ltcai 8.1.0 → 8.2.0

package/docs/CHANGELOG.md

@@ -6,6 +6,28 @@ existed at that release.
 
 ## [Unreleased]
 
+No unreleased changes.
+
+## [8.2.0] - 2026-06-27
+
+### Added
+- Added an evidence-backed Brain Brief to Brain Home so users can see what to
+  notice, what evidence supports it, and what action to take next.
+- Added `GET /api/memory/brain-brief`, backed by `MemoryService.brain_brief()`,
+  with honest empty-state guidance and recall/graph/model-proof actions.
+- Added unit coverage for Brain Brief service behavior and the memory API route.
+
+### Changed
+- Completed the remaining model loading/runtime extraction into
+  `model_loading.py` and `model_engines.py` while preserving compatibility
+  delegations from `model_runtime.py`.
+- Extracted WorkspaceOS graph trace, agent/workflow run, skill, and snapshot
+  comparison ownership into focused manager modules.
+- Wired Knowledge Graph embedding dimensions from the central resolved `Config`
+  embedder at app startup.
+- Synchronized Python, npm, VS Code extension, Tauri, workspace, readiness,
+  static asset, and current-release documentation versions to 8.2.0.
+
 ## [8.1.0] - 2026-06-27
 
 ### Changed
@@ -260,877 +282,19 @@ Package metadata, Tauri, frontend, Python all aligned to 7.7.0. UI/UX microcopy
 - Package/runtime/static metadata is synchronized to 7.0.0; package publish and
   deployment remain owner-run only.
 
-## [6.7.0] - 2026-06-18
-
-> Brain IA Cleanup. Makes rich product pages reachable from Brain Home,
-> separates visible product IA from legacy compatibility aliases, and
-> code-splits heavy pages out of the first Brain bundle.
-
-### Changed
-- Brain Home now exposes direct action routes for adding documents, searching
-  the knowledge graph, changing models, and opening settings, so the rich
-  Capture, Brain explorer, Library, and System pages are reachable from the
-  first Brain flow.
-- The post-onboarding app shell now mounts Capture, Brain explorer/graph,
-  Library, System, and Act pages inside a shared Brain shell navigation instead
-  of leaving those rich pages as compatibility-only routes.
-- `routes.ts` now keeps product shell routes and compatibility aliases in
-  separate maps while preserving old hash/entry URLs.
-- Rich pages now load through `React.lazy`, splitting Brain explorer, Capture,
-  Library, System, and Act chunks away from the initial Brain Home bundle.
-- Package/runtime/static metadata is synchronized to 6.7.0; package publish and
-  deployment remain owner-run only.
-
-## [6.6.0] - 2026-06-18
-
-> Brain Proof Runtime. Makes the first-five-minute Brain value visible by
-> showing backend-owned evidence that saved context can be recalled and survives
-> model changes.
-
-### Added
-- `/api/memory/brain-proof`, backed by `MemoryService.brain_proof()`, combining
-  Brain readiness, durable memory counts, graph/vector depth, active model id,
-  and a unified recall sample.
-- Brain Home proof strip for recallable context, model-continuity, and
-  knowledge-store state.
-- Recent recall proof card that surfaces the latest memory/item the Brain can
-  bring back after a useful exchange.
-- Brain Home document upload CTA in the empty state and composer, giving the
-  first session a direct way to grow the Brain with files.
-- Refreshed README-linked release evidence under `output/release/v6.6.0/`.
-- Unit coverage for model-independent Brain proof recall behavior.
-- Honest empty-state Brain proof: model-independent capability remains visible,
-  but continuity is only marked proven after durable evidence exists.
-
-### Changed
-- Interaction router runtime context now carries an active-model getter into
-  the Memory router, reducing app-factory inline wiring while preserving route
-  order.
-- Default Brain proof recall seeding is scoped to the current user/workspace
-  before falling back to conversation memory.
-- Chat-to-Knowledge-Graph ingestion now carries workspace scope into projected
-  graph nodes, and Brain proof conversation counts use scoped conversations so
-  another user/workspace cannot make an empty Brain look proven.
-- Direct Knowledge Graph ingest and document upload ingestion now resolve the
-  request workspace and project that scope into graph/source/document nodes.
-- Package/runtime/static metadata is synchronized to 6.6.0; package publish and
-  deployment remain owner-run only.
-
-## [6.5.0] - 2026-06-18
-
-> Brain Experience Readiness. Uses the 6.4.0 quality baseline to improve what
-> normal users can understand: how alive the Brain is, where they are in the
-> memory-to-graph journey, and what changed after a useful exchange is saved.
-
-### Added
-- Brain Home readiness strip that classifies the current Brain as waiting for
-  its first memory, forming topics, or ready for map exploration.
-- Backend-owned Brain readiness summary from `MemoryService`, exposed through
-  `/api/memory/manager` and `/api/memory/brain-quality`, so the UI no longer
-  has to infer Brain growth from local fragments.
-- Persistent depth progress rail across the Living Brain surface so the
-  memory -> topics -> relationships -> graph journey is visible and reversible.
-- Source-aware memory-save detail under chat feedback, making the recall loop
-  explicit after a conversation enriches the Brain.
-- Visual regression assertions for Brain readiness, depth progress, and
-  source-aware memory-save feedback.
-
-### Changed
-- Brain overview empty states now guide first-memory use without exposing graph
-  internals or administrator language.
-- Brain readiness now uses the same memory, graph, relationship, and source
-  health signals that the Memory Manager reports, keeping product UX aligned
-  with backend Brain state.
-- Package/runtime/static metadata is synchronized to 6.5.0; package publish and
-  deployment remain owner-run only.
-
-## [6.4.0] - 2026-06-17
-
-> