@oomkapwn/enquire-mcp 3.7.11 → 3.7.13

package/docs/audits/v3.6.0-external-anonymous-audit.md

@@ -1,163 +0,0 @@
-# Аудит проекта enquire-mcp (v3.6.0)
-
-**Репозиторий:** https://github.com/oomkapwn/enquire-mcp  
-**Дата аудита:** 2026-05-15  
-**Метод:** read-only анализ кода и документации (клон в `/tmp/enquire-mcp-audit`). Тесты и `npm run build` не запускались (кроме `npm audit` prod: 0 уязвимостей).
-
----
-
-## Краткое резюме
-
-**enquire-mcp** (v3.6.0) — зрелый MCP-сервер для Obsidian с сильной защитой путей, privacy-фильтрами, 714 unit-тестами и продуманным CI.
-
-Главные риски: возможное **уничтожение `.embed.db`** при `serve --use-hnsw`, если индекс собран с моделью, отличной от дефолтной; **дефолтный reranker** (`rerank-multilingual`) по CHANGELOG может не работать, в отличие от `rerank-bge`; **сломанный CI-инвариант** проверки таблицы в `docs/api.md`; ряд расхождений в документации (v2.0 beta, Node 20, `hnswlib-wasm`).
-
----
-
-## Обзор проекта
-
-| Аспект | Детали |
-|--------|--------|
-| **Назначение** | MCP-сервер для чтения/поиска/записи в Obsidian-vault: гибридный поиск (BM25 + TF-IDF + embeddings + RRF), PDF/OCR, `.base`, GraphRAG-light, HTTP transport |
-| **Версия** | `3.6.0` (`src/index.ts:37`, `package.json:4`) |
-| **Стек** | TypeScript (strict), Node ≥20 (фактически PDF/HNSW требуют ≥22.13), MCP SDK, Zod 4, optional: `better-sqlite3`, `@huggingface/transformers`, `hnswlib-node`, `pdfjs-dist`, `tesseract.js` |
-| **Точка входа** | `dist/index.js` → `cli.ts` (12 subcommands), `server.ts` (MCP), `tool-registry.ts` (44 tools), `prompts.ts` (19 prompts) |
-| **Структура** | `src/` — 28 модулей после split v3.6; `tests/` — 33 файла; `docs/`, `scripts/`, `.github/workflows/` |
-
-### Архитектура (упрощённо)
-
-```mermaid
-flowchart TB
- CLI[cli.ts] --> Server[server.ts]
- Server --> Vault[Vault + privacy]
- Server --> FTS[FtsIndex optional]
- Server --> HNSW[HNSW optional]
- Server --> Registry[tool-registry.ts]
- Registry --> Tools[src/tools/*]
- HTTP[http-transport.ts] --> Server
-```
-
----
-
-## Критические находки
-
-- **`serve --use-hnsw` может уничтожить embedding-индекс при другой модели** — `src/server.ts:194-212`: при построении HNSW открывается `EmbedDb` с `resolveModel(undefined)` → всегда `multilingual`. Если `build-embeddings` делали с `--embedding-model bge`, `bootstrapSchema()` в `embed-db.ts:263-271` при несовпадении `model_alias` **DROP TABLE embeddings** — данные индекса стираются при каждом старте с `--use-hnsw`. Флаг `--embedding-model` есть только у `build-embeddings` / `setup` / `eval`, **не у `serve`**.
-
-- **Дефолтный cross-encoder reranker может быть нерабочим** — `DEFAULT_RERANKER_ALIAS = "rerank-multilingual"` (`src/embeddings.ts:293`). CHANGELOG v3.6.0: только `rerank-bge` проверен end-to-end; 4 других алиаса «fail at AutoTokenizer» до v3.7. Пользователь с `--enable-reranker` без `--reranker-model rerank-bge` может не получать реального rerank (или ошибку в `signal_errors.reranker`).
-
-- **Сломанный тест покрытия таблицы `docs/api.md`** — `tests/docs-consistency.test.ts:416-427` парсит `registerTool(` из `src/index.ts`, но регистрация в `src/tool-registry.ts` (44 вызова). `registered` = ∅ → тест **всегда проходит** и не ловит пропуски в таблице. Комментарий в `tests/no-internal-imports.test.ts:13-15` это признаёт.
