, options = {}) {
3909
5124
  });
3910
5125
  return { fromGeneration: record.currentGen, generation: toGeneration };
3911
5126
  }
5127
+ function describingSummary(record) {
5128
+ const { summary } = record;
5129
+ if (summary === void 0 || summary.generation !== record.currentGen) return void 0;
5130
+ const keyed = record.wrappedDeks !== void 0 && record.wrappedDeks.length > 0;
5131
+ return "sealed" in summary === keyed ? summary : void 0;
5132
+ }
3912
5133
 
3913
5134
  // src/core/erase-id.ts
3914
- var DEFAULT_MAX_BITMAP_BYTES2 = 1 << 20;
3915
5135
  async function eraseIdFromSegment(ref, id, deps, options = {}) {
3916
5136
  validateUserRef(ref);
3917
5137
  const { chunkKey, remainder } = splitId(id);
3918
5138
  const codec = requireCodec(deps.codec, "eraseIdFromSegment");
3919
- const maxBytes = deps.maxBitmapBytes ?? DEFAULT_MAX_BITMAP_BYTES2;
5139
+ const maxBytes = deps.maxBitmapBytes ?? DEFAULT_MAX_BITMAP_BYTES;
3920
5140
  const base = { segment: ref.segment, namespace: ref.namespace };
5141
+ const read = (op) => retryRead(op, deps.readRetry);
3921
5142
  const record = await deps.registry.get(ref);
3922
5143
  if (record === null) return { ...base, erased: false, reason: "absent", collected: [] };
3923
5144
  if (record.status === "destroyed")
@@ -3954,13 +5175,18 @@ async function eraseIdFromSegment(ref, id, deps, options = {}) {
3954
5175
  return row.currentGen === from && row.token === fromToken ? null : "superseded";
3955
5176
  };
3956
5177
  const holds = async (generation2) => {
5178
+ const key2 = { ...base, generation: generation2 };
5179
+ const chunkIn = (crypto) => read(async () => (await openGenerationReader(deps.storage, key2, crypto)).getChunk(chunkKey));
3957
5180
  try {
3958
- const reader = await openGenerationReader(
3959
- deps.storage,
3960
- { ...base, generation: generation2 },
3961
- cryptoAt(generation2)
3962
- );
3963
- const bytes = await reader.getChunk(chunkKey);
5181
+ let bytes;
5182
+ try {
5183
+ bytes = await chunkIn(cryptoAt(generation2));
5184
+ } catch (err) {
5185
+ if (!isIntegrityError(err) || cryptoAt(generation2) === void 0 || await read(() => objectIsEncrypted(deps.storage, key2))) {
5186
+ throw err;
5187
+ }
5188
+ bytes = await chunkIn(void 0);
5189
+ }
3964
5190
  return bytes !== null && codec.safeDeserialize(bytes, maxBytes).has(remainder);
3965
5191
  } catch (err) {
3966
5192
  if (isNotFoundError(err)) return null;
@@ -4030,27 +5256,40 @@ async function eraseIdFromSegment(ref, id, deps, options = {}) {
4030
5256
  const stage = async () => {
4031
5257
  let written;
4032
5258
  try {
4033
- const reader = await openGenerationReader(deps.storage, fromKey, cryptoAt(from));
4034
- const bytes = await reader.getChunk(chunkKey);
5259
+ const reader = await read(() => openGenerationReader(deps.storage, fromKey, cryptoAt(from)));
5260
+ const bytes = await read(() => reader.getChunk(chunkKey));
4035
5261
  if (bytes === null) return await notInCurrent();
4036
5262
  const target = codec.safeDeserialize(bytes, maxBytes);
4037
5263
  assertRemaindersInRange(target, chunkKey);
4038
5264
  if (!target.has(remainder)) return await notInCurrent();
4039
5265
  target.remove(remainder);
