doubled-graph 2.0.2__tar.gz → 2.1.0__tar.gz

{doubled_graph-2.0.2 → doubled_graph-2.1.0}/INDEX.md

@@ -49,7 +49,7 @@
 |---|---|
 | Новый проект | [prompts/new-project/](prompts/new-project/) (8 файлов) |
 | Миграция | [prompts/migrate-existing-project/](prompts/migrate-existing-project/) (7 файлов) |
-| Обновление методики (1.9.x → 2.0) | [prompts/upgrade-methodology/](prompts/upgrade-methodology/) (2 файла) |
+| Обновление методики (2.0.x → 2.1) | [prompts/upgrade-methodology/](prompts/upgrade-methodology/) (2 файла) |
 | Поддержка | [prompts/maintenance/](prompts/maintenance/) (4 файла) |
 
 ---
@@ -59,9 +59,10 @@
 | Тема | Файл |
 |---|---|
 | Архитектура | [docs/SPEC.md](docs/SPEC.md) |
-| Контракт 4 публичных tool'ов | [docs/TOOLS.md](docs/TOOLS.md) |
+| Контракт MCP-tools (10) | [docs/TOOLS.md](docs/TOOLS.md) |
 | Hooks (git + Claude Code + optional) | [docs/HOOKS.md](docs/HOOKS.md) |
-| Исходник MVP | [src/doubled_graph/](src/doubled_graph/) |
+| Verification / Definition of Done | [docs/VERIFICATION.md](docs/VERIFICATION.md) |
+| Исходник | [src/doubled_graph/](src/doubled_graph/) |
 | Smoke-тесты | [tests/](tests/) |
 
 ---

{doubled_graph-2.0.2 → doubled_graph-2.1.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: doubled-graph
-Version: 2.0.2
+Version: 2.1.0
 Summary: MCP facade over CodeGraphContext + GRACE 4 (.grace artifacts): impact, context, drift-detect, migration for LLM-driven codebases.
 Project-URL: Homepage, https://github.com/rour/doubled-graph-mcp
 Project-URL: Repository, https://github.com/rour/doubled-graph-mcp
@@ -23,12 +23,14 @@ Classifier: Topic :: Software Development :: Code Generators
 Classifier: Topic :: Software Development :: Documentation
 Requires-Python: >=3.11
 Requires-Dist: anyio>=4.15.1
-Requires-Dist: codegraphcontext>=0.6.13
+Requires-Dist: codegraphcontext<0.7,>=0.6.13
 Requires-Dist: mcp<2,>=1.30.0
 Requires-Dist: pydantic>=2.14.1
 Requires-Dist: tree-sitter-language-pack>=1.22.1
 Provides-Extra: dev
+Requires-Dist: mutmut<4,>=3.8.1; extra == 'dev'
 Requires-Dist: pytest-asyncio>=0.23; extra == 'dev'
+Requires-Dist: pytest-cov>=5; extra == 'dev'
 Requires-Dist: pytest>=8.0; extra == 'dev'
 Requires-Dist: ruff>=0.5; extra == 'dev'
 Description-Content-Type: text/markdown
@@ -42,7 +44,7 @@ MCP-сервер для LLM-агентов, работающих с больши
 - **Computed graph** — AST-граф реального кода (через CodeGraphContext): кто вызывает что, где определено.
 - **Declared graph** — граф из артефактов GRACE 4 (`.grace/`): замысел, контракты, верификация.
 
-Четыре MCP-инструмента:
+Четыре доменных MCP-инструмента (плюс `health`, `migrate_check` и четыре gateway над `grace`):
 
 | Инструмент | Когда |
 |---|---|
@@ -51,6 +53,8 @@ MCP-сервер для LLM-агентов, работающих с больши
 | `context` | 360° вид: символ + контракт + вызовы |
 | `detect_changes` | дрейф кода от артефактов — **до** коммита |
 
+`detect_changes` возвращает шесть типов дрейфа: `code_without_module`, `module_without_code`, `contract_mismatch`, `stale_crosslinks`, `missing_verification`, `markup_missing`. `contract_mismatch` сверяет `<Exports>`, `MODULE_MAP` и `INPUTS` контрактов с AST-сигнатурой (имена; типы — только если CGC их отдаёт), `stale_crosslinks` — с реальными `CALLS`/`IMPORTS`-рёбрами между файлами модулей.
+
 ---
 