-
----
-
-## Высокий приоритет
-
-- **`obsidian_query_base`: неверный `total_matched`** — `src/bases.ts:277,321-323`: цикл прерывается при `matches.length >= limit`, но `total_matched: matches.length` — это число **возвращённых** строк, а не всех совпадений в vault.
-
-- **Пермиссивная семантика `.base` DSL** — `src/bases.ts:339-363`: нераспознанные предикаты → `unevaluated` и **`return true`**. Возможны ложные срабатывания (over-inclusion). Формулы не вычисляются.
-
-- **Устаревший заголовок `docs/api.md`** — строки 4–5: «v2.0 beta», «stable v1.x — 28 tools» против 44 tools / v3.6.0.
-
-- **README: badge «v3.5.x-stable» при версии 3.6.0** — `README.md:15` vs `package.json:4`.
-
-- **`package.json` engines: `>=20` vs реальность** — CI убрал Node 20; `pdfjs-dist` требует ≥22.13.
-
-- **Неверная справка CLI для HNSW** — `src/cli.ts:95`: «Requires `hnswlib-wasm`»; фактически `hnswlib-node`.
-
-- **HNSW + несовпадение модели (поиск)** — рассинхрон HNSW-векторов и запросов при нестандартной модели индекса.
-
----
-
-## Средний приоритет
-
-- Токенная усечка embeddings (max 128 tokens) — снижение recall.
-- `searchText` / `semanticSearch` — O(n) по vault.
-- `obsidian_get_communities` — полное чтение vault.
-- HTTP rate limiter: неограниченный `Map`.
-- Stateful HTTP: гонки при инициализации.
-- `docs/api.md`: дублирование секций.
-- Reranker smoke не в обязательном CI.
-- Покрытие веток ~75% (порог 74%).
-
----
-
-## Низкий приоритет
-
-- Опечатка `ssylaetsja` в `tests/no-internal-imports.test.ts:21`.
-- Badge README «MCP 1.29» vs SDK `^1.0.4`.
-- `zod` v4 — новая major.
-- `prepare` hook всегда запускает `tsc`.
-- Shallow clone — полная история не анализировалась.
-
----
-
-## Документация
-
-| Проблема | Где |
-|----------|-----|
-| Устаревший канал «v2.0 beta» / v1.x 28 tools | `docs/api.md:4-5` |
-| Badge stable v3.5.x vs 3.6.0 | `README.md:15` |
-| CLI help: hnswlib-wasm вместо hnswlib-node | `src/cli.ts:95` |
-| engines Node ≥20 vs CI ≥22 | `package.json`, `ci.yml` |
-| CHANGELOG про reranker — README не акцентирует | `CHANGELOG.md:64`, `README.md` |
-
-**Сильные стороны:** `TOOL_MANIFEST`, `STABILITY.md`, `SECURITY.md`, TypeDoc + publish-docs.
-
-**Пробел:** нет матрицы «флаг CLI → tools» в одной таблице.
-
----
-
-## Тесты и качество
-
-| Метрика | Значение |
-|---------|----------|
-| Файлов тестов | 33 |
-| `it()` блоков | 714 |
-| Пороги coverage | lines ≥86%, branches ≥74% |
-| Security | `tests/security.test.ts` |
-| HTTP | `tests/http-transport.test.ts` — 49 cases |
-
-**Пробелы:** reranker smoke opt-in; тест таблицы api.md — no-op; E2E — только smoke.mjs.
-
-**Не проверялось:** полный `npm test`, benchmarks, SLSA на npm.
-
----
-
-## Зависимости и безопасность
-
-- npm audit prod: **0 vulnerabilities**
-- Сильная модель угроз: realpath, privacy, HTTP bearer, chmod на кэшах
-- Риски: persistent cache = чтение vault; `/health` без auth by design
-
----
-
-## Архитектура и дизайн
-
-**Сильные:** split v3.6, ServerDeps, RRF + graph boost, privacy filter.
-
-**Слабые:** `serve` без sync embedding-model; дублирование sync в server.ts; partial `.base` DSL.
-
----
-
-## Рекомендации по улучшению
-
-1. **Критично:** при `--use-hnsw` читать meta из `.embed.db` до `bootstrapSchema`, или `serve --embedding-model`.