5266
+ assertRegistryCanWrite(deps.registry, "eraseIdFromSegment");
4040
5267
  const generation2 = await nextGeneration(ref, deps);
4041
5268
  const key2 = { ...base, generation: generation2 };
5269
+ const metadata = metadataToCarry(
5270
+ reader.metadata,
5271
+ aead === void 0 ? void 0 : usableSummary(ref, record, aead)
5272
+ );
4042
5273
  const tally = await writeCrbmGenerationStream(
4043
5274
  deps.storage,
4044
5275
  key2,
4045
- rewrite(reader, chunkKey, target, codec, maxBytes),
4046
- { crypto: cryptoAt(generation2), clock: deps.clock }
5276
+ rewrite(reader, chunkKey, target, codec, maxBytes, read),
5277
+ { crypto: cryptoAt(generation2), clock: deps.clock, metadata }
4047
5278
  );
4048
5279
  written = generation2;
4049
5280
  const beforeVerify = await deps.registry.get(ref);
4050
5281
  const early = rowVerdict(beforeVerify);
4051
5282
  if (early !== null) return refused(early, written);
4052
- await verifyGeneration(deps.storage, key2, tally, cryptoAt(generation2));
4053
- return { generation: generation2, key: key2 };
5283
+ await read(
5284
+ () => verifyGeneration(deps.storage, key2, { ...tally, metadata }, cryptoAt(generation2))
5285
+ );
5286
+ const summary2 = summaryOf(
5287
+ ref,
5288
+ generation2,
5289
+ { cardinality: tally.cardinality, metadata },
5290
+ aead
5291
+ );
5292
+ return { generation: generation2, key: key2, fingerprint: tally.fingerprint, summary: summary2 };
4054
5293
  } catch (err) {
4055
5294
  if (!isNotFoundError(err)) throw err;
4056
5295
  let now;
@@ -4075,10 +5314,16 @@ async function eraseIdFromSegment(ref, id, deps, options = {}) {
4075
5314
  if (!staged.erased && staged.generation !== void 0) await discardRefused(staged.generation);
4076
5315
  return staged;
4077
5316
  }
4078
- const { generation, key } = staged;
5317
+ const { generation, key, fingerprint, summary } = staged;
4079
5318
  const published = await publishGeneration(deps.registry, key, {
4080
5319
  expectFrom: from,
4081
- expectToken: fromToken
5320
+ expectToken: fromToken,
5321
+ summary,
5322
+ // A write that ends without an answer is settled by reading the row, and a pointer at this number is this
5323
+ // rewrite's only over the object it wrote: the footer proves it, so another incarnation's cannot pass for it.
5324
+ holdsOwnObject: () => provesOwnObject(deps.storage, key, fingerprint),
5325
+ clock: deps.clock,
5326
+ rng: deps.rng ?? deps.readRetry?.rng
4082
5327
  });
4083
5328
  if (!published) {
4084
5329
  await discardRefused(generation);
@@ -4103,14 +5348,21 @@ async function eraseIdFromSegment(ref, id, deps, options = {}) {
4103
5348
  if (left !== void 0) throw cannotRemove(left);
4104
5349
  return { ...base, erased: true, fromGeneration: from, generation, collected };
4105
5350
  }
4106
- async function* rewrite(reader, chunkKey, replacement, codec, maxBytes) {
5351
+ var REWRITE_READ_AHEAD = 32;
5352
+ async function* rewrite(reader, chunkKey, replacement, codec, maxBytes, read) {
4107
5353
  const keys = [...reader.chunkKeys()].sort((a, b) => a - b);
5354
+ const window = new ChunkWindow(
5355
+ keys.filter((k) => k !== chunkKey),
5356
+ (k) => read(() => reader.getChunk(k)),
