test(memory): establish semantic retrieval no-go baseline

[SunwardL]

Summary

• Add a versioned, synthetic P2-A retrieval dataset with 31 Memory records and 40 English/Chinese queries, including semantic paraphrase, cross-lingual, scope, lifecycle, authority, temporal, abstention, and multi-record cases.
• Add strict dataset/manifest validation, a common safety prefilter, deterministic metrics, resource accounting, an offline network guard, and a frozen lexical reference baseline.
• Evaluate two local q8 multilingual candidates through an evaluation-only adapter and record the locked E5 hybrid candidate, fallback behavior, Windows x64 resource evidence, and deterministic clean-process runs.
• Record a no-go decision: the candidate passes every safety, offline, determinism, precision, lexical-control, and desktop-resource gate, but does not meet the pre-registered holdout Recall@5 and MRR improvement thresholds.

This PR intentionally does not change production Memory retrieval. SQLite FTS5 and filesystem fallback remain the supported path.

Why a no-go result is useful

The decision thresholds were frozen before holdout execution. The selected candidate improved holdout Recall@5 from 0.500 to 0.625 (+0.125), below the required +0.150, and holdout MRR stayed at 0.500, below the required +0.100 improvement. Lowering the thresholds or tuning against the revealed holdout would invalidate the experiment, so P2-B is not proposed from this candidate.

Changes

• Add strict schemas and integrity checks for the frozen dataset and manifest.
• Add a common authorized-record prefilter before any candidate sees records.
• Add split/language/category Recall@K, Precision@K, MRR, abstention, false-positive, and hard-safety accounting.
• Add deterministic resource percentile helpers and an in-process offline network guard.
• Add the lexical control adapter and split machine-readable baseline artifacts.
• Add an evaluation-only normalized-vector adapter with deterministic ties, semantic-gated RRF, and exact lexical fallback for missing, corrupt, unsupported, or failed candidate initialization.
• Add development screening, resource evidence, the candidate lock, holdout/full decision evidence, and a threshold-by-threshold decision evaluator.
• Add a dedicated npm --prefix kun run eval:memory-retrieval command.
• Add a complete OpenSpec proposal, design, specification, candidate screening record, and task ledger.

Frozen candidate and result

Item	Result
Candidate	multilingual-e5-small-q8-hybrid-0.8-1
Model revision	761b726dd34fb83930e26aab4e9ac3899aa1fa78
Quantized ONNX SHA-256	f80102d3f2a1229f387d3c81909990d8945513e347b0eab049f7de3c6f98c193
Fusion	similarity >= 0.80, equal-weight RRF, rank constant 60
Holdout Recall@5	0.625 vs lexical 0.500 (+0.125; required +0.150)
Holdout MRR	0.500 vs lexical 0.500 (+0.000; required +0.100)
Full Recall@5 / Precision@5 / MRR	0.833 / 0.345 / 0.729
Safety violations	0 scope, lifecycle, authority, unknown, network, or fallback mismatches
Determinism	3 fresh processes, identical report SHA-256 4bee9d01b3bd48fe3da18917bd009b3fdb8b105681d653e377c6427dc608eada
Decision	no-go; keep SQLite FTS5/filesystem and do not start P2-B

Windows x64 resource evidence

• Cold readiness: 584.333 ms
• Warm query p50/p95: 3.581 / 4.738 ms over 30 samples
• 10,000-record index build: 17,011.378 ms
• Incremental update: 5.178 ms
• Model assets: 135,392,016 bytes
• Generated 10,000-record index: 15,590,000 bytes
• Peak RSS delta: 351,342,592 bytes
• Network attempts: 0

All frozen resource limits passed. Model assets and the generated stress index remain outside the repository.

Tests

• npm --prefix kun run eval:memory-retrieval — 7 files, 38 tests passed
• npm run build:kun — passed
• npm run typecheck — passed
• npm run build — passed
• npm run check:file-lines — 6,947 tracked text files passed
• npm run lint — 0 errors; 30 existing warnings
• npm exec -- openspec validate add-kun-semantic-memory-retrieval --strict — passed
• git diff --check — passed

The complete npm test run finished with 5,751 passed, 27 skipped, and 22 failures in seven files outside this change. Seven Chromium-cookie failures were caused by the isolated runner PATH omitting the installed sqlite3.exe; both cookie files passed on focused rerun after restoring the normal user PATH (22 tests). Eleven failures remain reproducible in five unchanged upstream files: eight file-session index assertions, two Windows/POSIX path expectations, and one POSIX permission-bit expectation. No failed file or its production module is changed here.

Non-goals

• No production embedding index or hybrid ranking wiring
• No canonical Memory schema change
• No model download, model binary, generated vector index, or new runtime dependency
• No hosted Memory upload or LLM reranking
• No hidden feature flag that can activate the candidate
• No real user Memory, account data, machine path, credential, or conversation excerpt