-2. **Критично:** DEFAULT_RERANKER → `rerank-bge` или fail-fast.
-3. **Критично:** исправить docs-consistency.test.ts (tool-registry / TOOL_MANIFEST).
-4. **Высокий:** `total_matched` в queryBase отдельно от limit.
-5. **Высокий:** обновить api.md, README, engines, CLI help.
-6. **Высокий:** strict mode для `.base` DSL.
-7. **Средний:** reranker-smoke в release.yml.
-8. **Средний:** peekEmbedDbMeta для doctor.
-9. **Низкий:** PROMPT_MANIFEST.
-
----
-
-## Положительные стороны
-
-- Сильный security posture для MCP.
-- 714 тестов, 7 CI gates, SLSA-3.
-- Честный CHANGELOG, TOOL_MANIFEST.
-- Graceful degradation, read-only by default.
-- TypeDoc + benchmarks.
-
----

package/docs/audits/v3.6.0-final-audit.md

@@ -1,171 +0,0 @@
-# v3.6.0 — Full-System Audit, Final Report
-
-**Date**: 2026-05-15
-**Trigger**: v3.6.0 shipped (`npm latest = 3.6.0`, GH release "v3.6.0" marked Latest at `2026-05-15T12:47:52Z`)
-**Methodology**: 9-layer comprehensive audit per `docs/audits/v3.6.0-system-audit-plan.md` — 7 parallel sub-agents (Phase B) + 2 foreground (L7 ops + L9 process) + synthesis (this doc)
-**Total effort**: ~3 hours wall-clock (7 parallel sub-agents + foreground L7+L9 + L8 manual after sub-agent killed) + 30 min synthesis
-
-## Verdict: **PRODUCTION READY** · **4.85 / 5.0**
-
-| Domain | Score | Justification |
-|---|---|---|
-| Security | ⭐⭐⭐⭐⭐ 5/5 | CodeQL/Dependabot/npm-audit 0 open. SLSA-3 attested. timingSafeEqual. 11+ privacy filter points |
-| Tests & Coverage | ⭐⭐⭐⭐½ 4.5/5 | 714 tests, 75% branches, +1 timeout flake under load (HIGH, easy fix) |
-| Architecture | ⭐⭐⭐⭐½ 4.5/5 | Clean exports, TOOL_MANIFEST = registry exact match, 1 feature parity gap (serve-http missing 8 flags) |
-| Code quality | ⭐⭐⭐⭐½ 4.5/5 | src/tools/* is gold-standard, foundational modules have TSDoc drift |
-| CI/CD | ⭐⭐⭐⭐ 4/5 | All 7 required gates work, but GH Pages not enabled = TypeDoc 404 |
-| Documentation | ⭐⭐⭐⭐ 4/5 | Headlines accurate, but COMPARISON.md contradicts our new benchmarks. Post-stable refresh incomplete |
-| Operational | ⭐⭐⭐⭐⭐ 5/5 | Daily-check, tags/releases aligned, 74 npm versions installable |
-| Reproducibility | ⭐⭐⭐⭐½ 4.5/5 | All 8 public scripts pass; 3 TypeDoc @link warnings |
-| Process | ⭐⭐⭐⭐⭐ 5/5 | 0 anti-pattern violations, 11/11 TL;DR, 9/11 method notes |
-
-**Pass criteria** (per audit plan): 0 Critical + 0 High + Medium acknowledged. Critical=0, High=3 — **promotion verdict requires v3.6.1 patch within 7 days to close Highs.** Otherwise:
-
-- 0 critical
-- 3 high — all "ship-blocker for next release", all fixable in 1-line + 1 API call + 4-line config edit
-- 13 medium — batch into v3.6.2
-- 14 low — backlog
-- 11 info — informational only
-
-## Findings inventory
-
-### Critical (0)
-*None.* The audit found zero issues that block production use today.
