ltcai 11.0.1 → 11.2.0

package/README.md

@@ -11,7 +11,7 @@
 [![CI Status](https://github.com/TaeSooPark-PTS/LatticeAI/actions/workflows/ci.yml/badge.svg?branch=main)](https://github.com/TaeSooPark-PTS/LatticeAI/actions/workflows/ci.yml)
 [![License](https://img.shields.io/badge/license-MIT-green)](LICENSE)
 
-![v11.0.1 Living Brain walkthrough](output/release/v11.0.1/gifs/v11.0.1-living-brain-walkthrough.gif)
+![v11.2.0 Living Brain walkthrough](output/release/v11.2.0/gifs/v11.2.0-living-brain-walkthrough.gif)
 
 Chat, files, folders, notes, and web pages all flow into one durable knowledge
 graph on your computer. Any model — local MLX or cloud — can speak with that
@@ -24,10 +24,10 @@ memory. Nothing leaves your machine without explicit consent.
 
 | | |
 | --- | --- |
-| **Chat with a Brain that remembers** — every conversation grows durable, source-linked memory ![Brain Chat](output/release/v11.0.1/screenshots/04-brain-chat-home.png) | **See how knowledge connects** — a real relationship graph, not a file list ![Memory Graph](output/release/v11.0.1/screenshots/05-memory-graph.png) |
-| **Capture anything** — files, whole folders, notes, screenshots, web pages ![Capture](output/release/v11.0.1/screenshots/06-capture.png) | **Automate with review** — agent changes become proposals you approve first ![Review Center](output/release/v11.0.1/screenshots/12-review-center.png) |
-| **Pick a model in one click** — recommended local models for your hardware ![Recommended Models](output/release/v11.0.1/screenshots/02-recommended-models.png) | **Stay in control** — audit, roles, retention in a separate admin surface ![Admin Console](output/release/v11.0.1/screenshots/10-admin-console.png) |
-| **Watch a file become memory** — three named steps, not a pipeline diagram ![Material to memory](output/release/v11.0.1/screenshots/11-knowledge-journey.png) | **Say how much it may do alone** — one dial in plain words; dangerous actions stay blocked either way ![Settings](output/release/v11.0.1/screenshots/08-system.png) |
+| **Chat with a Brain that remembers** — every conversation grows durable, source-linked memory ![Brain Chat](output/release/v11.2.0/screenshots/04-brain-chat-home.png) | **See how knowledge connects** — a real relationship graph, not a file list ![Memory Graph](output/release/v11.2.0/screenshots/05-memory-graph.png) |
+| **Capture anything** — files, whole folders, notes, screenshots, web pages ![Capture](output/release/v11.2.0/screenshots/06-capture.png) | **Automate with review** — agent changes become proposals you approve first ![Review Center](output/release/v11.2.0/screenshots/12-review-center.png) |
+| **Pick a model in one click** — recommended local models for your hardware ![Recommended Models](output/release/v11.2.0/screenshots/02-recommended-models.png) | **Stay in control** — audit, roles, retention in a separate admin surface ![Admin Console](output/release/v11.2.0/screenshots/10-admin-console.png) |
+| **Watch a file become memory** — three named steps, not a pipeline diagram ![Material to memory](output/release/v11.2.0/screenshots/11-knowledge-journey.png) | **Say how much it may do alone** — one dial in plain words; dangerous actions stay blocked either way ![Settings](output/release/v11.2.0/screenshots/08-system.png) |
 
 ## Why Lattice AI
 
@@ -58,53 +58,63 @@ First-run flow — wake the Brain, pick the owner, load a recommended model:
 
 | | | |
 | --- | --- | --- |
-| ![Login](output/release/v11.0.1/screenshots/01-login.png) | ![Model install](output/release/v11.0.1/screenshots/03-install-load-progress.png) | ![Model library](output/release/v11.0.1/screenshots/07-model-library.png) |
+| ![Login](output/release/v11.2.0/screenshots/01-login.png) | ![Model install](output/release/v11.2.0/screenshots/03-install-load-progress.png) | ![Model library](output/release/v11.2.0/screenshots/07-model-library.png) |
 
 Screenshot index and capture notes:
-[output/release/v11.0.1/SCREENSHOT_INDEX.md](output/release/v11.0.1/SCREENSHOT_INDEX.md)
+[output/release/v11.2.0/SCREENSHOT_INDEX.md](output/release/v11.2.0/SCREENSHOT_INDEX.md)
 
