khive-pack-knowledge 0.12.0

Knowledge verb pack — lore corpus (atoms/domains), TF-IDF retrieval, concept registration
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
//! Handler-level degrade regression tests for the ANN warm-wait path.
//!
//! These tests must live inside the crate because [`vamana::simulate_warming_in_flight`]
//! and [`vamana::set_warm_wait_timeout_override_ms`] are `pub(crate)` — inaccessible
//! from the external `tests/` directory.
//!
//! ## P1 — `search`, `suggest`, and `compose` degrade path
//!
//! When the ANN is warming but not yet loaded and the bounded wait times out,
//! `search` and `suggest` must set `ann_unavailable: true` rather than silently
//! returning zero results. `search` preserves its existing partial-result policy:
//! a lexical hit remains usable without the advisory. `compose` in auto-mode calls
//! `suggest` internally and must propagate the flag in `data["ann_unavailable"]`.
//!
//! The prerequisite for `ann_unavailable` is a non-empty corpus (vectors in the
//! store); we satisfy this by upsert + `knowledge.index` through the registry
//! before the handler call.  The warming-not-loaded state is forced via
//! `simulate_warming_in_flight` on a *fresh* `SharedAnn` (separate from the
//! registry's own).
//!
//! ## P2 — `warm_known_snapshots` end-to-end
//!
//! After `knowledge.index rebuild_ann=true` the persisted Vamana snapshot lives in
//! `retrieval_snapshots`.  Calling `warm_known_snapshots` on a *fresh* `SharedAnn`
//! must load the snapshot so `search_loaded` returns `Some`.

use crate::knowledge::{vamana, KnowledgeHandlers};
use async_trait::async_trait;
use khive_pack_kg::KgPack;
use khive_runtime::{
    AllowAllGate, BackendId, EmbedderProvider, KhiveRuntime, Namespace, RuntimeConfig,
    VerbRegistry, VerbRegistryBuilder,
};
use lattice_embed::{EmbedError, EmbeddingModel, EmbeddingService};
use serde_json::json;
use std::sync::atomic::{AtomicBool, Ordering};
use std::sync::Arc;

// ── fake embedder ─────────────────────────────────────────────────────────────
//
// Returns N distinct 384-dim unit vectors (one per text, differentiated by index
// position) so every indexed atom gets a valid embedding and the Vamana builder
// can produce a non-trivial index.

const MODEL_KEY: &str = "all-minilm-l6-v2";
pub(super) const DIM: usize = 384;

struct FakeDimService;

#[async_trait]
impl EmbeddingService for FakeDimService {
    async fn embed(
        &self,
        texts: &[String],
        _model: EmbeddingModel,
    ) -> Result<Vec<Vec<f32>>, EmbedError> {
        Ok(texts
            .iter()
            .enumerate()
            .map(|(i, _)| {
                let v = (i + 1) as f32;
                let norm = (DIM as f32 * v * v).sqrt();
                vec![v / norm; DIM]
            })
            .collect())
    }

    fn supports_model(&self, _model: EmbeddingModel) -> bool {
        true
    }

    fn name(&self) -> &'static str {
        "fake-dim"
    }
}

struct FakeDimProvider;

#[async_trait]
impl EmbedderProvider for FakeDimProvider {
    fn name(&self) -> &str {
        MODEL_KEY
    }

    fn dimensions(&self) -> usize {
        DIM
    }

    async fn build(&self) -> Result<Arc<dyn EmbeddingService>, khive_runtime::RuntimeError> {
        Ok(Arc::new(FakeDimService))
    }
}

// Controlled two-topic embedder for the degraded-candidate ranking test.  The
// query and genuinely relevant domain share the first axis; the domain whose
// title only collides lexically shares the second. The failing variant still
// embeds/indexes the corpus and embeds the ANN query, but rejects the next
// candidate batch so the fresh rerank is genuinely unavailable even though
// #2232 no longer resends the cached query in that batch.
struct ControlledRankingService {
    fail_fresh_rerank: bool,
    query_embedded: AtomicBool,
}

#[async_trait]
impl EmbeddingService for ControlledRankingService {
    async fn embed(
        &self,
        texts: &[String],
        _model: EmbeddingModel,
    ) -> Result<Vec<Vec<f32>>, EmbedError> {
        if self.fail_fresh_rerank && self.query_embedded.swap(false, Ordering::AcqRel) {
            return Err(EmbedError::InferenceFailed(
                "controlled fresh-rerank failure".into(),
            ));
        }
        Ok(texts
            .iter()
            .map(|text| {
                let lower = text.to_lowercase();
                if lower.contains("semantic_target")
                    || lower.contains("speculative decoding inference acceleration")
                {
                    let mut vector = vec![0.0; DIM];
                    vector[0] = 1.0;
                    vector
                } else if lower.contains("lexical_collision") {
                    let mut vector = vec![0.0; DIM];
                    vector[1] = 1.0;
                    vector
                } else {
                    vec![1.0 / (DIM as f32).sqrt(); DIM]
                }
            })
            .collect())
    }

    async fn embed_query(
        &self,
        texts: &[String],
        model: EmbeddingModel,
    ) -> Result<Vec<Vec<f32>>, EmbedError> {
        let result = self.embed(texts, model).await;
        if result.is_ok() {
            self.query_embedded.store(true, Ordering::Release);
        }
        result
    }

    fn supports_model(&self, _model: EmbeddingModel) -> bool {
        true
    }

    fn name(&self) -> &'static str {
        "controlled-ranking"
    }
}

struct ControlledRankingProvider {
    fail_fresh_rerank: bool,
}

#[async_trait]
impl EmbedderProvider for ControlledRankingProvider {
    fn name(&self) -> &str {
        MODEL_KEY
    }

    fn dimensions(&self) -> usize {
        DIM
    }

    async fn build(&self) -> Result<Arc<dyn EmbeddingService>, khive_runtime::RuntimeError> {
        Ok(Arc::new(ControlledRankingService {
            fail_fresh_rerank: self.fail_fresh_rerank,
            query_embedded: AtomicBool::new(false),
        }))
    }
}

// ── helpers ───────────────────────────────────────────────────────────────────

fn rt_with_fake_embedder() -> KhiveRuntime {
    let rt = KhiveRuntime::new(RuntimeConfig {
        wal_ceiling_bytes: 0,
        wal_ceiling_configured_bytes: 0,
        wal_ceiling_source: Default::default(),
        wal_ceiling_env_raw: None,
        web: Default::default(),
        telemetry: Default::default(),
        mounts: Vec::new(),
        brain: Default::default(),
        git_write: Default::default(),
        display_timezone: khive_runtime::config::resolve_default_display_timezone(),
        events_split: None,
        db_path: None,
        blob_hydration_bytes: khive_runtime::DEFAULT_BLOB_HYDRATION_BYTES,
        default_namespace: Namespace::local(),
        embedding_model: Some(EmbeddingModel::AllMiniLmL6V2),
        additional_embedding_models: vec![],
        gate: Arc::new(AllowAllGate),
        packs: vec!["kg".to_string(), "knowledge".to_string()],
        backend_id: BackendId::main(),
        brain_profile: None,
        visible_namespaces: vec![],
        allowed_outbound_namespaces: vec![],
        actor_id: None,
        exec: Default::default(),
        ..khive_runtime::RuntimeConfig::no_embeddings()
    })
    .expect("in-memory runtime");
    rt.register_embedder(FakeDimProvider);
    rt
}