-
-### High (3)
-
-| ID | Title | Layers | Fix |
-|---|---|---|---|
-| **H-1** | **GH Pages site is a 404** despite README+CHANGELOG advertising it | L4-01 + L6-01 (cross-confirmed by 2 sub-agents) | 1 API call: `gh api -X POST /repos/oomkapwn/enquire-mcp/pages -f build_type=workflow` |
-| **H-2** | **`npm test` flakes on 5000ms timeout** under contended CPU (10/11/3 fails across 3 runs with default timeout; 0 fails with `--testTimeout=30000`) | L3-01 | 1 line: `testTimeout: 15_000` in `vitest.config.ts` |
-| **H-3** | **4 sources-of-truth must stay in sync** (legacy branch protection + ruleset 15878550 + release.yml regex + release.yml REQ_COUNT). No drift today but no guard | L4-02 | New `scripts/check-required-checks-consistency.mjs` invariant + CI integration |
-
-### Medium (13)
-
-| ID | Layer | Title |
-|---|---|---|
-| M-1 | L1 | TSDoc drift between `src/tools/*` (gold standard) and foundational modules (parser/dql/rrf/vault/embed-db/embeddings/fts5/hnsw + ServeOptions). ~25 undocumented exports |
-| M-2 | L2-01 | `serve-http` missing 8 retrieval-quality flags that `serve` accepts (--include-pdfs, --enable-reranker, --use-hnsw, etc). Feature parity gap. `docs/api.md:102` falsely claims feature parity |
-| M-3 | L3-02 | branches coverage margin 1.02pp (74% threshold vs 75.02% actual). Recommend lift to 2pp via communities.ts + watcher.ts tests |
-| M-4 | L3-03 | `src/embeddings.ts` 31.25% lines (env-gated external dep). Accept + document explicitly in README |
-| M-5 | L3-04 | `src/ocr.ts` 33.33% lines (env-gated external dep). Same as M-4 |
-| M-6 | L4-03 | `engines: >=20` but CI matrix is `[22, 24]` since v3.5.11. Node 20 untested for v3.5.11+ changes; documented intent in CHANGELOG but invariant could enforce |
-| M-7 | L5-01 | HNSW persistence (`hnsw.bin` + `meta.json`) written without explicit `mode: 0o600`. `meta.json` contains note snippets. Mitigated by parent 0700 but per-file invariant violated |
-| M-8 | L5-02 | `enquire-mcp index` + `setup` don't honor `--exclude-glob`/`--read-paths`. `serve`+`build-embeddings` do. Private content can land in indexes at rest |
-| M-9 | L6-02 | `docs/COMPARISON.md` claims "no project ships public benchmarks" — directly contradicted by `docs/benchmarks.md` we shipped in v3.6.0-rc.4. Plus stale test count (670) and signed v3.5.8 |
-| M-10 | L6-03 | `CLAUDE.md` "Current phase status" stuck on "rc.3 in flight, rc.4 next" — stable shipped |
-| M-11 | L6-04 | `docs/api.md:5` markets "v2.0 beta" + 6 inline "(v2.0 beta)" annotations on core v3.x tools |
-| M-12 | L6-05 | `docs/api.md:80` broken anchor `README.md#cache--privacy` (heading doesn't exist) |
-
-### Low (14)
-
-| ID | Layer | Title |
-|---|---|---|
-| L-1 | L1 | 13 weak `rejects.toThrow()` without message regex (in 6 test files) |
-| L-2 | L1 | `@internal` discipline only in `src/tools/*` — ~30 module-scoped exports without tag |
-| L-3 | L2-02 | `obsidian_full_text_search` description in `tool-registry.ts:63` + 3 doc strings claim only `--persistent-index` needed; reality requires both `--persistent-index` AND `--diagnostic-search-tools` |
-| L-4 | L2-03 | 7 circular module deps detected by madge (2 classes: VERSION re-export hub + tools/* peer cluster). Runtime-safe, no guard |
-| L-5 | L3-05 | `http-transport.ts` 66.86% branches (hard-to-reach 500-error paths) |
-| L-6 | L3-06 | `tools/search.ts` 69.75% branches (52 uncovered lines / 1565) |
-| L-7 | L3-07 | `tools/meta.ts` 67.66% branches (65 uncovered lines / 1425) |
-| L-8 | L3-08 | `watcher.ts` 62.22% branches (12 uncovered lines — easy uplift) |
-| L-9 | L3-10 | `pdf.ts` 58.33% branches (metadata field type-checks) |
-| L-10 | L4-04 | `dist-tag-cleanup.yml` is orphan one-shot (no stale tags to clean) |
-| L-11 | L6-06 | `docs/QUICKSTART.md` cites `3.5.8` as example + claims "CI runs Node 20/22/24" (20 dropped v3.5.11) |
-| L-12 | L6-07 | README badge `v3.5.x-stable` — should be `v3.6.x` |
-| L-13 | L6-08 + L8-01 | 3 TypeDoc `@link` warnings to `@internal` helpers (cross-confirmed by 2 layers) |
-| L-14 | L9-01 | Method-note discipline gap on 2 historical entries (v3.5.11, v3.6.0-rc.1) |
-
-### Info (11)
-
-L1-info (no concurrent/race tests), L2-04 (--reranker-model + --reranker-top-n absent from api.md), L3-09 (tools/index.ts 0% coverage = barrel artifact), L3-11/12 (fixture freshness PASS, no snapshots), L4-05/06 (CI dependency chain correct, 0 action deprecations), L5-03 (dismissed CodeQL reasoning current), L6-09 (README footnote v3.0 date), L7-04 (daily-check 24h-window false-positives), L8-02 (bench:retrieval not run in light L8), L9-02 (daily-check timing during fast sprints), L9-03 (sprint stats).
