ltcai 9.9.9 → 10.2.0

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)
 
-![v9.9.9 Living Brain walkthrough](output/release/v9.9.9/gifs/v9.9.9-living-brain-walkthrough.gif)
+![v10.2.0 Living Brain walkthrough](output/release/v10.2.0/gifs/v10.2.0-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
@@ -24,9 +24,9 @@ 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/v9.9.9/screenshots/04-brain-chat-home.png) | **See how knowledge connects** — a real relationship graph, not a file list ![Memory Graph](output/release/v9.9.9/screenshots/05-memory-graph.png) |
-| **Capture anything** — files, whole folders, notes, screenshots, web pages ![Capture](output/release/v9.9.9/screenshots/06-capture.png) | **Automate with review** — agent changes become proposals you approve first ![Review Center](output/release/v9.9.9/screenshots/12-review-center.png) |
-| **Pick a model in one click** — recommended local models for your hardware ![Recommended Models](output/release/v9.9.9/screenshots/02-recommended-models.png) | **Stay in control** — audit, roles, retention in a separate admin surface ![Admin Console](output/release/v9.9.9/screenshots/10-admin-console.png) |
+| **Chat with a Brain that remembers** — every conversation grows durable, source-linked memory ![Brain Chat](output/release/v10.2.0/screenshots/04-brain-chat-home.png) | **See how knowledge connects** — a real relationship graph, not a file list ![Memory Graph](output/release/v10.2.0/screenshots/05-memory-graph.png) |
+| **Capture anything** — files, whole folders, notes, screenshots, web pages ![Capture](output/release/v10.2.0/screenshots/06-capture.png) | **Automate with review** — agent changes become proposals you approve first ![Review Center](output/release/v10.2.0/screenshots/12-review-center.png) |
+| **Pick a model in one click** — recommended local models for your hardware ![Recommended Models](output/release/v10.2.0/screenshots/02-recommended-models.png) | **Stay in control** — audit, roles, retention in a separate admin surface ![Admin Console](output/release/v10.2.0/screenshots/10-admin-console.png) |
 
 ## Why Lattice AI
 
@@ -57,48 +57,53 @@ First-run flow — wake the Brain, pick the owner, load a recommended model:
 
 | | | |
 | --- | --- | --- |
-| ![Login](output/release/v9.9.9/screenshots/01-login.png) | ![Model install](output/release/v9.9.9/screenshots/03-install-load-progress.png) | ![Model library](output/release/v9.9.9/screenshots/07-model-library.png) |
+| ![Login](output/release/v10.2.0/screenshots/01-login.png) | ![Model install](output/release/v10.2.0/screenshots/03-install-load-progress.png) | ![Model library](output/release/v10.2.0/screenshots/07-model-library.png) |
 
 Screenshot index and capture notes:
-[output/release/v9.9.9/SCREENSHOT_INDEX.md](output/release/v9.9.9/SCREENSHOT_INDEX.md)
+[output/release/v10.2.0/SCREENSHOT_INDEX.md](output/release/v10.2.0/SCREENSHOT_INDEX.md)
 