fn rt_with_controlled_ranking(fail_fresh_rerank: bool) -> KhiveRuntime {
    let rt = KhiveRuntime::new(RuntimeConfig {
        wal_ceiling_bytes: 0,
        wal_ceiling_configured_bytes: 0,
        wal_ceiling_source: Default::default(),
        wal_ceiling_env_raw: None,
        web: Default::default(),
        telemetry: Default::default(),
        mounts: Vec::new(),
        brain: Default::default(),
        git_write: Default::default(),
        display_timezone: khive_runtime::config::resolve_default_display_timezone(),
        events_split: None,
        db_path: None,
        blob_hydration_bytes: khive_runtime::DEFAULT_BLOB_HYDRATION_BYTES,
        default_namespace: Namespace::local(),
        embedding_model: Some(EmbeddingModel::AllMiniLmL6V2),
        additional_embedding_models: vec![],
        gate: Arc::new(AllowAllGate),
        packs: vec!["kg".to_string(), "knowledge".to_string()],
        backend_id: BackendId::main(),
        brain_profile: None,
        visible_namespaces: vec![],
        allowed_outbound_namespaces: vec![],
        actor_id: None,
        exec: Default::default(),
        ..khive_runtime::RuntimeConfig::no_embeddings()
    })
    .expect("in-memory runtime");
    rt.register_embedder(ControlledRankingProvider { fail_fresh_rerank });
    rt
}

/// File-backed variant. Required by tests that exercise v2 ANN persistence:
/// `knowledge.index(rebuild_ann=true)` only writes v2 segments when the backend
/// has a `data_dir`. An in-memory runtime has none, and ADR-079 removed the v1
/// `retrieval_snapshots` write path, so an in-memory rebuild persists nothing.
pub(super) fn file_rt_with_fake_embedder(db_path: std::path::PathBuf) -> KhiveRuntime {
    let rt = KhiveRuntime::new(RuntimeConfig {
        wal_ceiling_bytes: 0,
        wal_ceiling_configured_bytes: 0,
        wal_ceiling_source: Default::default(),
        wal_ceiling_env_raw: None,
        web: Default::default(),
        telemetry: Default::default(),
        mounts: Vec::new(),
        brain: Default::default(),
        git_write: Default::default(),
        display_timezone: khive_runtime::config::resolve_default_display_timezone(),
        events_split: None,
        db_path: Some(db_path),
        blob_hydration_bytes: khive_runtime::DEFAULT_BLOB_HYDRATION_BYTES,
        default_namespace: Namespace::local(),
        embedding_model: Some(EmbeddingModel::AllMiniLmL6V2),
        additional_embedding_models: vec![],
        gate: Arc::new(AllowAllGate),
        packs: vec!["kg".to_string(), "knowledge".to_string()],
        backend_id: BackendId::main(),
        brain_profile: None,
        visible_namespaces: vec![],
        allowed_outbound_namespaces: vec![],
        actor_id: None,
        exec: Default::default(),
        ..khive_runtime::RuntimeConfig::no_embeddings()
    })
    .expect("file-backed runtime");
    rt.register_embedder(FakeDimProvider);
    rt
}

pub(super) fn build_registry(rt: &KhiveRuntime) -> VerbRegistry {
    let mut builder = VerbRegistryBuilder::new();
    builder.register(KgPack::new(rt.clone()));
    builder.register(crate::KnowledgePack::new(rt.clone()));
    let registry = builder.build().expect("registry builds");
    rt.install_edge_rules(registry.all_edge_rules());
    registry
}

async fn lexical_timeout_fixture(
    atoms: Vec<serde_json::Value>,
    domain: serde_json::Value,
) -> (KhiveRuntime, VerbRegistry) {
    let rt = rt_with_fake_embedder();
    let registry = build_registry(&rt);
    if !atoms.is_empty() {
        registry
            .dispatch("knowledge.upsert_atoms", json!({"atoms": atoms}))
            .await
            .expect("upsert member atoms");
    }
    registry
        .dispatch("knowledge.upsert_domains", json!({"domains": [domain]}))
        .await
        .expect("upsert domain");
    registry
        .dispatch("knowledge.index", json!({"rebuild_ann": false}))
        .await
        .expect("index");
    (rt, registry)
}

/// RAII guard: reset the timeout override when the test exits (even on panic).
struct TimeoutOverrideReset;

impl Drop for TimeoutOverrideReset {
    fn drop(&mut self) {
        vamana::set_warm_wait_timeout_override_ms(0);
    }
}

/// Serializes the timeout-override tests. They mutate the process-global
/// `ANN_WARM_WAIT_TIMEOUT_OVERRIDE_MS`, and `TimeoutOverrideReset` clears it on
/// drop. Under Cargo's parallel test runner, one test's reset could otherwise
/// fire while another is mid-flight, dropping it back to the 5s production
/// timeout (a latency-order-dependent slow run). `tokio::sync::Mutex` is
/// await-safe (no `clippy::await_holding_lock`) and does not poison on panic,
/// so a failing test still releases the lock. Each test declares the guard
/// before `TimeoutOverrideReset`, so on exit the reset (override -> 0) runs
/// first and the lock releases only after, handing a clean state to the next.
static TIMEOUT_OVERRIDE_SERIAL: tokio::sync::Mutex<()> = tokio::sync::Mutex::const_new(());

// ── P1: search preserves its timeout fallback policy

/// `search` reports ANN unavailability only when warming times out, the corpus
/// is non-empty, and lexical search also found no result. A lexical hit remains
/// a valid partial result without the advisory.
#[tokio::test]
async fn search_preserves_timeout_fallback_policy() {
    let _serial = TIMEOUT_OVERRIDE_SERIAL.lock().await;
    vamana::set_warm_wait_timeout_override_ms(50);
    let _reset = TimeoutOverrideReset;

    let rt = rt_with_fake_embedder();
    let registry = build_registry(&rt);
    registry
        .dispatch(
            "knowledge.upsert_atoms",
            json!({
                "atoms": [{
                    "slug": "degrade-search-atom",
                    "name": "Degrade Search Atom",
                    "finalized": true,
                    "content": "lexicalsentinel transformer retrieval corpus benchmark search latency vector index nearest neighbor ranking fusion embedding cosine similarity attention encoder decoder positional normalization residual connection"
                }]
            }),
        )
        .await
        .expect("upsert atom");
    registry
        .dispatch("knowledge.index", json!({ "rebuild_ann": false }))
        .await
        .expect("index");

    let ann = vamana::new_shared();
    let key = vamana::AnnKey::new("local", rt.default_embedder_name());
    vamana::simulate_warming_in_flight(&ann, key);
    let token = rt.authorize(Namespace::local()).expect("authorize");

    // The domain gate drops the seeded non-domain atom, including from the
    // small-corpus full-scan fallback, so lexical search is genuinely empty.
    let empty = KnowledgeHandlers::search(
        &rt,
        &token,
        json!({ "query": "unfindablegossamer", "kind": "domain", "rerank": false }),
        &ann,
    )
    .await
    .expect("empty search must not Err");
    assert_eq!(empty["total"], 0, "result: {empty}");
    assert_eq!(empty["ann_unavailable"], true, "result: {empty}");

    let lexical = KnowledgeHandlers::search(
        &rt,
        &token,
        json!({ "query": "lexicalsentinel", "rerank": false }),
        &ann,
    )
    .await
    .expect("lexical fallback search must not Err");
    assert!(
        lexical["total"].as_u64().unwrap_or_default() > 0,
        "lexical fallback must retain its hit; result: {lexical}"
    );
    assert!(
        lexical.get("ann_unavailable").is_none(),
        "search preserves its existing no-advisory policy when FTS found hits; result: {lexical}"
    );

    // The simulated warm remains in flight: serving ANN cannot provide a
    // vector, but the persisted vec0 row still reranks the lexical hit.
    let reranked = KnowledgeHandlers::search(
        &rt,
        &token,
        json!({ "query": "lexicalsentinel", "rerank": true }),
        &ann,
    )
    .await
    .expect("cold ANN must not block stored-vector rerank");
    assert_eq!(
        reranked["rerank_provenance"]["stored_vector_lookup"], "supported",
        "result: {reranked}"
    );
    assert!(
        reranked["rerank_provenance"]["from_stored"]
            .as_u64()
            .unwrap_or_default()
            > 0,
        "cold serving ANN must still read persisted candidate vectors: {reranked}"
    );
}

// ── P1a: suggest sets ann_unavailable when warming times out ─────────────────

/// `suggest` must set `ann_unavailable: true` when:
/// 1. The ANN key is in the warming set but the index is not yet loaded
///    (`simulate_warming_in_flight` injects this state into a fresh `SharedAnn`).
/// 2. The bounded wait times out (50 ms override via `set_warm_wait_timeout_override_ms`).
/// 3. FTS hits are empty: `suggest` uses `type_filter = Some("domain")` internally,
///    and our seeded atom has no `type:domain` tag — `load_candidates_from_atoms`
///    drops it, so `hits.is_empty() == true`.
/// 4. The corpus has vectors: `compute_fingerprint().vector_count > 0` (satisfied
///    by running `knowledge.index` before the handler call).
#[tokio::test]
async fn suggest_sets_ann_unavailable_when_warming_times_out() {
    let _serial = TIMEOUT_OVERRIDE_SERIAL.lock().await;
    vamana::set_warm_wait_timeout_override_ms(50);
    let _reset = TimeoutOverrideReset;

    let rt = rt_with_fake_embedder();
    let registry = build_registry(&rt);

    // Seed a regular (non-domain) atom, then index to populate the vector store.
    registry
        .dispatch(
            "knowledge.upsert_atoms",
            json!({
                "atoms": [{
                    "slug": "degrade-suggest-atom",
                    "name": "Degrade Suggest Atom",
                    "content": "transformer neural network attention mechanism self-attention encoder decoder positional embedding layer normalization residual connection feed forward dense sparse retrieval vector index"
                }]
            }),
        )
        .await
        .expect("upsert atom");

    registry
        .dispatch("knowledge.index", json!({ "rebuild_ann": false }))
        .await
        .expect("index");

    // A fresh SharedAnn — separate from the registry's own — with the key in
    // warming but no index loaded.  This forces the degrade path in `suggest`.
    let ann = vamana::new_shared();
    let model = rt.default_embedder_name().to_string();
    let key = vamana::AnnKey::new("local", &model);
    vamana::simulate_warming_in_flight(&ann, key);

    let token = rt.authorize(Namespace::local()).expect("authorize");
    let result = KnowledgeHandlers::suggest(
        &rt,
        &token,
        // ≥5 words required by suggest; type_filter="domain" will drop the
        // non-domain atom, leaving FTS hits empty → ann_unavailable condition met.
        json!({ "query": "machine learning neural network transformer attention" }),
        &ann,
    )
    .await
    .expect("suggest must not Err");

    assert_eq!(
        result.get("ann_unavailable").and_then(|v| v.as_bool()),
        Some(true),
        "suggest must carry ann_unavailable=true when ANN warming times out \
         and FTS hits are empty; got: {result}"
    );
}

// ── P1b: compose propagates ann_unavailable from its internal suggest call ────

/// `compose` in auto-mode delegates to `suggest` and must surface
/// `data["ann_unavailable"] = true` when the underlying `suggest` sets the flag.
///
/// Auto-mode is triggered when `domain_ids` and `atom_ids` are absent. A live
/// lexical domain with no members is served by suggest but skipped by compose,
/// which reaches the no-domains early return while preserving the signal.
#[tokio::test]
async fn compose_propagates_ann_unavailable_in_auto_mode() {
    let _serial = TIMEOUT_OVERRIDE_SERIAL.lock().await;
    vamana::set_warm_wait_timeout_override_ms(50);
    let _reset = TimeoutOverrideReset;

    let rt = rt_with_fake_embedder();
    let registry = build_registry(&rt);

    registry
        .dispatch(
            "knowledge.upsert_domains",
            json!({
                "domains": [{
                    "slug": "degrade-empty-compose-domain",
                    "name": "Degrade Empty Compose Domain",
                    "description": "machine learning neural network transformer attention architecture multi head self attention retrieval composition domain with intentionally empty membership for reliable production operations",
                    "members": []
                }]
            }),
        )
        .await
        .expect("upsert empty domain");

    registry
        .dispatch("knowledge.index", json!({ "rebuild_ann": false }))
        .await
        .expect("index");

    let ann = vamana::new_shared();
    let model = rt.default_embedder_name().to_string();
    let key = vamana::AnnKey::new("local", &model);
    vamana::simulate_warming_in_flight(&ann, key);

    let token = rt.authorize(Namespace::local()).expect("authorize");
    // Auto-mode requires ≥10 words; no domain_ids/atom_ids.
    // type_weights are not reached because compose skips the zero-member domain.
    let result = KnowledgeHandlers::compose(
        &rt,
        &token,
        json!({
            "query": "machine learning neural network transformer attention architecture multi head self attention"
        }),
        &ann,
        std::collections::HashMap::new(),
    )
    .await
    .expect("compose must not Err");

    assert_eq!(
        result
            .get("data")
            .and_then(|d| d.get("ann_unavailable"))
            .and_then(|v| v.as_bool()),
        Some(true),
        "compose must propagate ann_unavailable=true from its internal suggest call; \
         got: {result}"
    );
    assert_eq!(
        result["data"]["markdown"],
        "# Knowledge Briefing\n\nNo matching domains found for auto-suggest.",
        "compose must skip the zero-member domain while preserving the signal; got: {result}"
    );
}

#[tokio::test]
async fn auto_compose_skips_zero_size_suggested_domains() {
    let rt = KhiveRuntime::memory().expect("in-memory runtime");
    let registry = build_registry(&rt);
    registry
        .dispatch(
            "knowledge.upsert_atoms",
            json!({"atoms": [{
                "slug": "compose-live-member",
                "name": "Compose Live Member",
                "content": "Stored text supplies a valid live member with a measurable cost for each domain while the separate domain descriptions determine candidate relevance and ordering.",
                "finalized": true
            }]}),
        )
        .await
        .expect("upsert live member");
    registry
        .dispatch(
            "knowledge.upsert_domains",
            json!({"domains": [
                {
                    "slug": "compose-empty-domain",
                    "name": "Compose Empty Domain",
                    "description": "Lexical domain member sizing search uses the live content of member atoms to calculate the estimated token cost before a fold selection admits the domain into its budget.",
                    "members": []
                },
                {
                    "slug": "compose-populated-domain",
                    "name": "Compose Populated Domain",
                    "description": "Lexical domain member sizing search uses the live content of member atoms to calculate the estimated token cost before a fold selection admits the domain into its budget.",
                    "members": ["compose-live-member"]
                }
            ]}),
        )
        .await
        .expect("upsert empty and populated domains");
    let response = registry
        .dispatch(
            "knowledge.compose",
            json!({"query": "explain lexical domain member sizing search and token cost for automatic composition"}),
        )
        .await
        .expect("auto-compose succeeds");
    let domains = response["data"]["domains"].as_array().expect("domains");
    assert_eq!(domains.len(), 1, "got: {response}");
    assert_eq!(domains[0]["slug"], "compose-populated-domain");
    assert_eq!(response["data"]["count"], 1, "got: {response}");
    assert_eq!(response["data"]["atoms"][0]["slug"], "compose-live-member");

    let explicit = registry
        .dispatch(
            "knowledge.compose",
            json!({"query": "empty domain", "domain_ids": ["compose-empty-domain"]}),
        )
        .await
        .expect("explicit empty-domain request remains supported");
    assert_eq!(
        explicit["data"]["markdown"], "# Knowledge Briefing\n\nNo atoms found.",
        "the candidate skip must not apply to explicit domain ids: {explicit}"
    );
}

// ── P2: warm_known_snapshots loads a persisted snapshot into a fresh SharedAnn ─

/// After `knowledge.index rebuild_ann=true` the Vamana snapshot is persisted in
/// `retrieval_snapshots`.  Calling `warm_known_snapshots` on a *fresh* `SharedAnn`
/// must load that snapshot so `search_loaded` returns `Some` (index is in memory).
#[tokio::test]
async fn warm_known_snapshots_loads_persisted_snapshot() {
    // File-backed runtime so knowledge.index(rebuild_ann=true) persists v2 ANN
    // segments to data_dir/ann/<hex>; warm_known_snapshots then enumerates and
    // loads them. (In-memory has no data_dir, and ADR-079 removed the v1 write
    // path, so nothing would be persisted to warm.)
    let dir = tempfile::TempDir::new().expect("tempdir");
    let rt = file_rt_with_fake_embedder(dir.path().join("test.db"));
    let registry = build_registry(&rt);

    // Seed two atoms so Vamana has enough vectors to build a non-trivial index.
    registry
        .dispatch(
            "knowledge.upsert_atoms",
            json!({
                "atoms": [
                    {
                        "slug": "warm-snap-atom-a",
                        "name": "Warm Snapshot Atom A",
                        "content": "dense retrieval corpus benchmark search latency gradient descent vector index nearest neighbor ranking fusion pipeline embedding rerank cosine similarity unique warm78a"
                    },
                    {
                        "slug": "warm-snap-atom-b",
                        "name": "Warm Snapshot Atom B",
                        "content": "ranking fusion pipeline embedding rerank cosine similarity unique warm78b transformer attention mechanism self-attention encoder decoder positional feed forward dense neural network gradient"
                    }
                ]
            }),
        )
        .await
        .expect("upsert atoms");

    // Index with rebuild_ann=true to persist the Vamana snapshot in retrieval_snapshots.
    let index_result = registry
        .dispatch("knowledge.index", json!({ "rebuild_ann": true }))
        .await
        .expect("index with rebuild_ann=true");

    // Guard: the index run must have actually embedded atoms (not just done nothing).
    let indexed = index_result
        .get("indexed")
        .and_then(|v| v.as_u64())
        .unwrap_or(0);
    assert!(
        indexed >= 2,
        "knowledge.index must embed at least 2 atoms for this test to be meaningful; \
         got indexed={indexed}"
    );

    // A fresh SharedAnn — no snapshot loaded yet.
    let ann = vamana::new_shared();
    let model = rt.default_embedder_name().to_string();
    let key = vamana::AnnKey::new("local", &model);

    // Precondition: the fresh ann has nothing loaded.
    let dummy_query = vec![1.0f32 / (DIM as f32).sqrt(); DIM];
    assert!(
        vamana::search_loaded(&ann, &key, &dummy_query, 1)
            .await
            .is_none(),
        "precondition: fresh SharedAnn must have no index loaded before warm_known_snapshots"
    );

    // warm_known_snapshots reads retrieval_snapshots, finds the persisted key, and
    // calls ensure_ann_for_model which restores the AnnBridge from the snapshot.
    vamana::warm_known_snapshots(&rt, &ann).await;

    assert!(
        vamana::search_loaded(&ann, &key, &dummy_query, 1)
            .await
            .is_some(),
        "search_loaded must return Some after warm_known_snapshots loads the snapshot; \
         model={model}, key={key:?}"
    );
}

// ── P3: issue #91 — degraded-mode escalation + consequence semantics ──────────

/// When ANN candidate retrieval times out, distinguish a successful fresh
/// dense rerank from a genuinely lexical-only result.  Both cases use the same
/// candidates: dense reranking promotes the semantically relevant domain,
/// while an unavailable rerank leaves the lexical title collision first.
#[tokio::test]
async fn suggest_reports_degraded_candidates_or_lexical_only_from_ranking_consequence() {
    let _serial = TIMEOUT_OVERRIDE_SERIAL.lock().await;
    vamana::set_warm_wait_timeout_override_ms(50);
    let _reset = TimeoutOverrideReset;

    for (fail_fresh_rerank, expected_mode, expected_winner) in [
        (false, "ann_candidates_degraded", "Opaque Systems Domain"),
        (
            true,
            "lexical_only",
            "Speculative Decoding Methods for Inference",
        ),
    ] {
        let rt = rt_with_controlled_ranking(fail_fresh_rerank);
        let registry = build_registry(&rt);
        registry
            .dispatch(
                "knowledge.upsert_domains",
                json!({"domains": [
                    {
                        "slug": "degrade-semantic-domain",
                        "name": "Opaque Systems Domain",
                        "description": "SEMANTIC_TARGET speculative serving throughput decoding latency batching inference cache scheduling acceleration gpu utilization techniques request queuing large scale parallelism language aware autoscaling models deployment topology load balancing production workloads resource management observability reliability"
                    },
                    {
                        "slug": "degrade-lexical-collision",
                        "name": "Speculative Decoding Methods for Inference",
                        "description": "LEXICAL_COLLISION techniques for large language models emphasize decoding speculation, inference methods, acceleration strategies, model techniques, language processing, and unrelated terminology repeated for lexical matching"
                    }
                ]}),
            )
            .await
            .expect("upsert domains");
        registry
            .dispatch("knowledge.index", json!({ "rebuild_ann": false }))
            .await
            .expect("index");

        let ann = vamana::new_shared();
        let key = vamana::AnnKey::new("local", rt.default_embedder_name());
        vamana::simulate_warming_in_flight(&ann, key);
        let token = rt.authorize(Namespace::local()).expect("authorize");
        let result = KnowledgeHandlers::suggest(
            &rt,
            &token,
            json!({
                "query": "speculative decoding inference acceleration techniques for large language models"
            }),
            &ann,
        )
        .await
        .expect("suggest must not Err");

        assert_eq!(result["ann_unavailable"], true, "result: {result}");
        assert_eq!(
            result["degraded"]["mode"], expected_mode,
            "result: {result}"
        );
        assert_eq!(
            result["results"][0]["name"], expected_winner,
            "result: {result}"
        );
    }
}