 ## Установка
@@ -104,6 +108,14 @@ doubled-graph setup
 - установит `git post-commit` хук для автообновления графа
 - запустит первый полный `analyze`
 
+> **Важно:** `setup` при каждом запуске **перезаписывает** скопированные
+> `prompts/` и `methodology/` версиями из установленного пакета
+> (`_copy_bundle` всегда пишет файлы заново, чтобы после `pipx upgrade`
+> не оставались устаревшие workflow). Не редактируйте их на месте — правки
+> будут потеряны при следующем `setup`. Для локальных изменений скопируйте
+> нужный файл в отдельный путь проекта и ссылайтесь на него оттуда;
+> исходные `prompts/`/`methodology/` в репозитории считайте managed-контентом.
+
 Посмотреть план без реальных изменений:
 
 ```bash
@@ -149,12 +161,12 @@ head -1 "$(which doubled-graph)"  # вывести shebang — это и ест
 
 Сценарий 2. Миграция существующего — `prompts/migrate-existing-project/00-entry-prompt.md`.
 
-Сценарий 3. Обновление версии методики (1.9.x → 2.0) — `prompts/upgrade-methodology/00-entry-prompt.md`.
+Сценарий 3. Обновление версии методики (2.0.x → 2.1) — `prompts/upgrade-methodology/00-entry-prompt.md`.
 
 ### 4. Проверить, что всё встало
 
 ```bash
-doubled-graph --version          # 2.0.2
+doubled-graph --version          # 2.1.0
 doubled-graph status --repo .    # artifacts.project_kind = grace4, grace_engine.upgrade_required = false
 doubled-graph lint --repo .      # summary.errors = 0
 doubled-graph migrate check --repo .   # map_pending / placeholders — что осталось заполнить
@@ -164,6 +176,10 @@ doubled-graph migrate check --repo .   # map_pending / placeholders — что
 Если `upgrade_required` = `true` — движок старее 4.x: `npm install -g @osovv/grace-cli@4.2.1`, затем `setup` повторно.
 Если `lint` вернул `markup.module-map-missing` — это содержательная работа, промпт `prompts/upgrade-methodology/01-content-gaps.md`.
 
+Полная проверка (ruff + unit + real-CGC E2E + diff/version): `bash scripts/verify.sh`; критерии готовности — [docs/VERIFICATION.md](docs/VERIFICATION.md).
+
+В 2.1: `detect_changes` проверяет `<Exports>`/`MODULE_MAP`/`INPUTS` и реальные CALLS/IMPORTS-рёбра, `impact` отдаёт `anchor`, `include_unchanged` расширяет охват, а `DG-Authored`-трейлер ставится через `prepare-commit-msg`. Установить все хуки: `doubled-graph init-hooks --all` (git + Claude Code + provenance); без Claude Code — `--post-commit --prepare-commit-msg`.
+
 
 ---
 
@@ -234,6 +250,8 @@ doubled-graph setup          # 2. обновить движок и подгот
 создаёт недостающие обязательные документы, добавляет `LINKS: M-*` в
 `MODULE_CONTRACT` (без этого файл не привязан к модулю), приводит `MAP_MODE`
 в соответствие с `ROLE`. Legacy `docs/` не изменяются и не удаляются.
+`setup` также дописывает в `AGENTS.md` управляемый блок anti-patterns
+(идемпотентно; существующий текст и BOM сохраняются).
 
 Если миграцию не сделать:
 

{doubled_graph-2.0.2 → doubled_graph-2.1.0}/README.md

@@ -7,7 +7,7 @@ MCP-сервер для LLM-агентов, работающих с больши
 - **Computed graph** — AST-граф реального кода (через CodeGraphContext): кто вызывает что, где определено.
 - **Declared graph** — граф из артефактов GRACE 4 (`.grace/`): замысел, контракты, верификация.
 