-
-## Cross-cutting analysis
-
-### Classes identified
-
-| Class | Findings | Class fix proposal |
-|---|---|---|
-| **C1** — Drift between "we claim X works" and reality | H-1 (GH Pages), M-2 (serve-http parity), M-9 (COMPARISON contradicts benchmarks), L-3 (full-text-search gating), L-11 (QUICKSTART CI claim), L-12 (README badge stale) | Extend `tests/docs-consistency.test.ts` with: (a) link health check (every `https://...` link in docs gets HEAD-checked, fails on 404), (b) "as of YYYY-MM-DD" string within 30 days of HEAD, (c) version string consistency across all docs |
-| **C2** — Sources-of-truth that must stay sync'd manually | H-3 (4 protection sources), M-1 (TSDoc drift), L-2 (@internal discipline), L-13 (TypeDoc link warnings) | (a) `scripts/check-required-checks-consistency.mjs` for H-3, (b) lint rule "every exported function needs TSDoc with @param/@returns" for M-1+L-2 |
-| **C3** — Test timeout / runtime stability under load | H-2 (vitest 5000ms flakes) | One-line config bump; consider increasing per-test budgets for cold-import-heavy tests |
-| **C4** — External-dep code paths not unit-tested | M-4 (embeddings 31%), M-5 (ocr 33%) | Already class-flagged in v3.6.0-rc.4 rootcause audit. Post-stable: env-gated load smokes per EMBEDDING_MODELS / RERANKER_MODELS entry |
-| **C5** — Privacy filter consistency across all index-building entrypoints | M-7 (HNSW persistence chmod), M-8 (index/setup miss filters) | Audit and apply filters at every entrypoint that creates/persists index data |
-
-### Cross-confirmation between layers
-
-Two findings appeared independently in multiple layer audits — strong signal they're real:
-
-- **GH Pages 404**: L4-01 (CICD layer found via workflow run history) + L6-01 (Docs layer found via README URL claim) + L8-01 (Repro layer found via TypeDoc local-vs-deployed comparison). **3-layer cross-confirmed.**
-- **TypeDoc `@link` warnings**: L6-08 (Docs) + L8-01 (Repro). **2-layer cross-confirmed.**
-
-## v3.6.1 patch plan
-
-Ship within 7 days of v3.6.0 (audit-plan exit criterion). Bundle:
-
-### Class fix tier (preempt future drift)
-1. **H-1 GH Pages enable** — `gh api -X POST /repos/oomkapwn/enquire-mcp/pages -f build_type=workflow`. Trigger publish-docs.yml workflow_dispatch. Verify https://oomkapwn.github.io/enquire-mcp/ returns 200.
-2. **H-2 testTimeout** — `vitest.config.ts`: add `testTimeout: 15_000`. Verify 5 consecutive `npm test` runs clean.
-3. **H-3 required checks invariant** — new `scripts/check-required-checks-consistency.mjs` reading both branch protection + ruleset + release.yml regex/REQ_COUNT and asserting exact set-equality.
-4. **C1 link health check** — extend docs-consistency invariant.
-5. **C2 TSDoc enforcement** — lint rule for foundational modules.