/// The zero-hit case (P1a's scenario) must ALSO carry the new `degraded`
/// object with `mode: "no_match"`, distinguishing "couldn't check" from
/// "lexical matching ran and found this particular set" (the P3 case above).
#[tokio::test]
async fn suggest_flags_degraded_no_match_when_hits_empty() {
    let _serial = TIMEOUT_OVERRIDE_SERIAL.lock().await;
    vamana::set_warm_wait_timeout_override_ms(50);
    let _reset = TimeoutOverrideReset;

    let rt = rt_with_fake_embedder();
    let registry = build_registry(&rt);

    // Non-domain atom: `suggest`'s type_filter="domain" drops it, so lexical
    // hits are empty (mirrors P1a).
    registry
        .dispatch(
            "knowledge.upsert_atoms",
            json!({
                "atoms": [{
                    "slug": "degrade-suggest-atom-2",
                    "name": "Degrade Suggest Atom 2",
                    "content": "transformer neural network attention mechanism self-attention encoder decoder positional embedding layer normalization residual connection feed forward dense sparse retrieval vector index"
                }]
            }),
        )
        .await
        .expect("upsert atom");

    registry
        .dispatch("knowledge.index", json!({ "rebuild_ann": false }))
        .await
        .expect("index");

    let ann = vamana::new_shared();
    let model = rt.default_embedder_name().to_string();
    let key = vamana::AnnKey::new("local", &model);
    vamana::simulate_warming_in_flight(&ann, key);

    let token = rt.authorize(Namespace::local()).expect("authorize");
    let result = KnowledgeHandlers::suggest(
        &rt,
        &token,
        json!({ "query": "machine learning neural network transformer attention" }),
        &ann,
    )
    .await
    .expect("suggest must not Err");

    assert_eq!(result.get("total").and_then(|v| v.as_u64()), Some(0));
    assert_eq!(
        result.get("ann_unavailable").and_then(|v| v.as_bool()),
        Some(true)
    );
    let degraded = result
        .get("degraded")
        .unwrap_or_else(|| panic!("degraded object must be present; got: {result}"));
    assert_eq!(
        degraded.get("mode").and_then(|v| v.as_str()),
        Some("no_match"),
        "zero-hit degraded suggest must report mode=no_match, distinct from lexical_only; got: {result}"
    );
    assert_eq!(
        degraded.get("cache_safe").and_then(|v| v.as_bool()),
        Some(false)
    );
}

// ── P4: issue #1930 — lexical-timeout degraded arm ────────────────────────────

/// A timed-out lexical read must retain the independently completed ANN
/// candidate. Inject only the lexical timeout so the shared request deadline
/// cannot also cancel ANN under host-dependent load. The real fresh-tail scan
/// remains in use; the candidate is served or explicitly excluded if sizing
/// could not measure it.
#[tokio::test(start_paused = true)]
async fn lexical_timeout_degrades_suggest_to_ann_backed_results() {
    let (rt, _registry) = lexical_timeout_fixture(
        Vec::new(),
        json!({
            "slug": "degrade-lexical-timeout-domain",
            "name": "Degrade Lexical Timeout Domain",
            "description": "a domain seeded only so ANN has a real vector to serve from the fresh-tail scan path when the lexical reader times out during this regression test",
            "members": []
        }),
    )
    .await;

    // A fresh SharedAnn: nothing warmed, no snapshot. `search_eligible_ann_with_refill`
    // falls back to the fresh-tail vector-store scan, which still finds the
    // domain seeded above.
    let ann = vamana::new_shared();
    let token = rt.authorize(Namespace::local()).expect("authorize");

    let query = "term0 term1 term2 term3 term4 term5 term6 term7";
    let result = crate::knowledge::lexical_timeout::tests::with_timeout(
        vec![crate::knowledge::lexical_timeout::LexicalPhase::ReaderOpen],
        std::time::Duration::from_millis(1),
        KnowledgeHandlers::suggest(&rt, &token, json!({ "query": query }), &ann),
    )
    .await
    .expect("suggest must not Err on a lexical-stage read timeout");

    assert_eq!(
        result["degraded"]["lexical_timeout"], true,
        "suggest must flag degraded.lexical_timeout when the lexical fetch times out; got: {result}"
    );
    assert_eq!(
        result["total"], 1,
        "ANN must survive lexical timeout: {result}"
    );
    assert_eq!(
        result["results"][0]["name"],
        "Degrade Lexical Timeout Domain"
    );
    // The ANN survivor is never dropped silently: it is served in `results`
    // when its member size was measured, and reported under
    // `degraded.member_sizing_timeout.excluded` (rank and score intact, no
    // fabricated size) if sizing was unmeasured.
    let surfaced: Vec<&str> = result["results"]
        .as_array()
        .into_iter()
        .flatten()
        .chain(
            result["degraded"]["member_sizing_timeout"]["excluded"]
                .as_array()
                .into_iter()
                .flatten(),
        )
        .filter_map(|hit| hit["name"].as_str())
        .collect();
    assert_eq!(
        surfaced,
        vec!["Degrade Lexical Timeout Domain"],
        "the only vector-backed candidate must be surfaced exactly once, served or withheld; got: {result}"
    );
    for hit in result["results"].as_array().into_iter().flatten() {
        assert!(
            hit["size"].is_u64(),
            "every served result carries a measured size; got: {result}"
        );
    }
}

/// Issue #2396 fix 3 (revised): the seeded domain has a real member atom, so a
/// real (non-zero) size would be computed if member sizing ran. A separate
/// scoped read timeout at the sizing boundary must withhold that domain
/// from `results` and report it under `degraded.member_sizing_timeout.excluded`
/// instead — never left in `results` with `size: 0` (which would let a
/// caller pass it into `knowledge.fold` as a free item) or `size: null`
/// (which breaks the documented suggest -> fold passthrough, issue #105).
#[tokio::test(start_paused = true)]
async fn lexical_timeout_reports_member_sizing_as_unmeasured_not_zero() {
    let (rt, registry) = lexical_timeout_fixture(
        vec![json!({
            "slug": "degrade-sizing-member-atom",
            "name": "Degrade Sizing Member Atom",
            "finalized": true,
            "content": "enough body content to price a non-zero token size for the owning domain if member sizing ever ran to completion"
        })],
        json!({
            "slug": "degrade-sizing-domain",
            "name": "Degrade Sizing Domain",
            "description": "a domain with a real member atom, so a real (non-zero) size would be computed if member sizing ran, proving a timeout is reported as unmeasured rather than a coincidental zero",
            "members": ["degrade-sizing-member-atom"]
        }),
    )
    .await;

    let ann = vamana::new_shared();
    let token = rt.authorize(Namespace::local()).expect("authorize");