-Четыре MCP-инструмента:
+Четыре доменных MCP-инструмента (плюс `health`, `migrate_check` и четыре gateway над `grace`):
 
 | Инструмент | Когда |
 |---|---|
@@ -16,6 +16,8 @@ MCP-сервер для LLM-агентов, работающих с больши
 | `context` | 360° вид: символ + контракт + вызовы |
 | `detect_changes` | дрейф кода от артефактов — **до** коммита |
 
+`detect_changes` возвращает шесть типов дрейфа: `code_without_module`, `module_without_code`, `contract_mismatch`, `stale_crosslinks`, `missing_verification`, `markup_missing`. `contract_mismatch` сверяет `<Exports>`, `MODULE_MAP` и `INPUTS` контрактов с AST-сигнатурой (имена; типы — только если CGC их отдаёт), `stale_crosslinks` — с реальными `CALLS`/`IMPORTS`-рёбрами между файлами модулей.
+
 ---
 
 ## Установка
@@ -69,6 +71,14 @@ doubled-graph setup
 - установит `git post-commit` хук для автообновления графа
 - запустит первый полный `analyze`
 
+> **Важно:** `setup` при каждом запуске **перезаписывает** скопированные
+> `prompts/` и `methodology/` версиями из установленного пакета
+> (`_copy_bundle` всегда пишет файлы заново, чтобы после `pipx upgrade`
+> не оставались устаревшие workflow). Не редактируйте их на месте — правки
+> будут потеряны при следующем `setup`. Для локальных изменений скопируйте
+> нужный файл в отдельный путь проекта и ссылайтесь на него оттуда;
+> исходные `prompts/`/`methodology/` в репозитории считайте managed-контентом.
+
 Посмотреть план без реальных изменений:
 
 ```bash
@@ -114,12 +124,12 @@ head -1 "$(which doubled-graph)"  # вывести shebang — это и ест
 
 Сценарий 2. Миграция существующего — `prompts/migrate-existing-project/00-entry-prompt.md`.
 
-Сценарий 3. Обновление версии методики (1.9.x → 2.0) — `prompts/upgrade-methodology/00-entry-prompt.md`.
+Сценарий 3. Обновление версии методики (2.0.x → 2.1) — `prompts/upgrade-methodology/00-entry-prompt.md`.
 
 ### 4. Проверить, что всё встало
 
 ```bash
-doubled-graph --version          # 2.0.2
+doubled-graph --version          # 2.1.0
 doubled-graph status --repo .    # artifacts.project_kind = grace4, grace_engine.upgrade_required = false
 doubled-graph lint --repo .      # summary.errors = 0
 doubled-graph migrate check --repo .   # map_pending / placeholders — что осталось заполнить
@@ -129,6 +139,10 @@ doubled-graph migrate check --repo .   # map_pending / placeholders — что
 Если `upgrade_required` = `true` — движок старее 4.x: `npm install -g @osovv/grace-cli@4.2.1`, затем `setup` повторно.
 Если `lint` вернул `markup.module-map-missing` — это содержательная работа, промпт `prompts/upgrade-methodology/01-content-gaps.md`.
 
+Полная проверка (ruff + unit + real-CGC E2E + diff/version): `bash scripts/verify.sh`; критерии готовности — [docs/VERIFICATION.md](docs/VERIFICATION.md).
+
+В 2.1: `detect_changes` проверяет `<Exports>`/`MODULE_MAP`/`INPUTS` и реальные CALLS/IMPORTS-рёбра, `impact` отдаёт `anchor`, `include_unchanged` расширяет охват, а `DG-Authored`-трейлер ставится через `prepare-commit-msg`. Установить все хуки: `doubled-graph init-hooks --all` (git + Claude Code + provenance); без Claude Code — `--post-commit --prepare-commit-msg`.
+
 
 ---
 
@@ -199,6 +213,8 @@ doubled-graph setup          # 2. обновить движок и подгот
 создаёт недостающие обязательные документы, добавляет `LINKS: M-*` в
 `MODULE_CONTRACT` (без этого файл не привязан к модулю), приводит `MAP_MODE`
 в соответствие с `ROLE`. Legacy `docs/` не изменяются и не удаляются.
+`setup` также дописывает в `AGENTS.md` управляемый блок anti-patterns
+(идемпотентно; существующий текст и BOM сохраняются).
 
 Если миграцию не сделать:
 