 ## Current Release
 
-The current release is **11.0.1 — Both Branches**:
-
-11.0.0 put every line under test and documented, unfixed, the defects that
-work surfaced. 11.0.1 is the settling of that account: **all eleven
-documented defects are fixed, the code they proved dead is gone, and the CI
-floor now holds 100% of branch arcs as well as lines.**
-
-- **Every 11.0.0 finding is fixed.** The Telegram helper that sent the local
-  server's auth header to api.telegram.org, the unload-all that fabricated
-  success, the HTML inspector whose stylesheet collection could never fire,
-  the review-item id that collided within one second, the vLLM zombie that
-  read as "already running", the security dashboard's listing/detail
-  redaction gaps and invalid-JSON exports, the embedding `model_id` frozen at
-  the wrong dimension, the fast-path model-name mismatch, and the workspace
-  routes' dead literal and unreachable 404 arms. Each fix flipped the test
-  that had pinned the broken behaviour and gained a regression test.
-- **Branch coverage joined the floor.** `branch = true` is now in the
-  coverage config: both directions of all 9,828 conditionals execute under
-  the suite (5,798 tests, up from 5,426), and `fail_under = 100` fails CI on
-  a single missed arc. Exclusions stay honest: eight reasoned
-  `pragma: no cover` lines and two reasoned `pragma: no branch` lines.
-- **Dead code left the tree.** The unused auto-approved write-policy factory,
-  an unused snapshot-import normalizer, the workspace 404 arms shadowed by
-  the anti-enumeration 403, a provably-dead version guard, a provably-dead
-  MLX condition, and the vLLM silent-success recheck — each removal proved by
-  AST/reference scan before deletion.
-- **Verified where CI runs.** The full suite passes with the branch gate on
-  macOS 3.14, a fresh-resolve python 3.11 environment (fastapi 0.141), and a
-  clean linux python:3.14 container — the three environments that caught
-  11.0.0's release-day failures.
+The current release is **11.2.0 — All Systems On**:
+
+11.1.0 built the intelligence layer; 11.2.0 makes sure **every part of it
+actually works, ships today's models, and puts the switches in your hand.**
+
+- **The model catalog is current and verified.** Every recommended model
+  was checked against the live Hugging Face API — no weights downloaded, no
+  model loaded, nothing run on your machine. Two models that had vanished
+  from the Hub (phi-3.5-vision, moondream2), four gated repos and three
+  obsolete generations left the catalog; the ten recommendations are all
+  2025–2026 builds (Gemma 4 line, Qwen3.5/3.6, gpt-oss-20b, LFM2.5 with
+  Korean support), RAM-tiered, with exact-case ids, real byte sizes (one
+  entry claimed 11.8 GB for a 61 GB repo) and a static-loadability verdict
+  whose limits are stated, not hidden. Already-downloaded older models stay
+  recognised so nobody's local weights become orphans.
+- **Every switch is yours now.** A new 기능 drawer on the Brain home's dock
+  lists every opt-in capability — multimodal memories, video, brain
+  network sharing, vault watch, photo semantic search, RRF fusion, graph
+  expansion, automatic synthesis, background indexing, and the vector
+  backend choice — each with a plain-language line, its current source
+  (default / env / you), and a live switch. Toggles apply at runtime
+  through injected gate seams; an uninstalled backend says "install
+  needed" instead of pretending.
+- **The scope-outs are in.** Notion-export, Git-history and mail/calendar
+  bridges (all through the single ingestion gate with approval and
+  dry-run), recipient public-key encryption for subgraph shares (X25519
+  sealed box; passphrase mode stays), video ingestion via guarded ffmpeg
+  keyframes with subtitle companions, vault watch mode, bulk review
+  approval, the Self-Model summary reaching the agent loop, and text→image
+  semantic photo search.
+- **A 58-row evidence audit** verified every documented feature against
+  the running app: 51 worked as documented; the rest were fixed (a
+  permanently blank briefing health section, cloud memory write-back that
+  never reached the Review Center, an empty Brain scoring 100/excellent)
+  or the docs were lowered to the truth. The audit ships as
+  [docs/FEATURE_AUDIT_v11.2.0.md](docs/FEATURE_AUDIT_v11.2.0.md).
+
+All of it lands with the floor intact: **6,490 tests, 100.00% of 39,054
+statements and 11,014 branches**, verified on macOS 3.14, a fresh-resolve
+python 3.11 environment, and a clean linux python:3.14 container.
 
 Release notes: [RELEASE.md](RELEASE.md) · Full history: [docs/CHANGELOG.md](docs/CHANGELOG.md)
 
-Expected artifacts for 11.0.1 release must use exact filenames:
+Expected artifacts for 11.2.0 release must use exact filenames:
 
-- `dist/ltcai-11.0.1-py3-none-any.whl`
-- `dist/ltcai-11.0.1.tar.gz`
-- `ltcai-11.0.1.tgz`
-- `dist/ltcai-11.0.1.vsix`
-- `src-tauri/target/release/bundle/dmg/Lattice AI_11.0.1_aarch64.dmg`
+- `dist/ltcai-11.2.0-py3-none-any.whl`
+- `dist/ltcai-11.2.0.tar.gz`
+- `ltcai-11.2.0.tgz`
+- `dist/ltcai-11.2.0.vsix`
+- `src-tauri/target/release/bundle/dmg/Lattice AI_11.2.0_aarch64.dmg`
 
 Do not use wildcard artifact uploads. Package registry publishing remains owner-run.
 
@@ -139,6 +149,8 @@ See [ARCHITECTURE.md](ARCHITECTURE.md) for details and
 
 | Version | Theme |
 | --- | --- |
+| 11.2.0 | All Systems On |
+| 11.1.0 | Product Intelligence |
 | 11.0.1 | Both Branches |
 | 11.0.0 | Full Measure |
 | 10.10.0 | Quiet Station |

package/docs/CHANGELOG.md

@@ -4,6 +4,67 @@ The top entry is either the current unreleased main-branch work or the current
 release line. Older entries are historical and may describe behavior as it
 existed at that release.
 