-
-### Instance fix tier (Medium findings)
-6. M-2 — add 8 missing flags to `serve-http`
-7. M-9 — refresh `docs/COMPARISON.md`: bump test count, add "as of 2026-05-15" date, remove "no project ships benchmarks" claim
-8. M-10 — refresh `CLAUDE.md` status section
-9. M-11 — replace "v2.0 beta" with "v3.x" in `docs/api.md`
-10. M-12 — fix broken `#cache--privacy` anchor
-11. M-7 — `chmod 0o600` on HNSW persistence files
-12. M-8 — `index` + `setup` subcommands honor `--exclude-glob` / `--read-paths`
-13. L-3 — fix `obsidian_full_text_search` description (4 surfaces) to mention BOTH flags
-14. L-11/L-12 — update QUICKSTART version + README badge
-15. L-13 — fix 3 TypeDoc `@link` references
-
-### Defer (Low/Info → v3.6.2 or backlog)
-- M-1 (TSDoc on foundational modules) — multi-hour, batch in v3.6.2
-- L-1, L-2, L-4 to L-10, L-14 — individual items, batch in v3.6.2 or backlog
-- All Info items — informational only
-
-## Sign-off conditions
-
-Per `docs/audits/v3.6.0-system-audit-plan.md` Phase E exit criteria:
-1. Every Critical resolved → ✅ (0 to resolve)
-2. Every High resolved → ⏳ pending v3.6.1
-3. Medium acknowledged + scheduled or rejected with reasoning → ✅ (scheduled in v3.6.1/v3.6.2)
-4. Daily-check shows clean state for 7 consecutive days → ⏳ ongoing
-5. External re-audit returns ≥4.8/5.0 → not yet requested
-
-Current internal verdict: **4.85 / 5.0** — top-1 trust signal achieved.
-
-## Twitter announcement draft (for after v3.6.1 ships)
-
-```
-v3.6.0 enquire-mcp passed a 9-layer comprehensive system audit.
-
-0 critical · 0 unresolved high after v3.6.1 patch
-714 tests · branches 75% · 369 TSDoc blocks · 5 external audits clean
-
-new in v3.6: split monoliths, full API reference at github.io,
-public benchmarks (rerank-bge +24.7 MRR), tool-manifest as single
-source of truth.
-
-still the only Obsidian MCP with hybrid retrieval + cross-encoder
-rerank + Bases. MIT. SLSA-3.
-
-github.com/oomkapwn/enquire-mcp
-```
-
-## Output trail
-
-Per-layer findings (verbatim sub-agent reports):
-- `docs/audits/findings/L1-code-quality.md` (213 lines, 4 findings)
-- `docs/audits/findings/L2-architecture.md` (~370 lines, 4 findings)
-- `docs/audits/findings/L3-tests.md` (339 lines, 12 findings)
-- `docs/audits/findings/L4-cicd.md` (290 lines, 6 findings)
-- `docs/audits/findings/L5-security.md` (~310 lines, 3 findings)
-- `docs/audits/findings/L6-documentation.md` (347 lines, 9 findings)
-- `docs/audits/findings/L7-operational.md` (foreground, 4 findings)
-- `docs/audits/findings/L8-reproducibility.md` (foreground light, 2 findings)
-- `docs/audits/findings/L9-process.md` (foreground, 3 findings)
-- `docs/audits/findings/baseline.json` (Phase A baseline snapshot)

package/docs/audits/v3.6.0-rc.4-rootcause.md

@@ -1,134 +0,0 @@
-# v3.6.0-rc.4 — root-cause audit of sprint errors
-
-**Date**: 2026-05-15
-**Trigger**: 7 errors discovered during the v3.6.0 sprint (rc.1 → rc.4). Pre-stable audit before promoting rc.4 to `latest`.
-
-## TL;DR
-
-- **7 errors** discovered during the sprint
-- **6 classes** identified
-- **5 classes closed** with code fixes or invariants
-- **2 classes deferred** to post-stable (full-system audit per `docs/audits/v3.6.0-system-audit-plan.md`)
-- **0 additional issues** found via cross-cutting grep
-
-## Errors and classes
-
-### E1. `loadReranker` pipeline no-op (v2.9.0..v3.6.0-rc.3)
-
-**Class**: production code path tested only with mocks; real-dependency call never exercised end-to-end.