5357
+ REWRITE_READ_AHEAD,
5358
+ false
5359
+ );
4108
5360
  for (const k of keys) {
4109
5361
  if (k === chunkKey) {
4110
5362
  yield { chunkKey: k, bitmap: replacement };
4111
5363
  continue;
4112
5364
  }
4113
- const bytes = await reader.getChunk(k);
5365
+ const bytes = await window.take();
4114
5366
  if (bytes === null) continue;
4115
5367
  const bitmap = codec.safeDeserialize(bytes, maxBytes);
4116
5368
  assertRemaindersInRange(bitmap, k);
@@ -4129,8 +5381,8 @@ function assertRemaindersInRange(bitmap, chunkKey) {
4129
5381
  // src/drivers/crypto.ts
4130
5382
  import { createCipheriv, createDecipheriv, randomBytes } from "node:crypto";
4131
5383
  var KEY_BYTES = 32;
4132
- var NONCE_BYTES = 12;
4133
- var TAG_BYTES = 16;
5384
+ var NONCE_BYTES2 = 12;
5385
+ var TAG_BYTES2 = 16;
4134
5386
  var NodeAead = class {
4135
5387
  key;
4136
5388
  constructor(key) {
@@ -4139,14 +5391,14 @@ var NodeAead = class {
4139
5391
  this.key = Buffer.from(key);
4140
5392
  }
4141
5393
  seal(plaintext, aad) {
4142
- const nonce = randomBytes(NONCE_BYTES);
5394
+ const nonce = randomBytes(NONCE_BYTES2);
4143
5395
  const cipher = createCipheriv("aes-256-gcm", this.key, nonce);
4144
5396
  cipher.setAAD(aad);
4145
5397
  const ciphertext = Buffer.concat([cipher.update(plaintext), cipher.final()]);
4146
5398
  return { nonce, ciphertext, tag: cipher.getAuthTag() };
4147
5399
  }
4148
5400
  open(sealed, aad) {
4149
- if (sealed.nonce.length !== NONCE_BYTES || sealed.tag.length !== TAG_BYTES)
5401
+ if (sealed.nonce.length !== NONCE_BYTES2 || sealed.tag.length !== TAG_BYTES2)
4150
5402
  throw new IntegrityError("AEAD nonce/tag have the wrong length");
4151
5403
  const decipher = createDecipheriv("aes-256-gcm", this.key, sealed.nonce);
4152
5404
  decipher.setAAD(aad);
@@ -4166,11 +5418,11 @@ function wrapKey(dek, kek, keyId) {
4166
5418
  }
4167
5419
  function unwrapKey(wrapped, kek, keyId) {
4168
5420
  const blob = Buffer.from(wrapped, "base64");
4169
- if (blob.length < NONCE_BYTES + TAG_BYTES)
5421
+ if (blob.length < NONCE_BYTES2 + TAG_BYTES2)
4170
5422
  throw new IntegrityError("wrapped DEK is too short to be valid");
4171
- const nonce = blob.subarray(0, NONCE_BYTES);
4172
- const tag = blob.subarray(blob.length - TAG_BYTES);
4173
- const ciphertext = blob.subarray(NONCE_BYTES, blob.length - TAG_BYTES);
5423
+ const nonce = blob.subarray(0, NONCE_BYTES2);
5424
+ const tag = blob.subarray(blob.length - TAG_BYTES2);
5425
+ const ciphertext = blob.subarray(NONCE_BYTES2, blob.length - TAG_BYTES2);
4174
5426
  return new NodeAead(kek).open({ nonce, ciphertext, tag }, new TextEncoder().encode(keyId));
4175
5427
  }
4176
5428
  var InProcessKeystore = class {
@@ -4411,7 +5663,10 @@ async function shredSegment(ref, deps, allowCleartext, op = "destroySegment") {
4411
5663
  status: "destroyed",
4412
5664
  wrappedDeks: void 0,
4413
5665
  // ← the crypto-shred: the only copy of the DEK wrappings is gone
4414
- keyId: void 0
5666
+ keyId: void 0,
5667
+ // The current generation's cached count and metadata go with it: sealed, they cannot be opened without
5668
+ // the wrappings; clear, they would outlive the segment on its tombstone.