+## [11.2.0] - 2026-08-11 — All Systems On
+
+### Changed
+- 모델 카탈로그 전수 최신화: HF API 무부하 검증(가중치 다운로드 0, 실로드
+  0) 후 추천 10종을 2025–2026 세대(Gemma 4 · Qwen3.5/3.6 · gpt-oss-20b ·
+  LFM2.5)로 재구성. Hub 소멸 2종·gated 3종·vllm 전용·구세대 제거(완전
+  삭제), 기존 다운로드 모델은 인식 전용 목록으로 보호. 검증기에서
+  `--deep`/`--test-load` 제거(무부하 원칙 구조화). 크기 오기(11.8→61.1GB)
+  등 부수 결함 8건 수정.
+- 빈 Brain이 가용 1차원만으로 100/excellent를 받던 건강 채점을 정직화
+  (측정 불가 차원은 사유와 함께 unavailable, 전무 시 등급 없음).
+
+### Added
+- Brain 홈 dock **기능 서랍**: opt-in 10종(멀티모달·비디오·브레인 네트워크
+  ·볼트 감시·사진 의미 검색·RRF·이웃 확장·자동 합성·배경 인덱싱·벡터
+  백엔드)의 서버 카탈로그 라이브 토글 — `GET/POST /api/features`,
+  FeatureGate 시임으로 재시작 없이 반영, 사용자>env>기본.
+- Interop 브릿지: Notion export · Git 히스토리 · 메일(.eml)/캘린더(.ics)
+  — 단일 인제스트 게이트+승인+dry_run(`/api/ingestion/interop`).
+- 서브그래프 공유 수신자 공개키 암호화(X25519 sealed box, ephemeral 키)
+  + `GET /api/knowledge-graph/share/recipient-key`.
+- 비디오 인제스트(ffmpeg 가드 키프레임 + .srt/.vtt, NodeType.VIDEO),
+  볼트 감시 모드, 일괄 리뷰 승인/기각, Self-Model 요약의 에이전트 루프
+  도달, 텍스트→이미지 의미 검색(`/api/search/image-query`).
+- 58행 증거 기반 기능 감사 문서 docs/FEATURE_AUDIT_v11.2.0.md.
+
+### Fixed
+- 오늘의 브리핑 건강 섹션 영구 공백(스키마 불일치), 클라우드 유래 지식이
+  리뷰 센터에 도달하지 않던 write-back 미배선(+auto_commit 정책 연결),
+  vector-freshness 세분화 미노출, 멀티모달 context_quality 신호 미배선,
+  `kgv2_edges`가 canonical 엣지 타입을 ''로 읽던 뷰(`COALESCE(NULLIF)`),
+  죽은 자동화 저신뢰 게이트, 사어 grok id(→grok-4.5).
+
+## [11.1.0] - 2026-08-10 — Product Intelligence Layer
+
+### Added
+- 플러그형 벡터 인덱스 레이어(`lattice_brain/graph/vector_index/`):
+  BruteForce 기본, int8 Quantized·HNSW(`ltcai[hnsw]`) 옵트인, 영속 배경
+  임베딩 큐(`vector_jobs`), `vector_freshness_breakdown()`, RRF 융합·이웃
+  후보 확장 옵션. 하이브리드 p50 10k 299ms → 10.1ms, 50k 43.9ms
+  (recall@10 0.987, docs/PERFORMANCE.md).
+- Temporal 지식 모델: `valid_from`/`valid_to`/`superseded_by`(멱등 제자리
+  승급, NULL 규약) + `as_of(timestamp)` 슬라이스. 모순 감지 → 리뷰 제안
+  → 승인 시 temporal 스탬프. 이벤트 기반 합성(상위 개념·누락 엣지·
+  proactive Brief 제안, 25개 인제스트마다, 격리 배선) + 중요도/정리 제안.
+  새 API: `/api/brain/proactive-brief|importance|synthesize|contradictions/*`.
+- 멀티모달 1등 시민(`allow_multimodal`, 기본 꺼짐 — 꺼짐=바이트 동일):
+  Image/Audio 1등 노드, OCR·실캡션만(`caption_status`), 별도 이미지 벡터
+  공간+late fusion, Evidence 인라인 썸네일(승인 게이트 비우회), 비디오는
+  정직 거부.
+- Self-Model 서브그래프(제안-우선 생성, 사용자 직접 소유), 컨텍스트
+  예산 주입, 삭제 불가능 구조의 폴더 재구성 제안,
+  `/api/memory/self-model*` 5종.
+- Obsidian vault 브릿지(`POST /api/ingestion/obsidian`, 승인 게이트,
+  위키링크→엣지, 멱등) + 서명·암호화 선택적 서브그래프 공유 프로토타입
+  (`LATTICEAI_BRAIN_NETWORK` 기본 꺼짐, 수신=리뷰 제안).
+
+### Removed
+- `VisionStub`/`get_vision_embedder` — 파일명으로 캡션을 합성하고 그
+  문자열의 해시를 이미지 임베딩으로 저장하던 경로(정직성 위반) 삭제.
+
 ## [11.0.1] - 2026-08-10
 
 ### Fixed

package/docs/COMMUNITY_AND_PLUGINS.md

@@ -1,6 +1,6 @@
 # Community And Plugins
 
-Current release: **11.0.1 — Both Branches**.
+Current release: **11.2.0 — All Systems On**.
 
 LatticeAI defines the path from a strong local-first framework (8.4.0
 action-aware baseline, 8.5.0 registry+DI hardening, 8.6.0 capture/navigation

package/docs/DEVELOPMENT.md

@@ -3,7 +3,7 @@
 > **Status: canonical** — current contributor guidance, kept in sync with the
 > current release.
 
-Current release: **11.0.1 — Both Branches**.
+Current release: **11.2.0 — All Systems On**.
 
 This document is for contributors working on the local-first Digital Brain
 codebase. Product positioning and quick start stay in `README.md`; release