doubled_graph-2.1.0/docs/ACCEPTANCE_CERTIFICATE.md

@@ -0,0 +1,251 @@
+# Acceptance certificate — doubled-graph 2.1
+
+Status: **AUDIT CANDIDATE — NOT PASS** (contract v2, 2026-11-08: the owner removed the
+full mutation score from the acceptance criteria; targeted probes remain mandatory. No
+independent audit has run yet under contract v2. Cycle-10 audit FAILED on an incomplete
+manifest (49/122 read by the auditor) and open P2 handoff decisions: 30 manifest files
+are not in the git index and 1293 tracked `.pyc`/`__pycache__` remain. Cycle 11 was a
+**maintenance repair**, not an audit: it fixed the CGC worker leak, the `python -m`
+exit-code loss and the `health` "no side effects" promise.)
+
+## Artifact
+
+| Field | Value |
+|---|---|
+| Tree hash | `7e26796626bfed4939d7f38d3fc9feabdbe2159073f308789d83c5bd7adf9d9e` (src/tests/scripts, AUDIT_PROMPT §1 command; `LC_ALL=ru_RU.UTF-8` and `LC_ALL=C` now produce the same hash on this tree) |
+| HEAD | `b7ed89b31879d7de3132ff66ece7107f5b0bcb09` (working tree intentionally dirty; `git status --short` is part of the hand-off) |
+| Manifest | 124 files / 41,071 lines (AUDIT_PROMPT §2 scope; recomputed after the final edit including this certificate) |
+| Python | `3.12.16` (`venv/bin/python`); gates ran with `HOME=/tmp/fix-home-c11` |
+| CGC | `0.6.13` (pin `>=0.6.13,<0.7`) — FalkorDB 1.2.0 / redislite |
+| Backend | **FalkorDB** — `codegraphcontext.core.database_falkordb.FalkorDBManager`, per-repo DB at `~/.codegraphcontext/repos/<repo>-<hash>/graph` |
+| Date | `2026-11-08` |
+
+The backend claim is now precise about **selection and cross-process behaviour**:
+
+- `cgc.backend="auto"` is resolved **before** CGC runs to one concrete backend — FalkorDB
+  Little on Unix/Python ≥3.12 when `redislite` is importable, KuzuDB on Windows/<3.12 —
+  and that value is pinned through `CGC_RUNTIME_DB_TYPE`. It is no longer "let CGC pick".
+- FalkorDB is constructed directly; CGC's silent fallback to KuzuDB at the same
+  `db_path` is rejected (`CGCUnavailable` names the requested and actual backend). A
+  failed start is an explicit, actionable `CGC_BACKEND_FAIL`, never a silent `NONE`/HIGH.
+- Lock-contention errors (`Could not set lock on file`, `docs.kuzudb.com/concurrency`)
+  are retried at most 3 times with bounded backoff; exhaustion raises the original error
+  wrapped as `CGCUnavailable` (no extra processes are forked).
+- Cross-process handoff is exercised for real: the E2E test releases the per-repo DB
+  (`CGC.close()` → persisted shutdown of the FalkorDB worker) before running
+  `python -m doubled_graph impact` in a subprocess, then reconnects. The same backend and
+  DB path are asserted (not a mock/Kuzu). If the host cannot start FalkorDB at all, the
+  failure is explicit (`CGC_BACKEND_FAIL`), which is a visible backend failure, not a
+  silent fallback.
+
+## Gate
+
+Measured on the frozen artifact (`7e2679…`, all commands run after the final
+source/test/script edit, sequentially, `HOME=/tmp/fix-home-c11`, Python 3.12.16):
+
+- `ruff check src tests scripts`: **PASS** (`All checks passed!`).
+- Unit/contract: **657 passed, 2 skipped** — the only skips are exactly the two allowed
+  engine-level ones (`DOUBLED_GRAPH_ENGINE_TS`/tsx), and `tests/conftest.py` enforces a
+  **strict allowlist** (exact node id + exact reason): any third skip fails the session.
+- Real-CGC E2E: **14 passed, 0 skipped** on this artifact (separate
+  `pytest tests/test_e2e_acceptance.py` run). The 5/5 stability evidence quoted by
+  cycle 8 belongs to an older artifact and is **not transferred**.
+- Branch coverage (six critical modules): **100.00%** — `TOTAL 1891 statements, 734
+  branches, 0 partial`; `--cov-fail-under=100` reached.
+- Targeted mutation probes: **6/6 killed** (`scripts/mutation_probes.py`, timeout +
+  sha256-restore verified).
+- Process-leak check (cycle-11 P1): before the gate series `pgrep -f falkor_worker` = 1
+  and `pgrep -f redis-server` = 1 (both foreign: a `WebFront` repo worker from the