5669
+ summary: void 0
4415
5670
  });
4416
5671
  return { ...base, destroyed: true, cryptoShredded: encrypted };
4417
5672
  } catch (err) {
@@ -4550,39 +5805,108 @@ function assertNotDestroyed(record, ref) {
4550
5805
 
4551
5806
  // src/core/retention-sweep.ts
4552
5807
  var DEFAULT_RETIRE_LIMIT = 100;
5808
+ var MAX_CONSECUTIVE_PURGE_FAULTS = 3;
4553
5809
  var DEFAULT_TOMBSTONE_GRACE_MS = 864e5;
4554
5810
  var DEFAULT_LOOKBACK_BUCKETS = 7;
4555
5811
  async function rowsFromDueIndex(registry, options) {
4556
5812
  const rows = [];
4557
- const seen = /* @__PURE__ */ new Set();
5813
+ const pointers = /* @__PURE__ */ new Map();
5814
+ const litter = [];
5815
+ const gone = /* @__PURE__ */ new Set();
4558
5816
  for (const bucket of dueBucketsAt(options.now, options.lookbackBuckets)) {
4559
5817
  for await (const pointer of registry.list(dueNamespace(bucket))) {
4560
5818
  const ref = decodeDueName(pointer.segment);
4561
5819
  if (ref === null) continue;
4562
5820
  if (isReservedNamespace(ref.namespace)) continue;
4563
5821
  if (options.namespace !== void 0 && ref.namespace !== options.namespace) continue;
4564
- const key = `${ref.namespace ?? ""}\0${ref.segment}`;
4565
- if (seen.has(key)) continue;
4566
- seen.add(key);
5822
+ const found = {
5823
+ ref: { namespace: pointer.namespace, segment: pointer.segment },
5824
+ target: ref,
5825
+ token: pointer.token
5826
+ };
5827
+ const key = segmentKey(ref);
5828
+ const known = pointers.get(key);
5829
+ if (known !== void 0) {
5830
+ known.push(found);
5831
+ continue;
5832
+ }
5833
+ if (gone.has(key)) {
5834
+ litter.push(found);
5835
+ continue;
5836
+ }
4567
5837
  if (rows.length >= options.maxScanSegments) {
4568
5838
  throw new BudgetExceededError(
4569
5839
  `retireExpired: the due index yielded more than ${options.maxScanSegments} segments \u2014 the scan was abandoned there rather than completed. Raise \`maxScanSegments\`, narrow with \`namespace\`, or reduce \`lookbackBuckets\`.`
4570
5840
  );
4571
5841
  }
4572
5842
  const live = await registry.get(ref);
4573
- if (live === null) continue;
5843
+ if (live === null) {
5844
+ gone.add(key);
5845
+ litter.push(found);
5846
+ continue;
5847
+ }
4574
5848
  rows.push(live);
5849
+ pointers.set(key, [found]);
4575
5850
  }
4576
5851
  }
4577
- return rows;
5852
+ return { rows, pointers, litter };
5853
+ }
5854
+ async function rowsFromFleet(registry, options) {
5855
+ const held = [];
5856
+ const rows = await drainRegistry(registry, {
5857
+ namespace: options.namespace,
5858
+ maxScanSegments: options.maxScanSegments,
5859
+ op: "retireExpired",
5860
+ onReserved: (pointer) => {
5861
+ if (held.length >= options.maxScanSegments) return;
5862
+ const target = decodeDueName(pointer.segment);
5863
+ if (target === null || isReservedNamespace(target.namespace)) return;
5864
+ held.push({
5865
+ ref: { namespace: pointer.namespace, segment: pointer.segment },
5866
+ target,
5867
+ token: pointer.token
5868
+ });
5869
+ }
5870
+ });
5871
+ const pointers = /* @__PURE__ */ new Map();
5872
+ const litter = [];
5873
+ if (held.length === 0) return { rows, pointers, litter };
5874
+ const present = new Set(rows.map((rec) => segmentKey(rec)));
5875
+ for (const found of held) {
5876
+ const key = segmentKey(found.target);
5877
+ if (!present.has(key)) litter.push(found);
5878
+ else if (pointers.has(key)) pointers.get(key).push(found);
5879
+ else pointers.set(key, [found]);
5880
+ }
5881
+ return { rows, pointers, litter };