package/docs/FEATURE_AUDIT_v11.2.0.md

@@ -0,0 +1,393 @@
+# Feature Audit — v11.2.0
+
+> **Status: reference** — a point-in-time, evidence-based sweep of every row in
+> [`FEATURE_STATUS.md`](../FEATURE_STATUS.md). It records what was true when it
+> was run, not what the current release claims; the canonical feature state
+> stays in `FEATURE_STATUS.md`.
+
+Audited against the `release/v11.2.0` working tree on 2026-08-10, starting from
+the 11.1.0 feature table. Continues the evidence-based audit tradition started
+in 3.3.0: **every row gets a verdict and a way to reproduce it**, and a row
+whose sentence outruns its code is lowered rather than defended.
+
+## Method
+
+Claims were checked by running the product, not by reading about it.
+
+* **Hermetic full-app boot.** `latticeai.app_factory.build_runtime()` inside
+  `scripts/export_openapi.py`'s `isolated_runtime_environment`, which points
+  `HOME`, `LATTICEAI_DATA_DIR`, `LATTICEAI_BRAIN_DIR`, the agent workspace, the
+  vault dir and the static dir at a throwaway temp tree. No real HOME, no real
+  network, no live server. Endpoints were exercised through `TestClient`.
+* **Three passes.** (1) anonymous read surface, (2) corrected paths with
+  redirect-following off plus API introspection, (3) a registered-and-signed-in
+  session driving the local-file approval dance end to end.
+* **Route inventory** from the app's own route table and its generated OpenAPI
+  schema, cross-checked against the checked-in `frontend/openapi.json` that
+  `scripts/check_openapi_drift.mjs` already proves is regenerated from the live
+  app.
+* **Code-level verification** for claims an endpoint cannot show (constants,
+  defaults, guard order, which collaborator writes which key).
+
+Where a claim is pinned by an existing regression test, that test is the
+evidence: it is cheaper to keep true than a paragraph.
+
+## Cross-cutting smoke
+
+Measured on the hermetic boot described above.
+
+| Measure | Value |
+| --- | --- |
+| Routers registered (`include_router` calls, all distinct) | 35 |
+| Total routes on the app | 451 |
+| Of those, API routes | 446 |
+| Non-API routes | 4 (`/openapi.json`, `/docs`, `/docs/oauth2-redirect`, `/redoc`) |
+| Static mounts | `/static` always; `/icons` only when `static/icons/` exists (it does in the repo; it did not in the sandboxed static dir) |
+| OpenAPI paths | 404 |
+| OpenAPI operations | 445 |
+| `frontend/openapi.json` drift | clean (`node scripts/check_openapi_drift.mjs`) |
+| Health snapshot | `status: ok`, `version 11.1.0`, `mode local`, `graph_enabled true`, `telegram_enabled false`, `autoload_models false` |
+| Probes run | 70 (pass 1) + 42 (pass 2) + the signed-in pass; **zero 5xx, zero unhandled exceptions** |
+
+Every non-200 in the sweep was a *designed* refusal, and each named its
+reason: `403` on the share routes while `LATTICEAI_BRAIN_NETWORK` is off,
+`401` on local-file routes without a session, `405` on a POST-only path,
+`422` on a malformed body.
+
+### Frontend ↔ real routes
+
+320 server-path literals were extracted from `frontend/src`. **Every one
+resolves to a real route.** Ten looked missing on a first pass and all ten are
+template routes the client builds with an interpolated id (`/models/switch` →
+`/models/switch/{model_id:path}`, `/mcp/connectors` → `/mcp/connectors/{mcp_id}`,
+and so on); the eleventh, `/workspace-admin`, is a client-side hash route, not a
+server path.
+
+### mock_server ↔ real routes
+
+`tests/visual/mock_server.cjs` serves the twelve release-capture screens.
+
+| Measure | Value |
+| --- | --- |
+| Exact path literals in the mock | 132 |
+| Matching a real route (exact or template) | 131 |
+| Mock-only | **1** — `/v3`, a legacy alias for `/app` that the server no longer serves |
+| Prefix handlers in the mock | 17, all resolving to real routes or mounts |
+| Real routes the frontend can call that the mock does not serve | 183 |
+
+The 183 are capture-fidelity headroom, not product defects: the mock answers
+what the twelve screens need and falls through for the rest. The gap is bounded
+by a real gate — `tests/visual/v3.spec.js` asserts
+`service-unavailable-banner` count is zero across all ten plain-mode routes, so
+a mock gap that reaches a core panel fails the visual suite rather than shipping
+a screenshot of a working feature rendered as broken.
+
+## Verdict table
+
+`WORKS` — the claim was exercised and held. `PARTIAL` — the mechanism exists
+and is tested, but a surface the sentence implies does not reach it.
+`MISMATCH` — the doc said something the code does not do.
+
+| # | Row | Verdict | Evidence |
+| --- | --- | --- | --- |
+| 1 | Brain Home | **MISMATCH → doc fixed** | Living Brain + composer in-viewport asserted at 1280×800 and 390×780 (`tests/visual/v3.spec.js`); the Brain Brief renders only inside `BrainHomeDock`'s `stats` tab, and `BrainHome.test.tsx` asserts the dock opens closed |
+| 2 | Automation Intelligence | WORKS | `GET /api/automation/{overview,patterns,suggestions}` → 200 with `consent.default_state: "draft_disabled"`, `requires_user_enable: true`, `review_before_run: true` |
+| 3 | Brain Intelligence | **PARTIAL → code fixed** | `GET /api/brain/health` → 200 with all four dimensions; on an empty Brain three read `unavailable` yet the composite reported `overall_score: 100, grade: "excellent"` (see Finding 3) |
+| 4 | Temporal Knowledge | WORKS | `KnowledgeGraphStore.as_of(timestamp, …)` and `neighbors(…, as_of=None)` present with the documented defaults; `tests/unit/test_t2_temporal_schema.py`, `test_t2_temporal_reads.py`, `test_kg_temporal.py` |
+| 5 | Proactive Synthesis | WORKS | `POST /api/brain/synthesize` → 200 with the four proposal kinds; `SYNTHESIS_THRESHOLD_ENV` / `DEFAULT_SYNTHESIS_THRESHOLD = 25`; `tests/unit/test_t2_synthesis_job.py` |
+| 6 | Memory Decay | WORKS | `GET /api/brain/importance` → 200 (`half_life_days: 30.0`, `access_source: "metadata"`); `/api/brain/quality-report` carries `importance` + `tidying`; `tests/unit/test_t2_importance_decay.py` |