 ## Current Release
 
-The current release is **9.9.9 — Lean Shell**:
-
-- **No `✖` left in the parity matrix.** Every gap 9.9.6 wrote down is closed,
-  and every remaining `—` states *why* it is a boundary
-  ([docs/SURFACE_PARITY.md](docs/SURFACE_PARITY.md)).
-- **VS Code watches the loop live.** `POST /agent` streams the same
-  `agent_step` frames the web app gets, and cited evidence turns into one-click
-  follow-ups in the editor.
-- **Telegram badges its recall and reviews changes.** The server's grounding
-  verdict is shown as-is, and `/review` approves or rejects staged proposals
-  inline.
-- **The browser extension recalls, not just captures.** Ask your Brain from the
-  popup with the same badge, and see how many runs await approval — still
-  posting only to `127.0.0.1`.
-- **A knowledge garden in four beds** — what arrived, what disagrees, what went
-  stale, what the rest leans on.
-- **A profile for small local models.** Under 4B, the loop shortens its window,
-  escalates sooner, and — when JSON tool calls keep failing — writes the
-  planned files directly instead of ending empty-handed.
-- **Folder memory state.** How much of each connected folder is really in your
-  Brain, what failed, and the reason it failed.
-- **Two skills that pay off on install** — meeting notes and a weekly review
-  written only from what the Brain actually recorded.
-- **Voice memos** go straight into the Brain; when no local transcriber exists
-  the memo is still stored and honestly marked not-searchable.
+The current release is **10.2.0 — Load-Bearing Fixes**:
+
+A full code review of 10.1.1 scored the codebase 71/100 and found two things
+that were true but invisible. This release is that review's response.
+
+- **A database connection leak that hid behind garbage collection.**
+  `with sqlite3.connect(...)` commits but never closes, and 70+ call sites
+  relied on it. Nothing showed until something held a stack frame alive — a
+  profiler, a logged traceback, a coverage tracer — and then descriptors ran
+  out. Every one of those sites now closes.
+- **Coverage is measurable for the first time: 71%.** It could not be run
+  before, because running it *was* the thing that exhausted descriptors. There
+  is now a floor in CI so it cannot silently fall.
+- **"Sensitive memories are never sent" is finally true.** The filter was real,
+  correct, and wired — but nothing in the product could mark a memory, and the
+  blocked-type list was empty, so it could never fire. You can now mark any
+  memory from the boundary panel, files under `.ssh`/`.aws`/`.env` and friends
+  are flagged automatically at ingestion, and credential-shaped node types are
+  blocked outright.
+- **Knowledge that leaves is now redacted and recorded.** Secret-shaped text is
+  stripped from the outbound payload, and every send — and every refusal —
+  writes an audit entry naming what went, how much, and where to.
+- **112 silent `except: pass` handlers now say something.** Behaviour is
+  unchanged; each records the failure at debug level, so a genuine bug in an
+  optional path is no longer indistinguishable from the optional thing being
+  absent.
+- **The lint net was widened to the bug classes that actually bit** — loop
+  variable capture, truncating `zip`, silent handlers — and mypy now runs on
+  the trust-critical modules. CI gained macOS and Python 3.14, so the machine
+  this is developed on and the platform it ships a `.dmg` for are both tested.
 
 Release notes: [RELEASE.md](RELEASE.md) · Full history: [docs/CHANGELOG.md](docs/CHANGELOG.md)
 
-Expected artifacts for 9.9.9 release must use exact filenames:
+Expected artifacts for 10.2.0 release must use exact filenames:
 
-- `dist/ltcai-9.9.9-py3-none-any.whl`
-- `dist/ltcai-9.9.9.tar.gz`
-- `ltcai-9.9.9.tgz`
-- `dist/ltcai-9.9.9.vsix`
-- `src-tauri/target/release/bundle/dmg/Lattice AI_9.9.9_aarch64.dmg`
+- `dist/ltcai-10.2.0-py3-none-any.whl`
+- `dist/ltcai-10.2.0.tar.gz`
+- `ltcai-10.2.0.tgz`
+- `dist/ltcai-10.2.0.vsix`
+- `src-tauri/target/release/bundle/dmg/Lattice AI_10.2.0_aarch64.dmg`
 
 Do not use wildcard artifact uploads. Package registry publishing remains owner-run.
 
@@ -118,18 +123,26 @@ See [ARCHITECTURE.md](ARCHITECTURE.md) for details and
 
 - External package registries are owner-published and can lag behind GitHub.
 - PostgreSQL/pgvector is optional scale/migration tooling. SQLite remains the
-  live local Brain store in 9.9.5.
+  live local Brain store in 10.2.0.
 - Docker, model downloads, cloud model calls, Telegram, Brain Network, and
   update checks require explicit user action.
 - Conversation does not fabricate answers when no model is loaded. Agent and
   workflow simulation without a loaded LLM is deterministic and LLM-free (it
   does not call a model) — labeled as such, never presented as autonomous
   model success.
+- Some backend-generated messages (for example the Postgres DSN notice) are
+  produced server-side in English and are shown as-is; server-side i18n is not
+  part of 10.2.0.
 