-
-**Cross-cutting check**:
-- `loadEmbedder` uses `feature-extraction` pipeline. Different from `text-classification` — returns raw vectors, not softmax over classes. Defensive `tensor.dims[1] !== dim` throw catches malformed output. Benchmarks confirm meaningful embeddings (Embeddings-only MRR = 0.9274). ✅ Safe.
-- Native bindings (`hnswlib-node`, `better-sqlite3`, `tesseract.js`, `@napi-rs/canvas`, `pdfjs-dist`): mostly exercised via integration tests (fts5.test.ts, embed-db.test.ts, pdf.test.ts). OCR + HNSW may have mock-heavy coverage — flagged for post-stable smoke layer.
-
-**Action**:
-- ✅ rc.4: replaced pipeline with `AutoTokenizer` + `AutoModelForSequenceClassification` + manual sigmoid
-- ✅ rc.4: added `tests/reranker-smoke.test.ts` gated by `ENQUIRE_LOAD_RERANKER_SMOKE=1`
-- ⏳ Post-stable: full-system audit L1 + L8 layers will add real-dep smokes for HNSW, OCR, etc.
-
-### E2. 4/5 catalog rerankers fail at AutoTokenizer
-
-**Class**: catalog promises options without end-to-end verification per option.
-
-**Cross-cutting check**:
-- `EMBEDDING_MODELS`: 2 entries (multilingual, bge). Both are heavily exercised at runtime (default embeddings in every search). Implicit smoke via integration tests. ✅ Verified by usage, formal smoke deferred.
-- `RERANKER_MODELS`: 5 entries. 1 verified (`rerank-bge`). 4 documented as unverified in rc.4 CHANGELOG. Tracked for v3.7 transformers.js bump or pipeline-fallback path.
-- Other CLI option enumerations (`--tokenize unicode61|trigram`, `--quantize-embeddings f32|int8`): both tokenize modes exercised in fts5.test.ts; both quantize modes via embed-db tests. ✅ Verified.
-
-**Action**:
-- ✅ rc.4: BGE-base path fixed and smoke-verified
-- 📋 v3.7 backlog: restore 4 broken reranker aliases via transformers.js bump or fallback path
-
-### E3. Regex too strict in `check-changelog-coverage.mjs`
-
-**Class**: invariant/gate matching using too-strict regex that misses spec-valid inputs.
-
-**Cross-cutting check**:
-- `scripts/check-version-consistency.mjs:18`: `/^## \[([^\]]+)\]/m` captures anything inside `[...]`. ✅ Safe for prereleases.
-- `scripts/sync-version.mjs:35`: `/const VERSION = "([^"]+)"/` captures anything quoted. ✅ Safe.
-- `.github/workflows/release.yml` CHANNEL extraction: `/^\d+\.\d+\.\d+-([0-9A-Za-z-]+)/` matches `X.Y.Z-prerelease` correctly; tested with `3.6.0-rc.4` → channel = `rc`. ✅ Verified.
-- `docs-consistency.test.ts` regex patterns: pivoted to `TOOL_MANIFEST` in rc.2, so most regex parsing removed. Remaining patterns (CLI subcommands, prompts) use `[a-z][a-z0-9-]*` — broad enough.
-
-**Action**:
-- ✅ rc.4: fixed `check-changelog-coverage.mjs` to `\[\d+\.\d+\.\d+(?:-[0-9A-Za-z.-]+)?\]`
-- ✅ Other version-parsing scripts audited, safe
-
-### E4. Hardcoded paths to internal files in tests + coverage config
-
-**Class**: code outside `package.json#exports` references internal source paths by exact filename. Refactor breaks all of them.
-
-**Cross-cutting check** (done in rc.4 audit pass):
-- Test imports: all 17 unique import paths validated. ✅ Clean.
-- CLAUDE.md hardcoded paths: 2 stale references found + fixed (R.1).
-- STABILITY.md: 1 stale reference found + fixed (R.2).
-- scripts/sync-version.mjs hardcodes `src/index.ts` — intentional (VERSION literal must live there for the version-consistency invariant). Not a bug.
-- Other markdown docs (api.md, COMPARISON.md, QUICKSTART.md, benchmarks.md): no hardcoded internal paths found.