    let query = "term0 term1 term2 term3 term4 term5 term6 term7";
    let result = crate::knowledge::search::with_member_sizing_read_timeout(
        crate::knowledge::lexical_timeout::tests::with_timeout(
            vec![crate::knowledge::lexical_timeout::LexicalPhase::ReaderOpen],
            std::time::Duration::from_millis(1),
            KnowledgeHandlers::suggest(&rt, &token, json!({ "query": query }), &ann),
        ),
    )
    .await
    .expect("suggest must not Err on a lexical-stage read timeout");

    assert_eq!(result["degraded"]["lexical_timeout"], true, "got: {result}");
    assert!(
        result["results"].as_array().is_some_and(Vec::is_empty),
        "an unmeasured domain must be withheld from `results` entirely, never \
         left in with size 0 or null; got: {result}"
    );
    let excluded = result["degraded"]["member_sizing_timeout"]["excluded"]
        .as_array()
        .cloned()
        .unwrap_or_default();
    assert_eq!(
        excluded.len(),
        1,
        "the degradation entry must list the excluded domain; got: {result}"
    );
    assert_eq!(excluded[0]["rank"], 1, "got: {result}");
    assert_eq!(excluded[0]["name"], "Degrade Sizing Domain");
    assert!(
        excluded[0]["id"].is_string()
            && excluded[0]["name"].is_string()
            && excluded[0]["score"].is_number()
            && excluded[0].get("size").is_none(),
        "an excluded entry must carry id/name/rank/score for reference, and no \
         size; got: {result}"
    );

    registry
        .dispatch(
            "knowledge.fold",
            json!({ "candidates": result["results"].clone(), "budget": 10_000 }),
        )
        .await
        .expect(
            "suggest's results must feed fold unmodified even when a domain was \
             excluded for an unmeasured size (issue #105 passthrough)",
        );
}

/// Issue #2396 follow-up: `suggest`'s documented contract (issue #105) is
/// that a caller feeds its `results` straight into `knowledge.fold`'s
/// `candidates` unmodified. A member-sizing timeout must not break that
/// passthrough — the unpriced domain is withheld from `results` (and
/// reported under `degraded.member_sizing_timeout.excluded`), never left in
/// `results` with a `size` that `FoldCandidate`'s non-optional `usize`
/// cannot parse. Before the fix, this test's `fold` dispatch failed with a
/// parse error on the degraded domain's `size: null`.
#[tokio::test(start_paused = true)]
async fn suggest_results_under_a_member_sizing_timeout_still_pass_through_fold() {
    let (rt, registry) = lexical_timeout_fixture(
        vec![json!({
            "slug": "degrade-sizing-passthrough-atom",
            "name": "Degrade Sizing Passthrough Atom",
            "finalized": true,
            "content": "enough body content to price a non-zero token size for the owning domain if member sizing ever ran to completion"
        })],
        json!({
            "slug": "degrade-sizing-passthrough-domain",
            "name": "Degrade Sizing Passthrough Domain",
            "description": "a domain with a real member atom, so a real (non-zero) size would be computed if member sizing ran, proving a timeout is reported as unmeasured rather than a coincidental zero",
            "members": ["degrade-sizing-passthrough-atom"]
        }),
    )
    .await;

    let ann = vamana::new_shared();
    let token = rt.authorize(Namespace::local()).expect("authorize");

    let query = "term0 term1 term2 term3 term4 term5 term6 term7";
    let result = crate::knowledge::search::with_member_sizing_read_timeout(
        crate::knowledge::lexical_timeout::tests::with_timeout(
            vec![crate::knowledge::lexical_timeout::LexicalPhase::ReaderOpen],
            std::time::Duration::from_millis(1),
            KnowledgeHandlers::suggest(&rt, &token, json!({ "query": query }), &ann),
        ),
    )
    .await
    .expect("suggest must not Err on a lexical-stage read timeout");

    assert!(
        result["degraded"]["member_sizing_timeout"]["excluded"]
            .as_array()
            .is_some_and(|excluded| !excluded.is_empty()),
        "the degraded scenario must produce at least one excluded domain; got: {result}"
    );

    registry
        .dispatch(
            "knowledge.fold",
            json!({ "candidates": result["results"].clone(), "budget": 10_000 }),
        )
        .await
        .expect("suggest's results must feed fold unmodified (issue #105 passthrough)");
}

#[tokio::test]
async fn suggest_member_sizing_serves_a_domain_with_only_a_deleted_member_at_zero() {
    let rt = KhiveRuntime::memory().expect("in-memory runtime");
    let registry = build_registry(&rt);
    registry
        .dispatch(
            "knowledge.upsert_atoms",
            json!({"atoms": [{
                "slug": "sizing-deleted-member",
                "name": "Sizing Deleted Member",
                "content": "A live member body with enough content to have a nonzero estimated token cost before the atom is soft deleted.",
                "finalized": true
            }]}),
        )
        .await
        .expect("upsert member atom");
    registry
        .dispatch(
            "knowledge.upsert_domains",
            json!({"domains": [{
                "slug": "sizing-measured-domain",
                "name": "Sizing Measured Domain",
                "description": "Lexical domain member sizing search uses the live content of member atoms to calculate the estimated token cost before a fold selection admits the domain into its budget.",
                "members": ["sizing-deleted-member"]
            }]}),
        )
        .await
        .expect("upsert domain");
    let token = rt.authorize(Namespace::local()).expect("authorize");
    let ann = vamana::new_shared();
    let params = json!({"query": "lexical domain member sizing search", "limit": 1});
    let baseline = KnowledgeHandlers::suggest(&rt, &token, params.clone(), &ann)
        .await
        .expect("suggest with live member");
    assert_eq!(baseline["total"], 1, "got: {baseline}");
    assert!(baseline["results"][0]["size"].as_u64().unwrap_or(0) > 0);
    assert_eq!(baseline["results"][0]["members"], 1, "got: {baseline}");

    let deleted = registry
        .dispatch(
            "knowledge.delete_atoms",
            json!({"ids": ["sizing-deleted-member"]}),
        )
        .await
        .expect("soft-delete member atom");
    assert_eq!(deleted["deleted"], 1, "got: {deleted}");

    let result = KnowledgeHandlers::suggest(&rt, &token, params, &ann)
        .await
        .expect("suggest with deleted member");
    assert_eq!(result["total"], 1, "got: {result}");
    assert_eq!(result["results"][0]["id"], baseline["results"][0]["id"]);
    assert_eq!(result["results"][0]["size"], 0, "got: {result}");
    assert_eq!(result["results"][0]["members"], 0, "got: {result}");
    assert!(
        result["degraded"].get("member_sizing_timeout").is_none(),
        "a measured zero must not be excluded; got: {result}"
    );
    assert_ne!(result["degraded"]["lexical_timeout"], true);
    registry
        .dispatch(
            "knowledge.fold",
            json!({"candidates": result["results"], "budget": 10_000}),
        )
        .await
        .expect("member counts must preserve the unmodified suggest-to-fold input");
}