 ## Release History
 
 | Version | Theme |
 | --- | --- |
+| 10.2.0 | Load-Bearing Fixes |
+| 10.1.1 | Reachable Boundary |
+| 10.1.0 | Hybrid Brain |
+| 10.0.1 | One Source of Truth |
+| 10.0.0 | Plain Language |
 | 9.9.9 | Lean Shell |
 | 9.9.8 | Autonomy Dial |
 | 9.9.7 | No Gaps Left |

package/docs/CHANGELOG.md

@@ -4,6 +4,203 @@ 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.
 
+## [10.2.0] - 2026-07-29
+
+Response to a full code review of 10.1.1 (71/100). All twelve findings.
+
+### Fixed
+- **SQLite connections were committed but never closed** at 70+ sites.
+  `sqlite3.Connection.__exit__` commits or rolls back and leaves the connection
+  open; `KnowledgeGraphStore._connect`, `ConversationStore._connect`,
+  `WorkspaceOSStore._connect_state_db`, both storage engines, `portability`,
+  `archive`, `migration` and `users` all relied on it. Refcounting hid the leak
+  until something held a frame alive (profiler, logged traceback, coverage
+  tracer), at which point the process hit `EMFILE`. Added
+  `StorageEngine.session()` and made the store helpers closing context
+  managers; `tests/unit/test_connection_lifecycle.py` asserts it.
+- **The cloud sensitivity guard could not fire.** `HARD_BLOCK_NODE_TYPES` was
+  an empty set and nothing in the product could set the metadata flags
+  `is_node_blocked_for_cloud` matches on, so "sensitive memories are never
+  sent" was unfalsifiable. Populated the type list, added
+  `lattice_brain/sensitivity.py` (path rules, stamped at ingestion), added
+  `KnowledgeGraphStore.set_node_sensitivity` and
+  `POST /api/network-boundary/node-sensitivity`, and a per-memory hold-back
+  control in the boundary panel.
+- **Outbound cloud text was not redacted.** `redact_secret_text` ran on logs,
+  audit records and API previews but not on the payload that actually leaves.
+- **Vector similarity truncated silently on a dimension mismatch**, producing a
+  plausible but meaningless score. Now raises with both dimensions named.
+- **`tempfile.mktemp()` (TOCTOU race) in the Telegram bridge** replaced with
+  `mkstemp` at two sites.
+- **Three closures captured loop variables by reference** (`chat_agent_http`,
+  `computer_use`, `model_runtime`). Safe today because each is invoked within
+  its iteration; bound explicitly so deferring the call cannot break them.
+- **The cloud turn ran retrieval twice**, discarding the first result except
+  for its quality dict.
+
+### Added
+- `latticeai/services/cloud_egress_audit.py` — every cloud send and every
+  refusal writes an audit entry *before* the provider call, naming node ids,
+  count, token estimate, provider and mode. Shape, never content.
+- `latticeai/core/quiet.py` + `lattice_brain/quiet.py` — `quiet()` keeps the
+  behaviour of `except: pass` and records the exception with its file,
+  function and line at DEBUG. 112 handlers converted; `except: pass` is now a
+  lint error (S110/S112).
+- Coverage measurement: `pytest-cov` in the `dev` extra, `[tool.coverage]`
+  config, a **70% floor** in CI. Measured **71%** — the first known figure.
+- `tests/unit/test_cloud_egress_guard.py` — end-to-end proof that a flagged
+  memory never reaches the payload, that retrieval still carries `metadata`
+  (whose loss would silently disable the filter), and that egress is audited.
+- mypy over 13 trust-critical modules (`npm run typecheck:python`), wired into
+  `npm run lint:python`.
+- CI matrix: macOS runner (the platform the `.dmg` ships for, previously
+  untested) and Python 3.14 (the version this is developed on).
+
+### Changed
+- ruff widened from pyflakes-only to `E4,E7,E9,F,B,S,I,SIM,RET,C901`. Every
+  remaining finding is either fixed or carries a written reason; complexity
+  ceiling set to 30 with named exemptions.
+- `latticeai/core/workspace_os.py` 1354 → 1130 lines; vocabulary moved to