4578
5882
  }
4579
- async function forgetDuePointer(registry, ref, expiresAt) {
5883
+ async function forgetDuePointer(registry, ref, expiresAt, onFault) {
4580
5884
  if (!canIndex(ref)) return;
4581
5885
  try {
4582
5886
  await registry.delete(dueIndexRef(dueBucket(expiresAt), ref));
5887
+ } catch (err) {
5888
+ onFault(err);
5889
+ }
5890
+ }
5891
+ async function filePurgePointer(registry, ref, purgeAt) {
5892
+ if (!canIndex(ref)) return;
5893
+ try {
5894
+ await registry.create(dueIndexRef(dueBucket(purgeAt), ref), { currentGen: null });
4583
5895
  } catch {
4584
5896
  }
4585
5897
  }
5898
+ async function forgetPurgePointers(registry, ref, purgeAt, found, onFault) {
5899
+ if (!canIndex(ref)) return;
5900
+ const filed = dueIndexRef(dueBucket(purgeAt), ref);
5901
+ const all = found.some((p) => p.ref.namespace === filed.namespace) ? found : [...found, { ref: filed, target: ref }];
5902
+ for (const pointer of all) {
5903
+ try {
5904
+ await registry.delete(pointer.ref, pointer.token);
5905
+ } catch (err) {
5906
+ onFault(err);
5907
+ }
5908
+ }
5909
+ }
4586
5910
  async function retireExpired(deps, options) {
4587
5911
  if (options.namespace !== void 0) validateUserNamespace(options.namespace);
4588
5912
  const now = options.now;
@@ -4615,32 +5939,40 @@ async function retireExpired(deps, options) {
4615
5939
  }
4616
5940
  const dryRun = options.dryRun === true;
4617
5941
  const scan = options.scan ?? "fleet";
4618
- const rows = scan === "index" ? await rowsFromDueIndex(deps.registry, {
5942
+ const scanned = scan === "index" ? await rowsFromDueIndex(deps.registry, {
4619
5943
  now,
4620
5944
  lookbackBuckets: options.lookbackBuckets ?? DEFAULT_LOOKBACK_BUCKETS,
4621
5945
  namespace: options.namespace,
4622
5946
  maxScanSegments
4623
- }) : await drainRegistry(deps.registry, {
4624
- namespace: options.namespace,
4625
- maxScanSegments,
4626
- op: "retireExpired"
4627
- });
5947
+ }) : await rowsFromFleet(deps.registry, { namespace: options.namespace, maxScanSegments });
5948
+ const rows = scanned.rows;
5949
+ const removesRows = deps.registry.capabilities().conditionalDelete === true;
4628
5950
  const shards = options.shards;
4629
5951
  const totalShards = options.totalShards ?? 0;
4630
- const mine = shards === void 0 || totalShards <= 1 ? rows : rows.filter((r) => shards.includes(shardOf(segmentKey(r), totalShards)));
5952
+ const owned = (ref) => shards === void 0 || totalShards <= 1 || shards.includes(shardOf(segmentKey(ref), totalShards));
5953
+ const mine = rows.filter(owned);
4631
5954
  const entries = [];
4632
5955
  let eligible = 0;
4633
5956
  let retired = 0;
4634
5957
  let wouldRetire = 0;
4635
5958
  let tombstonesPurged = 0;
4636
5959
  let limited = false;
5960
+ let purgeFaults = 0;
5961
+ let firstPurgeFault;
5962
+ const noteFault = (err) => {
5963
+ if (isWriteConflictError(err)) return;
5964
+ purgeFaults += 1;
5965
+ firstPurgeFault ??= failureReason(err);
5966
+ };
5967
+ let purgeRun = 0;
5968
+ let purging = true;
4637
5969
  let attempted = 0;
4638
5970
  for (const rec of mine) {
4639
5971
  const ref = { namespace: rec.namespace, segment: rec.segment };
4640
5972
  const base = { segment: rec.segment, namespace: rec.namespace };
4641
5973
  const policy = readRetentionPolicy(rec.retention);
4642
5974
  if (rec.status === "destroyed") {
4643
- if (!purgeTombstones) continue;
5975
+ if (!purgeTombstones || !purging) continue;
4644
5976
  const retiredAt = retirementStamp(rec.retention);