+| 7 | Hybrid Recall | WORKS | `POST /api/memory/recall` → 200 with `quality_gate.gate == "hybrid-evidence/v2"`; `GET /api/search/hybrid` reports per-channel weights and recall detail |
+| 8 | Folder Ingestion | WORKS | `GET /api/ingestion/jobs` → 200; `/api/ingestion/folder`, `/jobs/{id}`, `/jobs/{id}/resume` present |
+| 9 | Extraction Quality | WORKS | `assess_extraction_quality` + observe-mode `_observe_quality_gate` (every failure path returns `None` and the ingest proceeds); `tests/unit/test_ingestion_quality_jobs.py` |
+| 10 | Vector Freshness | **PARTIAL → code fixed** | `GET /api/brain/vector-freshness` returned only `{status, pending_items, total_items, detail}` while `vector_freshness_breakdown()` had **no product caller** (see Finding 4) |
+| 11 | Vector Index Backends | WORKS | `resolve_vector_index({})` → `brute`/exact; an unknown name falls back **and says why** in `detail`; choices are exactly `brute`/`quantized`/`hnsw` |
+| 12 | Background Embedding | WORKS | `IngestionPipeline.drain_vector_queue` present; the Known Limitation that nothing schedules it is accurate |
+| 13 | Fusion Strategy | WORKS (off) | `LATTICEAI_FUSION_STRATEGY` unset in a clean boot; alpha fusion is what `/api/search/hybrid` reports |
+| 14 | Graph Candidate Expansion | WORKS (off) | `graph_expansion_enabled()` false with no env; `GRAPH_EXPANSION_ENV` as documented |
+| 15 | Self-Model (Personal Ontology) | WORKS | `POST /api/memory/self-model/propose` with `저는 …를 선호합니다` returned 1 candidate and **one pending review item**, nothing written; `apply` calls `review_queue.approve()` before `_write_fact`; summary capped at half the context budget |
+| 16 | Workspace Reorganization | WORKS | one `folder_reorganization` proposal, targets always `topics/<주제>/`, `unplaced` reasons, **no delete path exists in the module**, skip-and-report on apply |
+| 17 | Change Governance | WORKS | `MUTATING_TOOL_INVENTORY` + `assert_governance_coverage`; base-SHA conflict → 409; `_atomic_write` via `os.replace` under a lock; verifier fails closed to `NEEDS_REVIEW` |
+| 18 | Brain Brief | WORKS | `/api/memory/brain-brief` and `/api/brain/proactive-brief` → 200; the proactive brief counts existing proposals and raises none |
+| 19 | Conversation | WORKS | `tests/unit/test_chat_no_model.py`; the loaded-model list is empty in a clean boot and chat refuses rather than inventing output |
+| 20 | Knowledge Graph | WORKS | `/api/graph`, `/knowledge-graph/stats`, `/api/knowledge-graph/provenance` → 200; v2 schema reported alongside v1 |
+| 21 | Source Capture | WORKS | unified pipeline reachable from files/folders/notes/web; provenance keyed by origin (no wall-clock component) |
+| 22 | Local Models | WORKS | `GET /models`, `/models/recommendations` → 200 with a real device profile; `allow_model_downloads: false` in a clean boot |
+| 23 | Installer Audit | WORKS | `auto_setup` refuses to apply on a confirmation-token mismatch (`CommandConfirmationError`); `tests/unit/test_setup_wizard.py`, `test_cov_wp04_wizard_install.py` |
+| 24 | Cloud Models | WORKS | `_load_cloud_model` raises `Missing API key env var: …` **before** the client is constructed; egress is audited before any call |
+| 25 | Telegram | WORKS | `run_bot` returns without polling on any of: no token, empty allowlist, no server session; ACL check precedes `register_chat_id` on both message and callback paths |
+| 26 | Agent Runtime | WORKS | `/agents/api/runtime/health` → `status: "unavailable"` with the honest reason ("No LLM-backed model is loaded; product execution API refuses simulation runs"); preview records no run; unknown roles rejected |
+| 27 | Tool Registry / MCP | WORKS | `/tools/registry/diagnostics` → `registered_tools 47 / governed_tools 47`, empty `governance_without_handler` and `handler_without_governance`; `/mcp/installed`, `/mcp/tools` → 200 |
+| 28 | Workspaces | WORKS | `/workspace/os` → 200 with the personal workspace default; admin surfaces on separate routes |
+| 29 | VS Code Extension | WORKS | `/workspace/vscode/status` → 200 (`connected:false, status:"offline"`); `POST /agent` `stream:true` dispatches `agent_step`; 4xx never reads as success |
+| 30 | Browser Extension | WORKS | every outbound call is `http://127.0.0.1:${port}`; an absent verdict renders `근거 확인 불가`; no approve/reject call exists in the extension |
+| 31 | Telegram Review | WORKS | `/review` reads the same `/api/proposals`; 409 → "아무것도 쓰지 않았습니다", no retry |
+| 32 | Knowledge Garden | WORKS | `GET /api/brain/garden` → 200 with exactly four beds; an unavailable graph yields empty beds |
+| 33 | Agent File Tasks | WORKS | `executor_prompt_for(base, *, profile, self_model_summary)`; `EXECUTOR_PROMPT` itself carries no file-task hints; summary injected only when passed |
+| 34 | Agent Profiles | WORKS | `COMPACT_MAX_PARAMS_B = 4.0`; an id naming no size keeps `standard`; a failed direct-path write is reported as not written |
+| 35 | Folder Memory State | WORKS | `GET /knowledge-graph/local/health` → 200; vector freshness reported once and labelled `vector_freshness_global` |
+| 36 | Voice Capture | WORKS | `GET /api/capture/voice/status` → `capture:true, transcription:false` with a plain-language reason; the transcriber really is one shared port (`ctx.MULTIMODAL_PORTS.transcriber` feeds both the pipeline and `VoiceCaptureService`) |
+| 37 | Multi-modal Memories | WORKS | `ALLOW_MULTIMODAL_ENV` default `"0"`; both modes asserted in `tests/unit/test_t3_ingest_routing.py`; `NodeType.AUDIO` and `IMAGE` are first-class |
+| 38 | Vision OCR / Caption / Embedding | WORKS | `VisionStub` is gone from product code and from the shipped wheel/sdist, with a test asserting its absence; `resolve_vision_embedder` has no hash fallback |