+  `workspace_os_constants.py` and state/migration to `workspace_os_state.py`,
+  with delegating methods so the public surface is unchanged.
+- Package classifiers now advertise Python 3.13 and 3.14, matching what CI tests.
+
+### Known
+- `app_factory._build` remains one 1,343-line function. Its sections share ~34
+  local bindings, so decomposition requires a runtime-context object rather
+  than a mechanical extraction; deferred rather than rushed, and named in the
+  C901 exemptions so the choice is visible.
+
+## [10.1.1] - 2026-07-28
+
+### Added
+- `NetworkBoundaryPanel` (`frontend/src/components/NetworkBoundaryPanel.tsx`),
+  mounted in the System settings tab beside `PermissionModePanel`. Renders the
+  server's mode catalog, gates `cloud_allowed` behind the acknowledgement the
+  server requires, and exposes the hybrid write-back policy switches only while
+  the boundary permits cloud.
+- Context preview in that panel: a question is sent to
+  `POST /api/network-boundary/preview` and the panel lists the actual node
+  titles that would leave, the token estimate, and the token guard's verdict.
+  Available in `local_only` too, labelled as hypothetical.
+- API client bindings and types: `networkBoundary`, `setNetworkBoundary`,
+  `setHybridPolicy`, `previewCloudContext`; `NetworkBoundaryOption`,
+  `NetworkBoundaryState`, `HybridPolicy`, `CloudContextPreview`,
+  `CloudTokenBudget`. The `networkBoundary` fallback is `local_only`, so a
+  failed read never renders as if cloud were already permitted.
+- 22 `system.network.*` copy keys in the `workspace` i18n namespace (ko + en).
+- `/api/network-boundary*` routes in `tests/visual/mock_server.cjs` so release
+  evidence shows the panel working rather than an unavailable state.
+- `NetworkBoundaryPanel.test.tsx` (12 tests) and two Playwright specs covering
+  the settings-tab route, the acknowledgement gate, and the preview.
+
+### Removed
+- `static/app/network-boundary-panel.js` — the unmounted standalone module
+  10.1.0 shipped as this feature's UI. Nothing referenced it, it was absent
+  from the asset manifest, and it lived in Vite's output directory, which
+  `npm run build:assets` wipes; it had in fact already been deleted by the
+  previous release's asset rebuild. The React panel replaces it.
+
+### Changed
+- ARCHITECTURE, FEATURE_STATUS, README, and the hybrid design doc no longer
+  carry the "API-only surface" caveat, and describe the in-app control instead.
+
+## [10.1.0] - 2026-07-28
+
+### Added
+- Local-first hybrid path: the Knowledge Graph stays on-device and cloud LLMs
+  are an opt-in worker. `NetworkBoundaryMode` and `MinimalContext` contracts
+  (`latticeai/core/network_boundary.py`, `latticeai/api/network_boundary.py`)
+  with a persisted dial wired via `latticeai/runtime/network_boundary_wiring.py`.
+- Cloud streaming worker: `latticeai/services/cloud_streaming.py`,
+  `cloud_extraction.py`, `cloud_token_guard.py`, and an OpenAI-compatible stream
+  adapter (`openai_compatible_adapter.py`) with token budgets and guardrails.
+- Hybrid chat pipeline: `latticeai/api/chat_hybrid.py` plus
+  `services/hybrid_chat.py`, `hybrid_context.py`, and `hybrid_policy.py`; the
+  `/chat` route branches through it and writes back through the Review Queue.
+- Multimodal streaming contracts (`services/multimodal_streaming.py`) and a
+  network-boundary control panel (`static/app/network-boundary-panel.js`).
+- Tests: `test_network_boundary.py`, `test_hybrid_phase2.py`,
+  `test_hybrid_phase3.py`.
+
+### Notes
+- Default network boundary is `local_only`; cloud requires explicit
+  acknowledgement. Minimal related nodes only leave the machine; streamed
+  answers expand the local Brain with provenance.
+
+## [10.0.1] - 2026-07-28
+
+### Changed
+- `latticeai/core/agent.py` holds only the state machine (1769 → 1326 lines).