+  pipx 3.14 install, not this workspace); after the full suite, the separate E2E run,
+  coverage, probes and a manual `analyze` → `health` → `detect-changes` CLI series they
+  are back to exactly 1/1, and `pgrep -f 'Grace/venv'` is 0. Transient foreign
+  `falkor_worker` processes (pipx 3.14, other workspaces) appeared and exited
+  independently and were not touched.
+- Full mutation: **out of acceptance scope (owner decision 2026-11-08, contract v2).**
+  It is reported for information only and is **not** a PASS criterion. Last complete run
+  on `1793bb33…`: `3740 killed / 590 survived / 2 timeout / 192 suspicious / 4524 total`
+  (**82.67%**, threshold 100%), exit code `1`, duration `6063.9 s`,
+  `--max-children 1 --retries 0`; not transferred to `7e2679…`. Evidence:
+  `mutation-report.json`, `mutation-report.md`, `mutation-run.log`,
+  `mutants/mutmut-cicd-stats.json`. The full run remains a release backlog, not an
+  acceptance gate.
+- `scripts/verify.sh`: **PASS** (exit 0) — ruff + 657 passed/2 skipped + coverage 100% +
+  probes 6/6 + `git diff --check`/`--cached` + `doubled-graph 2.1.0`.
+- `git diff --check` + `--cached`: clean.
+
+## Audits
+
+| Audit | Scope | P0/P1/P2/P3 | Status |
+|---|---|---|---|
+| Independent full audit — cycle 1 | 109 files / 31,999 lines | 0 / 2 / 9 / 12 | **FAILED** (all findings addressed; see `.tmp/fix/cycle-01/RESOLUTION.md`) |
+| Independent full audit — cycle 2 | 109 files | 0 / 2 / 4 / 7 | **FAILED** (all findings addressed in code/tests/docs; see `.tmp/fix/cycle-02/RESOLUTION.md`) |
+| Independent full audit — cycle 3 | 118 files / 36,620 lines | 0 / 1 / 5 / 9 | **FAILED** (all findings addressed; see `.tmp/fix/cycle-03/RESOLUTION.md`) |
+| Independent full audit — cycle 4 | 119 files / 37,254 lines | 0 / 2 / 4 / 4 | **FAILED** — P1-1 + P2-1…P2-4 + P3-1…P3-4 addressed; P1-2 (full mutation score) remains open; see `.tmp/fix/cycle-04/RESOLUTION.md` |
+| Independent full audit — cycle 5 | 78/119 files read (auditor process failure, not an author claim) | 0 / 1 / 0 / 12 | **FAILED** — P1-1 and P3-1…P3-12 addressed in code/tests/docs; full mutation score still not re-measured; see `.tmp/fix/cycle-05/RESOLUTION.md` |
+| Independent full audit — cycle 6 | 51/120 files read fully (auditor process failure, not an author claim) | 0 / 2 / 2 / 4 | **FAILED** — P1-1, P1-2, P2-1, P2-2, P3-1…P3-4 addressed in code/tests/docs; full mutation score still not re-measured; see `.tmp/fix/cycle-06/RESOLUTION.md` |
+| Independent full audit — cycle 7 | 120/120 files read (manifest complete) | 0 / 2 / 2 / 10 | **FAILED** — P1-1, P1-2, P2-1, P2-2 and P3-1…P3-10 addressed in code/tests/docs; full mutation score still not re-measured on the new artifact; see `.tmp/fix/cycle-07/RESOLUTION.md` |
+| Independent full audit — cycle 8 | 121/121 files read (manifest complete) | 0 / 3 / 5 / 6 | **FAILED** — P1-1, P1-2, P2-1…P2-5 and P3-1…P3-6 addressed in code/tests/docs; full mutation score still not re-measured on the cycle-8 artifact; see `.tmp/fix/cycle-08/RESOLUTION.md` |
+| Independent full audit — cycle 9 | 53/122 files read (auditor process failure, not an author claim; the rest is N/R) | 0 / 0 / 5 / 4 | **FAILED** — P1 remains unconfirmed; P2-3, P2-5 and P3-1…P3-4 fixed in code/tests/docs; P2-1 (full mutation), P2-2 (tracked `.pyc`) and P2-4 (CI `gate` vs full mutation) remain open **user decisions**; see `.tmp/fix/cycle-09/RESOLUTION.md` |
+| Independent full audit — cycle 10 | 49/122 files read (auditor process failure, not an author claim; the rest is N/R) | 0 / 0 / 5 / 6 | **FAILED** — no P1 in the read scope; FAILED on the contract's incomplete-manifest rule plus P2 handoff items (mutation out of scope only after this cycle; 28 untracked manifest files; 1293 tracked `.pyc`). After it, the owner removed full mutation from the acceptance criteria (contract v2, 2026-11-08). |
+
+Cycle-8 findings closed in this artifact:
+
+- **P1-1** the CGC backend is deterministic across processes and never silently