4645
5977
  if (retiredAt === null) continue;
4646
5978
  if (now - retiredAt < grace) continue;
@@ -4657,13 +5989,31 @@ async function retireExpired(deps, options) {
4657
5989
  continue;
4658
5990
  }
4659
5991
  }
4660
- if (!dryRun) await deps.registry.delete(ref, rec.token);
4661
- if (!dryRun) tombstonesPurged += 1;
5992
+ if (!dryRun) {
5993
+ await deps.registry.delete(ref, rec.token);
5994
+ tombstonesPurged += 1;
5995
+ purgeRun = 0;
5996
+ if (removesRows) {
5997
+ await forgetPurgePointers(
5998
+ deps.registry,
5999
+ ref,
6000
+ retiredAt + grace,
6001
+ scanned.pointers.get(segmentKey(ref)) ?? [],
6002
+ noteFault
6003
+ );
6004
+ }
6005
+ }
4662
6006
  entries.push({
4663
6007
  ...base,
4664
6008
  action: dryRun ? "would-purge-tombstone" : "purged-tombstone"
4665
6009
  });
4666
6010
  } catch (err) {
6011
+ if (!isWriteConflictError(err)) {
6012
+ attempted -= 1;
6013
+ purgeRun += 1;
6014
+ if (purgeRun >= MAX_CONSECUTIVE_PURGE_FAULTS) purging = false;
6015
+ noteFault(err);
6016
+ }
4667
6017
  entries.push({ ...base, action: "skipped", reason: failureReason(err) });
4668
6018
  }
4669
6019
  continue;
@@ -4707,22 +6057,27 @@ async function retireExpired(deps, options) {
4707
6057
  }
4708
6058
  retired += 1;
4709
6059
  entries.push({ ...base, action: "retired", expiresAt: livePolicy.expiresAt, result });
4710
- await forgetDuePointer(deps.registry, ref, livePolicy.expiresAt);
6060
+ await forgetDuePointer(deps.registry, ref, livePolicy.expiresAt, noteFault);
4711
6061
  if (purgeTombstones && result.generationsDeleted.length === 0 && result.generationsRemaining.length === 0) {
4712
6062
  const tombstone = await deps.registry.get(ref).catch(() => null);
4713
6063
  if (tombstone?.status === "destroyed") {
4714
6064
  const deleted = await deps.registry.delete(ref, tombstone.token).then(
4715
6065
  () => true,
4716
- () => false
6066
+ (err) => {
6067
+ noteFault(err);
6068
+ return false;
6069
+ }
4717
6070
  );
4718
6071
  if (deleted) continue;
4719
6072
  }
4720
6073
  }
4721
- await stampRetirement(deps.registry, ref, now).catch(() => void 0);
6074
+ const stamped = await stampRetirement(deps.registry, ref, now).catch(() => false);
6075
+ if (stamped && removesRows) await filePurgePointer(deps.registry, ref, now + grace);
4722
6076
  } catch (err) {
4723
6077
  const after = await deps.registry.get(ref).catch(() => null);
4724
6078
  if (after?.status === "destroyed") {
4725
- await stampRetirement(deps.registry, ref, now).catch(() => void 0);
6079
+ const stamped = await stampRetirement(deps.registry, ref, now).catch(() => false);
6080
+ if (stamped && removesRows) await filePurgePointer(deps.registry, ref, now + grace);
4726
6081
  retired += 1;
4727
6082
  entries.push({
4728
6083
  ...base,
@@ -4743,6 +6098,20 @@ async function retireExpired(deps, options) {
4743
6098
  entries.push({ ...base, action: "skipped", reason: failureReason(err) });
4744
6099
  }
4745
6100
  }
6101
+ if (!dryRun && removesRows) {
6102
+ let attemptedLitter = 0;
6103
+ for (const pointer of scanned.litter) {
6104
+ if (!owned(pointer.target)) continue;
6105
+ if (attemptedLitter >= limit) break;
6106
+ attemptedLitter += 1;
6107
+ try {
6108
+ if (await deps.registry.get(pointer.target) !== null) continue;
6109
+ await deps.registry.delete(pointer.ref, pointer.token);
6110
+ } catch (err) {
6111
+ noteFault(err);
6112
+ }
6113
+ }
6114
+ }
4746
6115
  return {
4747
6116
  scanned: mine.length,
4748
6117
  eligible,