+  Its pure functions move to `latticeai/core/agent_helpers.py`: action parsing
+  (`extract_action`, `extract_action_details`), plan normalization
+  (`normalize_plan`), learning filters, transcript compaction, artifact and
+  requirement reporting, and the `TranscriptBudget` / `PhaseBudgets` dataclasses.
+- `AgentState` and `AGENT_TERMINAL_STATES` move to a new
+  `latticeai/core/agent_state.py`. Both `agent.py` and `agent_helpers.py` import
+  from it, which is what lets the helpers reference the enum instead of a
+  literal state string.
+- `latticeai.core.agent` re-exports every moved name as the same object and
+  declares the set in `__all__`. No caller changed: the HTTP layer, `run_store`,
+  `computer_use`, `bench_agent_smoke`, `bench_models`, and eight test modules
+  import exactly as before.
+- The home's secondary control row is visually demoted (opacity, tighter gaps,
+  softer divider) and gets larger touch spacing under 640px. The sticky top bar
+  border is softened. CSS only.
+
+### Fixed
+- `files_written` and `artifact_checklist` matched transcript steps against the
+  literal `"EXECUTING"` instead of `AgentState.EXECUTING.value`. A rename of the
+  enum value would have made both silently return empty results with no failing
+  test. Both now reference the enum.
+
+### Verification
+- pytest 1747 passed / 11 skipped; `scripts/agent_eval.py` 23/23 (100%); ruff
+  clean across `latticeai/`, `lattice_brain/`, `scripts/`, `tests/`.
+- All 18 extracted symbols were AST-compared against their originals before the
+  originals were removed.
+
+## [10.0.0] - 2026-07-28
+
+### Changed
+- The Brain home is four zones: the living Brain, one composer, the autonomy
+  dial, and capture. `BrainHomeHero` replaces the knowledge-flow canvas on the
+  home; the graph is reached by clicking the Brain.
+- File / folder / note / web capture render inside the composer toolbar
+  (`BrainIngestionDock variant="inline"`), on both the home and the active
+  conversation. The standalone "Brain에게 가르치기" panel is removed.
+- The top bar carries the language switch and the appearance toggle, so both
+  are reachable from every screen.
+- Provenance coverage is reported as counts with an explanation
+  (`CoverageMeter`) instead of a bare percentage.
+- Model identifiers are humanized for display (`humanizeModelId`); the catalog
+  display name wins when the id is known.
+- Settings splits appearance (light/dark) from detail level (basic/advanced/
+  admin); they were one card labeled "화면 모양".
+- Model list is ordered by what can be used now: loaded, ready, downloaded,
+  then the rest.
+
+### Fixed
+- The folder button did nothing in a browser: `pickFolder` called the
+  desktop-only `selectFolder`. `frontend/src/lib/folderPicker.ts` now owns the
+  desktop / File System Access / file-input decision for every surface.
+- The conversation header Brain rendered at 311px — `size="trace"` had no size
+  rule and inherited the base `clamp(220px, 28vw, 320px)`.
+- The sticky composer covered the end of the newest answer; the stream now
+  reserves room, and nothing renders after the composer.
+- `ValuePreview` printed a nested object's field names as its value
+  ("Runtime → Ready, Version, Execution Mode, Mode +4").
+- `white-space: nowrap` on controls was inherited by descriptive text inside
+  them, clipping a 347px sentence to 288px in the autonomy panel.
+- Markdown written by a model leaked into search-result titles and graph node
+  labels; `plainText` strips it.
+- Duplicate `ui.field.health` keys in the shell namespace.
+
+### Added
+- ko/en parity across memory tiers, agent roles, automation recipes and their
+  outputs, entity types, status badges, and backend payload field labels.
+- `plainText` and `humanizeModelId` in `frontend/src/lib/utils.ts`, with tests.
+
 ## [9.9.9] - 2026-07-27
 