+  swaps engines. `cgc.backend="auto"` resolves to FalkorDB (Unix/3.12+, `redislite`
+  present) or KuzuDB otherwise and is pinned via `CGC_RUNTIME_DB_TYPE`; FalkorDB is
+  constructed directly and a CGC-selected fallback is rejected with `CGCUnavailable`.
+  Lock errors are retried 3× with backoff, then surface explicitly. The post-commit
+  hook runs `analyze --strict` (no `--silent`) and writes the failure JSON plus an
+  explicit `post-commit analyze FAILED` line to `post-commit.log` while still exiting 0
+  (a post-commit hook must not break an already-created commit). The cross-process E2E
+  test releases the DB (`CGC.close()`) before the CLI subprocess, so it no longer races
+  for the same socket/dump. Tests: `test_cgc_backend.py` (pin, mismatch rejection,
+  bounded retry, exhaustion, `close`), updated `test_cgc_contract.py` auto/backend tests,
+  `test_e2e_impact_isolated_class_is_none_not_query_fail` (5/5), `test_e2e_workspace_path_real_backend`.
+- **P1-2** `drift_md.open_entries` is parsed from real markdown: `## D-*` and `### D-*`
+  entries, open/closed sections, recorded statuses (explicit status wins over section,
+  negations like "not resolved" count as open), and a preamble is not an entry. Tests:
+  `test_health.py::test_drift_md_*` (open `###` under an open section → 1/false-green
+  fixed; intro + closed `## D-*` → 0/false-red fixed; valid empty → 0; recorded status;
+  negation).
+- **P2-1** the certificate's backend claim now names the deterministic pin, the rejected
+  fallback, the bounded lock retry and the cross-process handoff (this section).
+- **P2-2** “E2E 14 passed” was backed by 5/5 consecutive full-suite runs on the previous
+  artifact `923dc61c…`; that count is **not transferred** to `0d039dce…`/`7e2679…`, where
+  the freeze ran the E2E suite once (`14 passed, 0 skipped`, also reported above).
+- **P2-3** `docs/VERIFICATION.md` keeps the isolated-class invariant `done` because the
+  E2E branch was made deterministic (DB released before the subprocess) and was repeated
+  5/5 on the previous artifact; the cycle-9 changes do not touch that path, and the
+  current artifact’s single E2E run is reported honestly above rather than as 5/5.
+- **P2-4** `docs/TOOLS.md §7.2` describes the real DRIFT.md parser and the
+  `open_entries > 0 → ok:false/exit 1` contract.
+- **P2-5** the literal `§4.4` deviation (merge `gate` runs coverage 100 + six targeted
+  probes; full mutation is a separate release job) stays documented in
+  `docs/VERIFICATION.md` and the workflow comment; no false PASS or transferred score.
+- **P3-1** stale `_relationship_fallback` docstring (“edge walk stays
+  Function-to-Function”) updated to the label-agnostic walk.
+- **P3-2** `docs/SPEC.md §6.4` shows the real `lint(self, verbose: bool = False)`.
+- **P3-3** `docs/SPEC.md §5` tree now lists all source modules (`health.py`,
+  `hooks_installer.py`, `policy/agents_md.py`, `tools/_meta.py`, …) and all **23** test
+  modules (including `test_cgc_backend.py`) plus `conftest.py`. Cycle 9 found that the
+  cycle-8 wording still said “22” and omitted `scripts/`, `methodology/`, `prompts/`,
+  `for-*`, `docs/AUDIT_PROMPT.md` and this certificate; the tree is complete now, and
+  the earlier “all cycle-8 findings addressed” claim is corrected in this section.
+- **P3-4** `.gitignore` keeps new bytecode out (`__pycache__/`, `*.py[cod]`) and a
+  hygiene test pins those rules. 1293 `.pyc`/`__pycache__` files are still **tracked**;
+  untracking them is a **user decision** (`git rm --cached` was explicitly not run).
+- **P3-5** `impact --depth` is validated to the documented/MCP `1–10` range in the CLI
+  (`argparse` type), with a regression test.
+- **P3-6** `docs/TOOLS.md` gives the same complete `meta` (including
+  `index_freshness_s`) in §1.2, §2.2, §3.2 and §4.2.
+
+## Cycle-9 corrections
+
+The cycle-9 audit was itself incomplete (53/122 files read; the auditor's process
+failure — the remaining files are N/R, not “OK”), returned **FAILED** with P0=0,