@@ -4750,6 +6119,8 @@ async function retireExpired(deps, options) {
4750
6119
  wouldRetire,
4751
6120
  tombstonesPurged,
4752
6121
  limited,
6122
+ purgeFaults,
6123
+ ...firstPurgeFault === void 0 ? {} : { firstPurgeFault },
4753
6124
  dryRun,
4754
6125
  entries
4755
6126
  };
@@ -4765,16 +6136,17 @@ function retirementStamp(meta) {
4765
6136
  async function stampRetirement(registry, ref, now) {
4766
6137
  for (let attempt = 0; attempt < 3; attempt += 1) {
4767
6138
  const rec = await registry.get(ref);
4768
- if (rec === null || rec.status !== "destroyed") return;
6139
+ if (rec === null || rec.status !== "destroyed") return false;
4769
6140
  try {
4770
6141
  await registry.compareAndSwap(ref, rec.token, {
4771
6142
  retention: { ...rec.retention, [RETIRED_AT]: now }
4772
6143
  });
4773
- return;
6144
+ return true;
4774
6145
  } catch (err) {
4775
6146
  if (!isWriteConflictError(err)) throw err;
4776
6147
  }
4777
6148
  }
6149
+ return false;
4778
6150
  }
4779
6151
  async function isFullyReclaimed(deps, ref) {
4780
6152
  for await (const key of deps.storage.list(ref)) {
@@ -4795,8 +6167,45 @@ async function generationsPresent(storage, ref) {
4795
6167
  for await (const key of storage.list(ref)) present.add(key.generation);
4796
6168
  return present;
4797
6169
  }
6170
+ async function checkSummary(ref, live, deps) {
6171
+ const generation = live.currentGen;
6172
+ const keyed = live.wrappedDeks !== void 0 && live.wrappedDeks.length > 0;
6173
+ let aead;
6174
+ if (keyed) {
6175
+ if (deps.keystore === void 0) return { kind: "ok", unchecked: true };
6176
+ try {
6177
+ aead = await deps.keystore.openDek(live.wrappedDeks);
6178
+ } catch (err) {
6179
+ if (isIntegrityError(err) || isCloudRoaringError(err) && err.name === "KeyUnavailableError")
6180
+ return { kind: "ok", unchecked: true };
6181
+ throw err;
6182
+ }
6183
+ }
6184
+ const described = usableSummary(ref, live, aead);
6185
+ if (described === void 0) return { kind: "ok", unchecked: keyed };
6186
+ const reader = await openGenerationReader(
6187
+ deps.storage,
6188
+ { namespace: ref.namespace, segment: ref.segment, generation },
6189
+ aead === void 0 ? void 0 : { aead, aadFor: (scope) => aadFor(ref, generation, scope) }
6190
+ );
6191
+ let cardinality = 0;
6192
+ for (const n of reader.cardinalities().values()) cardinality += n;
6193
+ if (summaryAgrees(described, { cardinality, metadata: reader.metadata })) return { kind: "ok" };
6194
+ return {
6195
+ kind: "issue",
6196
+ issue: {
6197
+ segment: ref.segment,
6198
+ namespace: ref.namespace,
6199
+ currentGen: generation,
6200
+ issue: "summary-mismatch"
6201
+ }
6202
+ };
6203
+ }
4798
6204
  async function runConsistencyCheck(deps, options = {}) {
4799
6205
  if (options.namespace !== void 0) validateUserNamespace(options.namespace);
6206
+ if (options.summaries !== void 0 && typeof options.summaries !== "boolean") {
6207
+ throw new ValidationError(`summaries must be a boolean; got ${String(options.summaries)}`);
6208
+ }
4800
6209
  const concurrency = options.concurrency ?? DEFAULT_CHECK_CONCURRENCY;
4801
6210
  if (!Number.isInteger(concurrency) || concurrency < 1) {
4802
6211
  throw new ValidationError(`concurrency must be a positive integer; got ${concurrency}`);
@@ -4818,7 +6227,10 @@ async function runConsistencyCheck(deps, options = {}) {
4818
6227
  if (!live || live.status === "destroyed") return { kind: "ok" };
4819
6228
  if (live.currentGen === null) return { kind: "ok" };
4820
6229