#[tokio::test]
async fn suggest_member_sizing_counts_only_live_members() {
    const LIVE_NAME: &str = "Sizing Live Member";
    const LIVE_BODY: &str = "Stored text supplies a valid live member with a measurable cost for each domain while the separate domain descriptions determine candidate relevance and ordering.";
    let rt = KhiveRuntime::memory().expect("in-memory runtime");
    let registry = build_registry(&rt);
    registry
        .dispatch(
            "knowledge.upsert_atoms",
            json!({"atoms": [
                {
                    "slug": "sizing-live-member",
                    "name": LIVE_NAME,
                    "content": LIVE_BODY,
                    "finalized": true
                },
                {
                    "slug": "sizing-removed-member",
                    "name": "Sizing Removed Member",
                    "content": LIVE_BODY,
                    "finalized": true
                }
            ]}),
        )
        .await
        .expect("upsert members");
    registry
        .dispatch(
            "knowledge.upsert_domains",
            json!({"domains": [{
                "slug": "sizing-mixed-domain",
                "name": "Sizing Mixed Domain",
                "description": "Lexical domain member sizing search uses the live content of member atoms to calculate the estimated token cost before a fold selection admits the domain into its budget.",
                "members": ["sizing-live-member", "sizing-removed-member"]
            }]}),
        )
        .await
        .expect("upsert domain");
    let deleted = registry
        .dispatch(
            "knowledge.delete_atoms",
            json!({"ids": ["sizing-removed-member"]}),
        )
        .await
        .expect("delete one member");
    assert_eq!(deleted["deleted"], 1);

    let token = rt.authorize(Namespace::local()).expect("authorize");
    let result = KnowledgeHandlers::suggest(
        &rt,
        &token,
        json!({"query": "lexical domain member sizing search", "limit": 1}),
        &vamana::new_shared(),
    )
    .await
    .expect("suggest mixed membership");
    assert_eq!(result["total"], 1, "got: {result}");
    assert_eq!(result["results"][0]["members"], 1, "got: {result}");
    assert_eq!(
        result["results"][0]["size"],
        crate::knowledge::util::estimate_compose_item_tokens(LIVE_NAME, LIVE_BODY),
        "deleted members must not contribute to the measured cost; got: {result}"
    );
    assert!(result["degraded"].get("member_sizing_timeout").is_none());
}

#[tokio::test]
async fn suggest_member_sizing_counts_each_distinct_live_atom_once() {
    const NAME: &str = "Sizing Shared Member Name";
    const BODY: &str = "Stored text supplies a valid live member with a measurable cost for each domain while the separate domain descriptions determine candidate relevance and ordering.";
    let rt = KhiveRuntime::memory().expect("in-memory runtime");
    let registry = build_registry(&rt);
    registry
        .dispatch(
            "knowledge.upsert_atoms",
            json!({"atoms": [
                {
                    "slug": "sizing-distinct-member-a",
                    "name": NAME,
                    "content": BODY,
                    "finalized": true
                },
                {
                    "slug": "sizing-distinct-member-b",
                    "name": NAME,
                    "content": BODY,
                    "finalized": true
                }
            ]}),
        )
        .await
        .expect("upsert distinct atoms with identical names and content");
    registry
        .dispatch(
            "knowledge.upsert_domains",
            json!({"domains": [
                {
                    "slug": "sizing-repeated-domain",
                    "name": "Sizing Repeated Domain",
                    "description": "Lexical domain member sizing search uses the live content of member atoms to calculate the estimated token cost before a fold selection admits the domain into its budget.",
                    "members": ["sizing-distinct-member-a", "sizing-distinct-member-a"]
                },
                {
                    "slug": "sizing-distinct-domain",
                    "name": "Sizing Distinct Domain",
                    "description": "Lexical domain member sizing search uses the live content of member atoms to calculate the estimated token cost before a fold selection admits the domain into its budget.",
                    "members": ["sizing-distinct-member-a", "sizing-distinct-member-b", "sizing-distinct-member-a"]
                }
            ]}),
        )
        .await
        .expect("upsert repeated and distinct membership");

    let token = rt.authorize(Namespace::local()).expect("authorize");
    let result = KnowledgeHandlers::suggest(
        &rt,
        &token,
        json!({"query": "lexical domain member sizing search", "limit": 2}),
        &vamana::new_shared(),
    )
    .await
    .expect("suggest domains with repeated and distinct members");
    assert_eq!(result["total"], 2, "got: {result}");
    let results = result["results"].as_array().expect("suggest results");
    let single_atom_size = crate::knowledge::util::estimate_compose_item_tokens(NAME, BODY);
    for (name, expected_members) in [("Sizing Repeated Domain", 1), ("Sizing Distinct Domain", 2)] {
        let domain = results
            .iter()
            .find(|domain| domain["name"] == name)
            .expect("domain is served");
        assert_eq!(domain["members"], expected_members, "got: {result}");
        assert_eq!(
            domain["size"],
            single_atom_size * expected_members,
            "got: {result}"
        );
    }
    assert!(result["degraded"].get("member_sizing_timeout").is_none());
}

#[tokio::test]
async fn suggest_member_sizing_serves_empty_or_missing_members_at_zero() {
    for members in [json!([]), json!(["absent-member"])] {
        let rt = KhiveRuntime::memory().expect("in-memory runtime");
        let registry = build_registry(&rt);
        registry
            .dispatch(
                "knowledge.upsert_domains",
                json!({"domains": [{
                    "slug": "sizing-empty-domain",
                    "name": "Sizing Empty Domain",
                    "description": "Lexical domain member sizing search uses the live content of member atoms to calculate the estimated token cost before a fold selection admits the domain into its budget.",
                    "members": members
                }]}),
            )
            .await
            .expect("upsert domain without live members");
        let token = rt.authorize(Namespace::local()).expect("authorize");
        let result = KnowledgeHandlers::suggest(
            &rt,
            &token,
            json!({"query": "lexical domain member sizing search", "limit": 1}),
            &vamana::new_shared(),
        )
        .await
        .expect("suggest domain without live members");
        assert_eq!(result["total"], 1, "got: {result}");
        assert_eq!(result["results"][0]["name"], "Sizing Empty Domain");
        assert_eq!(result["results"][0]["size"], 0, "got: {result}");
        assert_eq!(result["results"][0]["members"], 0, "got: {result}");
        assert!(result["degraded"].get("member_sizing_timeout").is_none());
        assert_ne!(result["degraded"]["lexical_timeout"], true);
    }
}

#[tokio::test]
async fn suggest_member_sizing_withholds_a_missing_domain_with_rank_and_score() {
    let rt = KhiveRuntime::memory().expect("in-memory runtime");
    let registry = build_registry(&rt);
    registry
        .dispatch(
            "knowledge.upsert_domains",
            json!({"domains": [
                {
                    "slug": "sizing-kept-domain",
                    "name": "Sizing Kept Domain",
                    "description": "Lexical domain member sizing search keeps measured domains available for token budget selection, because a fold budget can only admit domains whose compose cost is known.",
                    "members": []
                },
                {
                    "slug": "sizing-missing-domain",
                    "name": "Sizing Missing Domain",
                    "description": "Lexical domain member sizing search withholds domains whose canonical row cannot be measured and lists them under the degraded exclusion key with their rank and score.",
                    "members": []
                }
            ]}),
        )
        .await
        .expect("upsert domains");
    let token = rt.authorize(Namespace::local()).expect("authorize");
    let ann = vamana::new_shared();
    let params = json!({"query": "lexical domain member sizing search", "limit": 2});
    let baseline = KnowledgeHandlers::suggest(&rt, &token, params.clone(), &ann)
        .await
        .expect("suggest with both domain rows");
    assert_eq!(baseline["total"], 2, "got: {baseline}");
    let baseline_hits = baseline["results"].as_array().expect("baseline results");
    let missing_rank = baseline_hits
        .iter()
        .position(|hit| hit["name"] == "Sizing Missing Domain")
        .expect("domain to remove");
    let missing = &baseline_hits[missing_rank];
    let kept = baseline_hits
        .iter()
        .find(|hit| hit["name"] == "Sizing Kept Domain")
        .expect("domain to retain");