 ### Changed

package/docs/COMMUNITY_AND_PLUGINS.md

@@ -1,6 +1,6 @@
 # Community And Plugins
 
-Current release: **9.9.9 — Lean Shell**.
+Current release: **10.2.0 — Load-Bearing Fixes**.
 
 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: **9.9.9 — Lean Shell**.
+Current release: **10.2.0 — Load-Bearing Fixes**.
 
 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/HYBRID_CLOUD_KG_STREAMING.md

@@ -0,0 +1,101 @@
+# Hybrid Cloud + Local Knowledge Graph Streaming
+
+> Status: **shipped in 10.1.0**; in-app control added in **10.1.1**  
+> Principle: **Local Brain is the asset. Cloud is an opt-in worker.**  
+> Surface: `환경설정 → 내 지식이 나가는 범위`
+> (`frontend/src/components/NetworkBoundaryPanel.tsx`), plus
+> `/api/network-boundary` and `LATTICEAI_NETWORK_MODE`.
+
+## Goal
+
+Local Knowledge Graph stays on-device. Cloud LLMs (and optional multimodal) are opt-in workers. Minimal related nodes leave the machine; streamed answers expand the local Brain with provenance and review gates.
+
+## Phase summary
+
+| Phase | Branch | Delivered |
+|-------|--------|-----------|
+| 0 | `feature/hybrid-cloud-kg-streaming` | Architecture + core contracts |
+| 1 | `…-phase1` | Dial, OpenAI adapter, hybrid helpers, basic expansion |
+| 2 | `…-phase2` | `/chat` branch, token guardrails, extraction, preview API |
+| **3** | `…-phase3` | **Policy, Review Queue write path, multimodal contracts, UI panel module** |
+
+## Phase 3 details
+
+### Hybrid policy (`HybridPolicyService`)
+Persisted dial for:
+- `blocked_node_types` / `blocked_metadata_flags` (unioned with hard circuit breakers)
+- `auto_commit` (default **false**)
+- `allow_multimodal` (default **false**)
+- `min_extraction_confidence`
+
+API:
+```
+GET  /api/network-boundary/policy
+POST /api/network-boundary/policy
+     { "auto_commit": false, "allow_multimodal": false, "blocked_node_types": ["Secret"] }
+```
+
+### Review Queue write path
+`CloudResponseIngestor` now:
+1. Always enqueues a `change_proposal` review item when a ReviewQueueService is bound
+2. Only auto-writes to the graph when `plan.auto_commit` is true **and** a store write API exists
+
+Users approve cloud-derived memory growth in the existing Review Center.
+
+### Multimodal / video
+`latticeai/services/multimodal_streaming.py`:
+- Same `NetworkBoundaryMode.CLOUD_ALLOWED` gate
+- Additional `allow_multimodal` policy flag
+- Adapter protocol for future Runway/Luma/Veo-compatible providers
+
+### UI
+`NetworkBoundaryPanel` (React) + `GET /api/network-boundary/ui-state`:
+- Toggle local ↔ cloud (with ack)
+- Transparency preview of nodes about to be sent
+- Shows current policy flags
+
+Mounted in `frontend/src/pages/System.tsx` beside `PermissionModePanel`, so it
+appears on the settings tab of every install. The standalone
+`static/app/network-boundary-panel.js` module 10.1.0 shipped was removed in
+10.1.1: nothing mounted it, and it lived in Vite's build output directory,
+which `npm run build:assets` wipes.
+
+## End-to-end flow (Phase 3)
+
+```
+User toggles cloud_allowed (UI panel or API)
+    → optional policy: auto_commit / multimodal / sensitivity
+User sends chat message
+    → /chat resolves mode
+    → if cloud_allowed: minimal KG context + token budget check
+    → cloud stream (SSE)
+    → rich extraction (Decision/Task/Concept)
+    → Review Queue item (change_proposal)
+    → optional auto-commit when policy allows
+```
+
+## Configuration
+
+```bash
+export LATTICEAI_CLOUD_API_KEY=sk-...
+export LATTICEAI_CLOUD_MODEL=gpt-4o-mini
+export LATTICEAI_NETWORK_MODE=local_only
+export LATTICEAI_CLOUD_MAX_TOKENS_PER_TURN=2500
+export LATTICEAI_CLOUD_MAX_TOKENS_PER_SESSION=50000
+```
+
+## Privacy (unchanged core rules)
+
+1. Default `local_only`
+2. Cloud requires explicit ack
+3. Only minimal compact context leaves the host
+4. Hard + user sensitivity filters
+5. Cloud-derived nodes go through Review Queue unless user opts into auto_commit
+6. Multimodal is off unless policy explicitly enables it
+
+## Success criteria (Phase 3)
+
+- User can configure sensitivity and auto-commit without code changes
+- Cloud KG growth appears in Review Center as actionable proposals
+- Multimodal path cannot fire without both cloud mode and policy flag
+- UI panel works without rebuilding the main React bundle