  const present = await generationsPresent(deps.storage, ref);
4821
- if (present.has(live.currentGen)) return { kind: "ok" };
6230
+ if (present.has(live.currentGen)) {
6231
+ if (options.summaries !== true || live.summary === void 0) return { kind: "ok" };
6232
+ return await checkSummary(ref, live, deps);
6233
+ }
4822
6234
  return {
4823
6235
  kind: "issue",
4824
6236
  issue: {
@@ -4841,53 +6253,13 @@ async function runConsistencyCheck(deps, options = {}) {
4841
6253
  });
4842
6254
  const inconsistent = [];
4843
6255
  const errored = [];
6256
+ let summariesUnchecked = 0;
4844
6257
  for (const r of results) {
4845
6258
  if (r.kind === "issue") inconsistent.push(r.issue);
4846
6259
  else if (r.kind === "error") errored.push(r.error);
6260
+ else if (r.unchecked === true) summariesUnchecked += 1;
4847
6261
  }
4848
- return { checked: recs.length, inconsistent, errored };
4849
- }
4850
-
4851
- // src/core/retry.ts
4852
- var DEFAULT_RETRY_POLICY = {
4853
- maxAttempts: 4,
4854
- baseDelayMs: 50,
4855
- maxDelayMs: 2e3,
4856
- backoffFactor: 2,
4857
- jitter: "full"
4858
- };
4859
- function backoffDelayMs(policy, attempt) {
4860
- const raw = policy.baseDelayMs * policy.backoffFactor ** (attempt - 1);
4861
- return Math.min(policy.maxDelayMs, raw);
4862
- }
4863
- function applyJitter(policy, delayMs, rng) {
4864
- if (policy.jitter === "none") return delayMs;
4865
- return rng.next() * delayMs;
4866
- }
4867
- async function withRetry(op, policy, deps) {
4868
- const retryable = deps.isRetryable ?? isTransientError;
4869
- if (!Number.isFinite(policy.maxAttempts) || policy.maxAttempts < 1) {
4870
- throw new ValidationError(
4871
- `retry.maxAttempts must be a finite number >= 1 (got ${String(policy.maxAttempts)})`
4872
- );
4873
- }
4874
- const attempts = Math.floor(policy.maxAttempts);
4875
- let lastErr;
4876
- for (let attempt = 1; attempt <= attempts; attempt++) {
4877
- try {
4878
- return await op();
4879
- } catch (err) {
4880
- lastErr = err;
4881
- if (attempt >= attempts || !retryable(err)) throw err;
4882
- const delayMs = applyJitter(policy, backoffDelayMs(policy, attempt), deps.rng);
4883
- try {
4884
- deps.onRetry?.({ attempt, delayMs, err });
4885
- } catch {
4886
- }
4887
- await deps.clock.sleep(delayMs);
4888
- }
4889
- }
4890
- throw lastErr;
6262
+ return options.summaries === true ? { checked: recs.length, inconsistent, errored, summariesUnchecked } : { checked: recs.length, inconsistent, errored };
4891
6263
  }
4892
6264
 
4893
6265
  // src/drivers/retry/retrying-chunk-source.ts
@@ -4909,6 +6281,7 @@ var RetryingStorageChunkSource = class {
4909
6281
  /** Present only when the inner source supports it — so capability detection stays honest. */
4910
6282
  sizeOf;
4911
6283
  cardinalities;
6284
+ summary;
4912
6285
  currentGeneration;
4913
6286
  invalidate;
4914
6287
  exists;
@@ -4933,6 +6306,10 @@ var RetryingStorageChunkSource = class {
4933
6306
  if (innerCardinalities) {
4934
6307
  this.cardinalities = (ref) => withRetry(() => innerCardinalities.call(inner, ref), this.policy, this.deps);
4935
6308
  }
6309
+ const innerSummary = inner.summary;
6310
+ if (innerSummary) {
6311
+ this.summary = (ref) => withRetry(() => innerSummary.call(inner, ref), this.policy, this.deps);
6312
+ }
4936
6313
  const innerCurrentGeneration = inner.currentGeneration;
4937
6314
  if (innerCurrentGeneration) {
4938
6315
  this.currentGeneration = (ref) => withRetry(() => innerCurrentGeneration.call(inner, ref), this.policy, this.deps);