+**P1 not confirmed**, P2=5 and P3=4. P1 is therefore neither accepted nor refuted by
+that cycle; this certificate does not claim P1=0 on its basis.
+
+Fixed in this artifact:
+
+- **P2-5** `analyze_incremental` updated `imports_map` before `add_file_to_graph` and
+  never rolled it back on failure, so later files and `link_function_calls` resolved
+  imports of a file that was not indexed. It now snapshots the map before parse and
+  restores it on parse/skip/delete/write failure, before the rollback re-add and before
+  the next file/link phase. Tests: `tests/test_cgc_contract.py`
+  `::test_incremental_failed_file_imports_do_not_leak_into_later_files` (mixed
+  success/failure + repeated incremental) and
+  `::test_incremental_rollback_restores_imports_map_for_committed_file`.
+- **P2-3** `docs/SPEC.md §5` tree/counters corrected (23 test modules, `scripts/`,
+  `methodology/`, `prompts/`, `for-*`, audit prompt and certificate); the cycle-8
+  P3-3 “22 test modules” claim was incomplete and is corrected above.
+- **P3-1** `server.py` docstring now names the pinned `mcp` 1.30.x line
+  (`mcp>=1.30,<2`), not 0.9.x; the missing-SDK error message says `>=1.30,<2`.
+- **P3-2** `ensure_gitignore_entries` compares normalized whole lines
+  (`_gitignore_patterns`) instead of substrings, so
+  `.doubled-graph/logs/archive/` or a comment cannot satisfy
+  `.doubled-graph/logs/`. Test:
+  `tests/test_mutation_kill_hooks.py::test_ensure_gitignore_requires_exact_pattern_not_substring`.
+- **P3-3** `_extract_contract` requires a left token boundary for the marker keyword
+  and the symbol name (`MYCONTRACT: f` no longer matches `CONTRACT: f`; a name suffix
+  is not accepted), while `START_CONTRACT`/`START CONTRACT` still work. Test:
+  `tests/test_mutation_kill_context.py::test_extract_contract_requires_keyword_and_name_boundaries`.
+- **P3-4** `KùzuDB` → `KuzuDB` everywhere it occurred (`docs/SPEC.md` §4 diagram and
+  §11, `methodology/tools.md`).
+
+Open, not fixed and **not silently accepted** (user decisions):
+
+- **P2-1** full mutation is out of acceptance scope under contract v2 (owner decision
+  2026-11-08); the last measured score is reported for information only.
+- **P2-2** 1293 tracked `.pyc`/`__pycache__` files; untracking requires
+  `git rm --cached`, explicitly a user decision and not executed here.
+- **P2-4** the merge `gate` runs coverage 100 + six targeted probes, matching the
+  contract-v2 §4.4 wording; full mutation remains a separate release/backlog job.
+
+## Cycle-11 corrections (maintenance repair, not an audit)
+
+Three user-reported defects were fixed in the working tree (no commit):
+
+- **P1-A — CGC worker leak / process-table exhaustion.** `CGC` now owns a lifecycle:
+  `atexit.register(self.close)` per instance (re-armed by `_ensure` after a `close()`),
+  a class-level `weakref.WeakSet` registry, `CGC.close_all()` (idempotent, swallows
+  per-instance close errors), `__enter__`/`__exit__`, and a `close()` that always sets
+  `_db=None`, unregisters the exit hook and **never raises** (last error in
+  `last_close_error`). `analyze`, `impact`, `context`, `detect_changes` and `health`
+  release the instance in `finally`; `tests/conftest.py` runs `CGC.close_all()` after
+  every test (autouse) and on `pytest_sessionfinish`. This is the root cause of the
+  ~980 orphaned `redis-server`/`falkor_worker` processes that produced
+  `spawn sandbox-exec EAGAIN`.
+  - Closing the DB after every incremental analyze exposed an upstream CGC/FalkorDB
+    index bug: once the `(name, path, line_number, occurrence_index)` range index is
+    populated, predicates on `occurrence_index=0` return no rows, so the writer's
+    `File-[:CONTAINS]->symbol` link silently finds nothing on a reopened graph.
+    `CGC._relink_file_symbols()` re-links by the correctly indexed
+    `(name, path, line_number)` prefix after every incremental add/rollback; this also
+    fixes cross-process incremental (hook after CLI full), which was previously broken.