package/docs/ONBOARDING.md

@@ -1,6 +1,6 @@
 # Lattice AI Onboarding
 
-Current release: **9.9.9 — Lean Shell**.
+Current release: **10.2.0 — Load-Bearing Fixes**.
 
 The first-run goal is a five-minute path from "I opened the app" to "my Brain
 has a source, a question, and proof." This page is the product contract behind

package/docs/OPERATIONS.md

@@ -1,4 +1,4 @@
-# Lattice AI — Operations Guide (v9.9.9)
+# Lattice AI — Operations Guide (v10.2.0)
 
 > **Status: canonical** — kept in sync with the current release. Storage layout
 > below reflects the SQLite live Brain store and workspace scoping, not the

package/docs/TRUST_MODEL.md

@@ -1,6 +1,6 @@
 # Lattice AI Trust Model
 
-Current release: **9.9.9 — Lean Shell**.
+Current release: **10.2.0 — Load-Bearing Fixes**.
 
 Lattice AI is local-first, explicit about external communication, and honest
 when a capability is unavailable.

package/docs/WHY_LATTICE.md

@@ -1,6 +1,6 @@
 # Why Lattice AI Exists
 
-Current release: **9.9.9 — Lean Shell**.
+Current release: **10.2.0 — Load-Bearing Fixes**.
 
 **Lattice AI is a local-first Digital Brain that keeps your knowledge durable
 across any AI model.**

package/docs/kg-schema.md

@@ -1,6 +1,6 @@
 # Knowledge Graph Schema
 
-Current release: **9.9.9 — Lean Shell**.
+Current release: **10.2.0 — Load-Bearing Fixes**.
 