+| 39 | Image Retrieval | **PARTIAL → code fixed** | late fusion, the separate `image_embeddings` table, width-mismatch skipping and the `multimodal` result block are all real and tested — but no production caller passed `multimodal=` to `context_quality_signal` (see Finding 5) |
+| 40 | Evidence Thumbnails | WORKS | `THUMBNAIL_EDGE = 96`; the 24,000-char cap is enforced on write **and** on read; only `data:image/…` is accepted, server-side and client-side |
+| 41 | Video Ingestion | WORKS | ingest returns `status:"unavailable"`, `indexing_status:"skipped"`, `node_id: None`; nothing is stored |
+| 42 | Frontend Reliability | WORKS | `CoreServiceUnavailableBanner`; `ActionButton` does not fire success callbacks for `ok:false`; vitest 100% on all four metrics |
+| 43 | Trusted Agent Loop | WORKS | `LoopTrace`, the `loop` payload on both the streaming and non-streaming terminals, python-literal repair, the agent-eval CI gate |
+| 44 | Command Center | **MISMATCH → code fixed** | `/api/command/briefing` returned `health: {available:false, grade:null, score:null}` on a Brain reporting `grade: "excellent"` (see Finding 1) |
+| 45 | Evidence → Action | WORKS | `POST /api/evidence/actions` with no resolvable citations → 200 with an explicit reason, never a silent empty list |
+| 46 | Run Explanation | WORKS | `"ok": code == "done"`, reachable only from `state == "DONE"`; rendered on web, VS Code and Telegram |
+| 47 | Project Sessions | WORKS | `GET /api/projects` → 200; `/agent` accepts `project_id`; a `NEEDS_REVIEW` run never becomes a project `done` |
+| 48 | Citation Precision | WORKS | `tests/unit/test_typed_chunking.py`, `tests/unit/test_citation_grounding.py`; `plain` chunking stays byte-identical to the legacy walk |
+| 49 | Graph Relation Evidence | WORKS | `tests/unit/test_relation_evidence.py` |
+| 50 | Funnel Alerts | WORKS | `GET /api/admin/funnel-metrics` → 200 with named counters, `rates: null` and `alerts: []` below the 10-sample floor |
+| 51 | Frontend Payload | **MISMATCH → doc fixed** | `npm run check:bundle` measures **103.0 KiB** gzip initial JS against the 150 KiB budget; the doc had read "~99 KiB" since 9.9.9 (see Finding 2) |
+| 52 | Permission Modes | WORKS | `GET /api/permission-mode` → `strict` default with `proposal_first: true`; circuit breakers mode-invariant; the selector renders the server catalog and blocks `bypass` until acknowledged. One wording nuance: the home dial uses a confirmation dialog rather than a checkbox — consent is still explicit |
+| 53 | Network Boundary | WORKS | `local_only` default; `hard_block_metadata_flags` = `do_not_share/local_only/private/sensitive` applied with no mode branch; `POST /api/network-boundary/preview` answers in `local_only` and reports `would_block` |
+| 54 | Hybrid Cloud Chat | WORKS | all seven named modules exist; the adapter raises before opening a socket without `LATTICEAI_CLOUD_API_KEY`; the local path is a pure early return |
+| 55 | Cloud Memory Write-Back | **MISMATCH → code fixed** | the enqueue was implemented and unit-tested, but both live call sites constructed `CloudResponseIngestor(store=knowledge_graph)` with no review sink (see Finding 6) |
+| 56 | Obsidian Vault Bridge | WORKS | driven end to end in the signed-in pass: `permission_required` + token → `/permissions/approve/{token}` → `status: "dry_run"`, `scanned 2 / notes 2 / tags 1 / links.resolved 1` (see Finding 7 for the anonymous case) |
+| 57 | Selective Brain Network | WORKS (flag off) | `GET /api/knowledge-graph/share` → `enabled:false` with the flag named; all four mutating share routes → 403 with the reason. **The encryption half of this row was being changed by a parallel track while this audit ran — see Caveat** |
+| 58 | Release Assets | WORKS | package/pyproject/tauri/vsix all 11.1.0; `check_current_release_docs`, `check_doc_status`, `validate_release_artifacts`, `check_release_evidence_bound` all clean. Rolling the markers to 11.2.0 is the release lead's step, not this audit's |
+
+**Totals as swept: 51 WORKS, 3 PARTIAL, 4 MISMATCH.** Of the four mismatches,
+one was fixed in code during the sweep (Command Center), two were fixed in the
+docs (Brain Home, Frontend Payload) and one was reported to its owning track
+(Cloud Memory Write-Back). All three `PARTIAL` rows were reported rather than
+changed: each needed a code change in a file another track owned this sprint,
+or a behaviour decision that does not belong inside an audit. One Known
+Limitation was retired as no longer true.
+
+**Dispositions after the sweep (v11.2.0 integration close).** The four rows the
+audit could not touch itself were closed once the parallel tracks had landed —
+Findings 3, 4, 5 and 6 are all fixed in code, each with a regression test that
+drives the shipped entry point rather than the collaborator (see
+`tests/unit/test_t8_integration_close.py`). The verdict column above keeps the
+*swept* verdict with its disposition appended, because a point-in-time audit
+that silently rewrites itself is worth nothing; `FEATURE_STATUS.md` carries the
+current state.
+
+## Findings
+
+### Finding 1 — the Command Center's health section was reading a key nothing writes (FIXED)
+
+`CommandCenterService._health_section` read the Brain's health out of a nested
+`overall` block:
+
+```python
+overall = report.get("overall") or {}
+"available": overall.get("score") is not None,
+```
+
+`BrainIntelligenceService.health_report()` — the only producer — publishes
+`overall_score` and `grade` at the **top level**. So the briefing's health
+section was permanently `{available: false, grade: null, score: null}` while
+`/api/brain/health` reported `grade: "excellent"` on the very same graph, and
+the `check-health` quick action below it could never fire.
+
+User-visible: `DailyBriefingPanel` renders
+`{health.available && health.grade ? String(health.grade) : "—"}`, so Today's
+Briefing showed a permanent dash with no reason beside it — the exact failure
+mode the 9.9.7 rule ("a `—` must state why") exists to prevent.