+    Best-effort: failures are reported as `relink …` incremental warnings.
+  - Tests: `tests/test_cgc_lifecycle.py` (registry/atexit/close/close_all/`with`,
+    per-tool `close()` spies, relink unit tests). The absolute worker count is verified
+    by the CLI series recorded above, not by a flaky `pgrep` test (the strict skip
+    allowlist forbids a new platform skip, and the host already runs foreign workers).
+- **P1-B — `python -m doubled_graph` lost every exit code.** `__main__.py` now does
+  `raise SystemExit(main(argv_default=["serve"]))`. Regression tests run the real
+  module in a subprocess: `--version` → 0, `init-hooks --post-commit` outside git → ≠0,
+  malformed `phase set` → 2 (`PHASE_INVALID`), `detect-changes --scope all
+  --fail-on-drift` on a real tmp repo with drift → 1, plus the installed console script
+  (`doubled-graph`) rc 0/1. `tests/test_cli_entrypoint.py`.
+- **P2-C — `health` claimed "Read-only, no side effects".** It never changes sources or
+  `.grace`, but `detect_changes` writes `.doubled-graph/cache/` and the CGC parser probe
+  initialises the per-repo graph DB under `~/.codegraphcontext/`. The MCP description,
+  the `health.py` docstring, the CLI docstring and `docs/TOOLS.md §7.2` now say
+  "read-only for the project" and name the local caches/DBs;
+  `tests/test_health.py::test_health_description_is_read_only_for_project_not_side_effect_free`
+  fails if the unconditional promise returns.
+
+The normative rule is now in `docs/VERIFICATION.md` ("Правило: CGC всегда
+закрывается") and in the `CGC` class docstring: any code that constructs `CGC` must
+close it (`with`/`close`/`close_all`), otherwise it leaks a process.
+
+## Known gaps at candidate time
+
+| Gap | Class | Evidence |
+|---|---|---|
+| full mutation score | **out of acceptance scope** (owner decision 2026-11-08) | Removed from the contract's PASS criteria; targeted probes (6/6) remain mandatory. Last measured score (`1793bb33…`): 82.67%, exit 1 — release backlog, not an acceptance blocker. |
+| artifact not committed / 30 manifest files untracked | handoff, **user decision** | HEAD `b7ed89b` predates this work; `scripts/`, `docs/ACCEPTANCE_CERTIFICATE.md`, `docs/AUDIT_PROMPT.md`, `docs/VERIFICATION.md`, `tests/conftest.py`, `tests/test_cli_entrypoint.py`, `tests/test_cgc_lifecycle.py`, `tests/test_e2e_acceptance.py`, `tests/test_health.py`, `health.py`, `provenance.py`, `post_tool.py`, `anchors.py`, `agents_md.py` and the other untracked tests are outside the index. A clean clone of HEAD cannot run `scripts/verify.sh`. The user forbade `git add`/commit, so the artifact must be committed by the owner. |
+| tracked `.pyc` / `__pycache__` backlog | repository hygiene, **user decision** | 1293 files are still in the index. `.gitignore` prevents new ones and a test pins the rules; `git rm --cached` is forbidden for the automated repair and must be decided by the user. |
+| `arg_types` positional semantics | external CGC | CGC does not expose a positional contract; equal-length names/types comparison stays conservative |
+| `ARTICLE.md` / `extra/*` | out-of-scope | archive/article |
+
+No known gap is labelled `external` unless the limitation is outside this codebase.
+The cycle-3 `internal/partial` gaps (untracked-file rollback, single-file export parity)
+remain **closed**; their regression tests are named in `docs/VERIFICATION.md`.
+
+## Signature
+
+NOT SIGNED — no independent audit has run under contract v2 yet; the cycle-10 audit
+FAILED on an incomplete manifest and open handoff items. Cycle 11 was a maintenance
+repair (CGC worker leak, `python -m` exit codes, `health` side-effect claim), not an
+independent audit, so it does not change the verdict. The full mutation score is out
+of acceptance scope per the owner's 2026-11-08 decision; targeted probes remain green.
+The artifact is uncommitted and 1293 tracked `.pyc` files remain; both are user
+decisions. This file is an audit candidate, not a PASS.