 명세 출처: `lattice_ai_full_spec.pptx` 슬라이드 20·21·22
 구현: `lattice_brain/graph/schema.py`

package/lattice_brain/__init__.py

@@ -26,7 +26,7 @@ from .storage import (
     storage_from_env,
 )
 
-__version__ = "9.9.9"
+__version__ = "10.2.0"
 
 __all__ = [
     "AgentRuntime",

package/lattice_brain/archive.py

@@ -16,17 +16,18 @@ import shutil
 import sqlite3
 import tempfile
 import zipfile
+from contextlib import closing
 from dataclasses import dataclass
 from pathlib import Path, PurePosixPath
 from typing import Any, Dict, Iterable, List, Optional
 
+from cryptography.exceptions import InvalidTag
 from cryptography.hazmat.primitives import hashes
 from cryptography.hazmat.primitives.ciphers.aead import AESGCM
-from cryptography.exceptions import InvalidTag
 from cryptography.hazmat.primitives.kdf.pbkdf2 import PBKDF2HMAC
 
-from .utils import sha256_file as _sha256_file, utc_now_iso as _now
-
+from .utils import sha256_file as _sha256_file
+from .utils import utc_now_iso as _now
 
 ARCHIVE_FORMAT = "latticebrain.encrypted"
 ARCHIVE_VERSION = 2
@@ -106,7 +107,7 @@ def _checkpoint_sqlite(db_path: Path) -> None:
     if not db_path.exists():
         return
     try:
-        with sqlite3.connect(str(db_path)) as conn:
+        with closing(sqlite3.connect(str(db_path))) as conn, conn:
             conn.execute("PRAGMA wal_checkpoint(FULL)")
     except sqlite3.Error:
         return

package/lattice_brain/conversations.py

@@ -19,8 +19,9 @@ import json
 import logging
 import sqlite3
 import threading
+from contextlib import contextmanager
 from pathlib import Path
-from typing import Any, Dict, List, Optional
+from typing import Any, Dict, Iterator, List, Optional
 
 
 def _message_hash(item: Dict[str, Any]) -> str:
@@ -40,11 +41,21 @@ class ConversationStore:
         self._lock = threading.RLock()
         self._init_db()
 
-    def _connect(self) -> sqlite3.Connection:
+    @contextmanager
+    def _connect(self) -> Iterator[sqlite3.Connection]:
+        """Transactional connection that is closed when the block exits.
+
+        ``with sqlite3.connect(...)`` commits but never closes; see
+        :meth:`lattice_brain.storage.base.StorageEngine.session`.
+        """
         conn = sqlite3.connect(str(self.db_path))
         conn.row_factory = sqlite3.Row
         conn.execute("PRAGMA journal_mode=WAL")
-        return conn
+        try:
+            with conn:
+                yield conn
+        finally:
+            conn.close()
 
     def _init_db(self) -> None:
         with self._lock, self._connect() as conn:

package/lattice_brain/embeddings.py

@@ -10,7 +10,6 @@ import struct
 from dataclasses import dataclass
 from typing import Any, Dict, Iterable, List, Optional
 
-
 DEFAULT_EMBEDDING_DIM = int(os.getenv("LATTICEAI_VECTOR_DIM", "384"))
 EMBEDDING_MODEL_ID = f"lattice-local-hash-v1:{DEFAULT_EMBEDDING_DIM}"
 
@@ -71,7 +70,18 @@ class LocalEmbeddingModel:
         return [value / norm for value in vector]
 
     def similarity(self, left: Iterable[float], right: Iterable[float]) -> float:
-        return float(sum(a * b for a, b in zip(left, right)))
+        # strict=True: a dimension mismatch means the two vectors came from
+        # different embedding models. Truncating to the shorter one produces a
+        # plausible-looking similarity that is meaningless — exactly the silent
+        # wrongness this codebase keeps finding. Callers that can hit a model
+        # swap already handle failure and fall back to lexical search.
+        left_v, right_v = list(left), list(right)
+        if len(left_v) != len(right_v):
+            raise ValueError(
+                f"embedding dimension mismatch: {len(left_v)} vs {len(right_v)}; "
+                "the vector index was built with a different model"
+            )
+        return float(sum(a * b for a, b in zip(left_v, right_v, strict=True)))
 
     def encode(self, vector: Iterable[float]) -> bytes:
         values = list(vector)