+
+Why the suite stayed green: `tests/unit/test_command_center.py`'s `FakeBrain`
+returned `{"overall": {"score": …, "grade": …}}` — a shape no real collaborator
+has ever produced. The test and the code agreed with each other and neither
+agreed with production.
+
+Fixed in `latticeai/services/command_center.py` (read the real shape), the fake
+corrected to match its real counterpart, and pinned by
+`tests/unit/test_a3_feature_audit.py`, which wires the two **real** services
+together — the only arrangement that can catch this class of drift.
+
+### Finding 2 — the initial-JS figure had not been re-measured since 9.9.9 (FIXED)
+
+FEATURE_STATUS claimed "~99 KiB gzip". `node scripts/check_bundle_budget.mjs`
+measures **103.0 KiB** (index 66.1 + utils 29.2 + useQuery 7.6), against a build
+`scripts/check_frontend_build_freshness.mjs` confirms is byte-identical to
+source. The budget gate itself passes with room to spare — only the sentence was
+stale. Lowered to the measured figure with the tool that measures it named.
+The same stale number also appears in `ARCHITECTURE.md`.
+
+### Finding 3 — an empty Brain grades itself 100 / "excellent" (FIXED after the sweep)
+
+`health_report()` averages only the dimensions it could measure. On a
+first-run Brain, freshness, connectivity and consistency all report
+`status: "unavailable", score: null` and embedding coverage reports 100 (zero
+of zero indexed) — so the composite is `overall_score: 100, grade: "excellent"`
+from a single dimension over an empty graph.
+
+Each dimension is individually honest; the composite is not. This is a
+behaviour question rather than a documentation one (the row does not promise a
+specific empty-Brain grade), and changing the score would move a number the
+frontend and its tests render — so it is named here rather than changed inside
+an audit.
+
+**Disposition.** The lie was upstream of the composite: an index holding
+nothing reports `coverage_ratio: 1.0`, which is 100% of nothing. That dimension
+now reads `unavailable` with the reason `"no indexable items yet"` whenever
+`source_items` is zero, so the empty Brain has *no* measurable dimension and
+the existing `scores or None` path already returns `overall_score: None,
+grade: null` — which the frontend already renders as 데이터 없음 / "No data
+yet". Two additions keep the partial cases honest too: every unavailable
+dimension now carries a `reason` (the graph could not be read vs. nothing saved
+yet vs. no relationships yet), and the report gains a `coverage` block
+(`measured` / `total` / `unavailable` / `partial`) so a verdict drawn from two
+of four checks says so. When nothing could be measured a top-level `reason`
+names each gap — the 9.9.7 rule that a "—" must state why.
+
+### Finding 4 — `vector_freshness_breakdown()` has no surface (FIXED after the sweep)
+
+The Vector Freshness row explains that "12 pending" hides the difference
+between twelve never-embedded imports and twelve edits whose current answers
+are quietly wrong, and names `vector_freshness_breakdown()` as the answer. The
+method exists on the store and is tested — and has **zero product callers**.
+`GET /api/brain/vector-freshness` still returns the aggregate
+(`{status, pending_items, total_items, detail}`), so the distinction the
+sentence sells is not visible anywhere in the product.
+
+Either the endpoint should carry the breakdown or the row should say the split
+is store-level only. The fix lives in store/API files owned by another track
+this sprint, so it is reported rather than edited.
+
+**Disposition.** The endpoint carries it. `BrainIntelligenceService.vector_freshness`
+now returns the four-key contract from an unchanged private method and attaches
+`breakdown` beside it when the store can compute one. The key is **omitted**
+rather than zero-filled for every reason it cannot be had — graph off, an older
+store without the method, an unreadable index, an empty answer — because a
+zeroed backlog nobody measured is exactly the dishonesty the split exists to
+remove. The frontend chip reads `pending_items` and is untouched.
+
+### Finding 5 — `context_quality` never gains its `multimodal` key in production (FIXED after the sweep)
+
+The Image Retrieval row says `context_quality` "gains a `multimodal` key only
+when image nodes are really in the context". The mechanism is real and tested,
+but the only call site that passes `multimodal=` to `context_quality_signal` is
+`context_for_query(with_meta=True)`, which has no production caller. Both live
+producers of `context_quality` — `api/chat_helpers.build_context_quality` and
+`core/context_builder` — discard the hybrid result's `multimodal` block. Net
+effect: on every shipped surface `context_quality` keeps the four-key shape
+even when the answer cites `Image` nodes. Owned by the multi-modal track.
+
+**Disposition.** Both producers pass it now. `build_context_quality` calls
+`multimodal_signal(matches)` on **both** of its arms — the hybrid store and the
+lexical-only fallback, since a picture found by keyword is still a picture —
+and `retrieve_context_for_generation` does the same for the document path. The
+present-only-when-true contract is unchanged: `multimodal_signal` returns
+`None` for an all-text result set and `context_quality_signal` omits the key,
+so an all-text answer is byte-identical to what 11.1.0 returned.
+
+### Finding 6 — cloud write-back is never enqueued on the live path (FIXED after the sweep)
+
+`CloudResponseIngestor.ingest` stages a `change_proposal` **only when a
+`review_queue` sink is bound**. Both production call sites bind none:
+
+```
+latticeai/services/hybrid_chat.py:  ingestor = CloudResponseIngestor(store=knowledge_graph)
+latticeai/services/hybrid_chat.py:  ingest_status = CloudResponseIngestor(store=knowledge_graph).ingest(plan)
+```
+
+With `plan.auto_commit` hardcoded `False`, `ingest()` returns