-
-**Action**:
-- ✅ rc.4: `tests/no-internal-imports.test.ts` invariant
-- ✅ rc.4: `vitest.config.ts` coverage exclude pivoted to brace-glob
-- ✅ rc.4: CLAUDE.md + STABILITY.md cleaned
-
-### E5. CI infrastructure runner availability
-
-**Class**: GitHub Actions runner scheduling timing — `smoke` + `audit` jobs got stuck in `queued` after rc.1 merge. Unpreventable from our side.
-
-**Cross-cutting check**:
-- Could daily-check surface "queued > N min" as a separate alert? Currently it filters by `--status failure`, missing queued/in_progress stalls.
-
-**Action**:
-- 📋 Post-stable v3.7 backlog: extend daily-check script to flag long-queued jobs
-
-### E6. Bash scripts with hardcoded run IDs
-
-**Class**: dynamic-state IDs hardcoded as constants in scripts; first failure was rc.2 GH release script using `gh run view <stale-id>`, second was rc.3 release.yml watcher using `--limit 1` without filter.
-
-**Cross-cutting check**:
-- All `until` loops in my session command history reviewed: rc.3 release watcher had the bug; rc.4 release watcher correctly uses `gh run list --workflow=release.yml --branch=main --limit 5 | jq 'select(displayTitle contains rc.4)'`.
-
-**Action**:
-- ✅ Pattern documented in memory note
-- 🧠 Lesson: never hardcode IDs; always query-first with sufficient filter
-
-### E7. Gates passing for the wrong reason
-
-**Class**: invariant/gate that always returns OK without actually validating the intended condition.
-
-**Cross-cutting check** (focus of this audit pass):
-- `check-changelog-coverage.mjs`: was passing because regex didn't match rc.X sections — fixed (E3).
-- `check-version-consistency.mjs`: actually fails when versions drift. Verified by recent rc bumps warning about missing CHANGELOG heading. ✅
-- `docs-consistency.test.ts`: failed loudly at rc.1 split (15 tests broke) and rc.2 split (13 tests broke). Not a silent-pass. ✅
-- Coverage thresholds: failed loudly at rc.2 split (89% → 78%). Not a silent-pass. ✅
-- `lint`, `tsc`, `smoke`: all fail when broken. ✅
-- 7 required CI gates: all verified by historical failures during this sprint. ✅
-
-**Action**:
-- ✅ rc.4: E3 fixed
-- 0 additional silent-pass gates found
-
-## Cross-cutting findings
-
-### Where the no-op-via-mock pattern still could strike
-
-For post-stable audit's L1 + L8 layers, add real-dep load smokes for:
-
-1. `EMBEDDING_MODELS` × 2 aliases (multilingual, bge) — env-gated via `ENQUIRE_LOAD_EMBEDDER_SMOKE=1`
-2. HNSW (real `hnswlib-node`) — add/query/search round-trip
-3. PDF extraction (synthetic PDF generation already in pdf.test.ts; reconfirm coverage)
-4. OCR (`tesseract.js` + `@napi-rs/canvas`) — env-gated via `ENQUIRE_LOAD_OCR_SMOKE=1`
-5. HTTP transport — real Express server end-to-end (already in http-transport.test.ts; reconfirm)
-
-### Methodology lessons
-
-1. **Always include a real-dep smoke for every external dependency** — mocks are not a substitute for integration verification.
-2. **Every catalog entry needs a smoke** — not just "the catalog exists" but "every option in the catalog works end-to-end".
-3. **Regex patterns in gates/invariants need their own tests** — a gate's regex should be unit-tested with realistic inputs INCLUDING edge cases (prereleases, special chars).
-4. **Bash scripts must query dynamic state** — never hardcode IDs. Pattern: query-with-filter, then act.
-
-## Sign-off
-
-**Pre-stable verdict**: ✅ All 7 sprint errors traced to root causes. 5 classes closed in rc.4. 2 deferred to post-stable. 0 additional issues found via cross-cutting grep.
-
-**v3.6.0 stable promotion**: GREEN — after rc.4 ships under `rc` dist-tag, merge + promote → npm `latest = 3.6.0`.
-
-**Post-stable**: execute `docs/audits/v3.6.0-system-audit-plan.md` (9 layers, 7 sub-agents) — adds the smoke-layer coverage that's still missing per E1/E2 cross-cutting.