    // Retain the retrieval mirror so the hit list carries a domain that
    // the canonical sizing query can no longer reach.
    let access = rt.sql();
    let mut writer = access.writer().await.expect("writer");
    let removed = writer
        .execute(khive_storage::types::SqlStatement {
            sql: "DELETE FROM knowledge_domains WHERE namespace = 'local' AND id = ?1".into(),
            params: vec![khive_storage::types::SqlValue::Text(
                missing["id"].as_str().expect("domain id").to_owned(),
            )],
            label: None,
        })
        .await
        .expect("remove canonical domain row");
    assert_eq!(removed, 1);
    drop(writer);

    let result = KnowledgeHandlers::suggest(&rt, &token, params, &ann)
        .await
        .expect("suggest with a missing domain row");
    assert_eq!(result["results"], json!([kept]), "got: {result}");
    assert_eq!(result["total"], 1, "got: {result}");
    assert_eq!(
        result["degraded"]["member_sizing_timeout"]["excluded"],
        json!([{
            "id": missing["id"],
            "name": missing["name"],
            "rank": missing_rank + 1,
            "score": missing["score"]
        }]),
        "the missing domain must be withheld with its original rank and score, without a size; got: {result}"
    );
    assert_ne!(result["degraded"]["lexical_timeout"], true);
}

/// Issue #1930 Amendment 2: a lexical-stage-*only* timeout (its own narrow
/// budget expiring, not the request's) must not skip `search`'s embedding
/// rerank. Paused time advances by the production stage budget after the
/// first term, while the real handler runs under a 30-second request deadline.
/// Removing the independent stage scope must fail this test even though the
/// outer request remains healthy. The indexed candidate's stored-vector
/// provenance proves the rerank ran even when it needs no candidate embed.
#[tokio::test]
async fn search_still_reranks_after_a_lexical_stage_only_timeout() {
    let rt = rt_with_fake_embedder();
    let registry = build_registry(&rt);

    registry
        .dispatch(
            "knowledge.upsert_atoms",
            json!({
                "atoms": [{
                    "slug": "degrade-lexical-budget-atom",
                    "name": "Degrade Lexical Budget Atom",
                    "finalized": true,
                    "content": "lexicalbudgetsentinel transformer retrieval corpus benchmark search latency vector index nearest neighbor ranking fusion embedding cosine similarity attention encoder decoder positional normalization residual connection"
                }]
            }),
        )
        .await
        .expect("upsert atom");
    registry
        .dispatch("knowledge.index", json!({ "rebuild_ann": false }))
        .await
        .expect("index");

    // These lexical-only rows are seeded after indexing; fake time, not
    // corpus size or machine load, determines when the stage expires.
    crate::knowledge::search::seed_low_overlap_corpus(&rt, 1_000, 20).await;

    let ann = vamana::new_shared();
    let token = rt.authorize(Namespace::local()).expect("authorize");

    let query = "term0 term1 term2 term3 term4 term5 term6 term7";
    // Eight query words expand beyond the per-pass scaling cap (4x base).
    let stage_budget =
        std::time::Duration::from_millis(crate::knowledge::search::LEXICAL_STAGE_BUDGET_MS * 4);
    assert!(stage_budget < std::time::Duration::from_secs(30));
    tokio::time::pause();
    let result =
        khive_storage::scope_request_read_deadline(std::time::Duration::from_secs(30), async {
            let result = crate::knowledge::search::with_fts_deadline_advance_after_term(
                1,
                stage_budget,
                KnowledgeHandlers::search(&rt, &token, json!({ "query": query }), &ann),
            )
            .await;
            assert!(khive_storage::ensure_request_read_active("test.search_stage").is_ok());
            result
        })
        .await
        .expect("search must not Err on a lexical-stage-only timeout");

    assert_eq!(
        result["degraded"]["lexical_timeout"], true,
        "search must flag degraded.lexical_timeout when only the lexical \
         stage's own budget expires; got: {result}"
    );
    assert!(
        result["total"].as_u64().unwrap_or(0) > 0,
        "the seeded, ANN-indexed atom must still surface as a candidate; got: {result}"
    );
    assert_eq!(
        result["rerank_provenance"]["stored_vector_lookup"], "supported",
        "rerank must read persisted vectors after a lexical-only timeout; got: {result}"
    );
    assert!(
        result["rerank_provenance"]["from_stored"]
            .as_u64()
            .unwrap_or(0)
            > 0,
        "rerank must use the indexed atom's stored vector after a lexical-only timeout; got: {result}"
    );
    assert!(
        result["results"].as_array().is_some_and(|hits| hits
            .iter()
            .any(|hit| hit["score_provenance"]["embedding_rerank"].as_bool() == Some(true))),
        "a result must carry the applied embedding-rerank flag; got: {result}"
    );
}

include!("ann_degrade_tests/suggest_lexical_timeout.rs");

/// Fail-open applies to a genuine read-deadline timeout only. A storage error
/// that is *not* `StorageError::Timeout` — such as a malformed FTS5 MATCH
/// expression the parser rejects — must still surface as a hard error, never
/// be silently treated as a degraded-but-ok response.
#[tokio::test]
async fn malformed_fts5_match_expression_still_errors() {
    let rt = KhiveRuntime::memory().expect("in-memory runtime");
    let sql = rt.sql();
    let mut reader = sql.reader().await.expect("reader");

    // An unbalanced quote plus a dangling boolean operator is a genuine FTS5
    // parse error, never something `khive-pack-knowledge`'s own query
    // construction would emit (`quote_fts5_phrase` always escapes/balances
    // quotes) — this reproduces what a malformed expression from any other
    // caller of the same `fts_knowledge` table would surface as.
    let result = reader
        .query_all(khive_storage::types::SqlStatement {
            sql: "SELECT a.rowid FROM fts_knowledge \
                  JOIN knowledge_atoms AS a ON a.rowid = fts_knowledge.rowid \
                  WHERE fts_knowledge MATCH ?1 AND fts_knowledge.namespace = ?2 \
                  ORDER BY bm25(fts_knowledge) LIMIT 10"
                .to_string(),
            params: vec![
                khive_storage::types::SqlValue::Text("\"unterminated OR OR".to_string()),
                khive_storage::types::SqlValue::Text("local".to_string()),
            ],
            label: None,
        })
        .await;

    let err = result.expect_err("a malformed FTS5 MATCH expression must not silently succeed");
    assert!(
        !matches!(err, khive_storage::StorageError::Timeout { .. }),
        "a syntax error must not be classified as StorageError::Timeout, got {err:?}"
    );
}