+`{"status": "staged", "review_item_id": None, "written_nodes": 0}` — knowledge
+extracted from a cloud answer is neither written nor queued. It is discarded
+after the `hybrid_done` frame. The unit tests pass a `review_queue` directly to
+the ingestor, so the wiring gap is invisible to CI. Related: the hybrid policy's
+`auto_commit` flag is settable through `POST /api/network-boundary/policy` but
+has no consumer.
+
+The row's "is enqueued as a Review Center `change_proposal`" is therefore not
+true as shipped. Threading a review sink through
+`stream_hybrid_cloud_turn` / `run_hybrid_cloud_turn` / `maybe_hybrid_stream_response`
+touches the chat router and the review-queue seam, both of which other tracks
+were editing during this sweep — reported, not changed.
+
+**Disposition.** Threaded, exactly there. Both turn functions take
+`review_queue` and `auto_commit` as arguments (defaults `None` / `False`, so a
+headless caller still stages nothing), `maybe_hybrid_stream_response` forwards
+the sink and resolves the flag through the new
+`chat_hybrid.resolve_hybrid_auto_commit`, and `/chat` supplies the live queue
+through a new `AppContext.review_queue` **provider** — a callable rather than a
+value, because `REVIEW_QUEUE` is produced in `phase_platform_features`, two
+phases after the context is built, and a captured value would have been `None`
+forever. `resolve_hybrid_auto_commit` treats an unreadable policy as the
+default rather than as permission.
+
+The reason CI could not see this — the unit tests handed the ingestor its
+collaborator directly — is also the reason the new test drives `POST /chat`
+with the dial on `cloud_allowed` and a fake adapter, and asserts the item
+landed in the queue with `source: "change_proposal"`.
+
+### Finding 7 — the local-file 401 is a gate, not a defect (CLEARED)
+
+Worth recording because it looks like a defect and is not. In the default local
+profile (`require_auth = false`), `POST /api/ingestion/obsidian` and every other
+local-file route answer **401** to an anonymous caller:
+`PermissionGateway.require_local_user` resolves the raw session identity and
+does **not** fall through to the anonymous local owner that `require_user`
+returns. Reading only that line suggests the vault bridge is unreachable out of
+the box.
+
+It is not. Driven end to end in the signed-in pass — register (first user
+becomes admin) → login → `POST /api/ingestion/obsidian` →
+`permission_required: true` with an `approval_token` →
+`POST /permissions/approve/{token}` → re-post with the token →
+`status: "dry_run", scanned: 2, notes: 2, tags: 1, links.resolved: 1`. That
+sequence *is* the "standard local-read approval dance" the row describes.
+Filesystem access deliberately demands a named account even where the rest of
+the local API accepts the anonymous owner. Verdict: WORKS.
+
+### Finding 8 — a retired Known Limitation
+
+"The approval card under 작업 → 실행 still labels raw payload fields (`Action`,
+`Action Label`, `User Email`) in plain mode" is no longer true. `Act.tsx`'s
+`HumanPermissionDetails` renders a translated action badge, the target
+filename, a free-text reason and a labelled requester line, with a regression
+test forbidding raw i18n keys from reaching the DOM; the repo carries no
+rendered `Action Label` / `User Email` literal. The screenshot the bullet cited
+as proof — `output/release/v10.6.3/screenshots/09-automation-runs.png` — no
+longer exists (release evidence now starts at v11.0.0). The bullet was removed;
+carrying a fixed defect as a standing limitation is its own kind of dishonesty.
+
+## prefix-static hermeticity (wp19 / wp36) — re-confirmed
+
+`Config.from_env` resolves a *missing* static dir to `sys.prefix/static` when
+that exists, which on a pip-installed machine it genuinely does. `_empty_prefix`
+in `tests/unit/test_cov_wp19_build_phases.py` and
+`tests/unit/test_cov_wp36_build_phases_late.py` neutralizes that so the
+assertions describe the resolver rather than the runner.
+
+Re-verified by simulating the release-runner condition for a whole session: a
+pytest plugin pointed `sys.prefix` at a temp directory that **does** carry
+`static/` (with a sentinel file), and the two suites plus
+`test_cov_wp19_cli_entrypoint.py` were run under it.
+
+* **81 passed.**
+* The simulated prefix's `static/` still contained only the sentinel afterwards
+  — no test wrote into it.
+* The hazard itself was confirmed live under the same plugin: with a missing
+  configured static dir, `Config.from_env` resolves to the fake prefix's
+  `static/`. The hermeticity is load-bearing, not decorative.
+
+## Caveat — a concurrently mutating tree
+
+This sweep ran while two other tracks were writing to the same working tree.
+Files under `lattice_brain/` (portability, sealed box, multimodal, ingestion,
+graph schema) and `latticeai/services/model_*` were changing during the run, and
+one intermediate state was observed directly (a partially-rewired
+`latticeai/services/obsidian_bridge.py` raising `NameError` mid-refactor, and a
+`KnowledgeGraphStore` attribute warning at boot). Rows 37–41, 55 and 57 are
+therefore reported **as of this snapshot**; their owning tracks are the
+authority on where they land. Everything measured for rows outside those files
+was stable across repeated runs.
+
+## Changes this audit made
+
+* `latticeai/services/command_center.py` — read the health shape the Brain
+  actually publishes (Finding 1).
+* `tests/unit/test_command_center.py` — corrected the `FakeBrain` fixture to
+  its real counterpart's shape.
+* `tests/unit/test_a3_feature_audit.py` — new. Pins the two drift classes this
+  audit found: a doc naming an endpoint the app does not serve, and the
+  briefing reading a health score the Brain never published.
+* `FEATURE_STATUS.md` — Brain Home (Brain Brief is on the dock, not the first
+  screen), Frontend Payload (103 KiB, measured), and the retired approval-card
+  limitation.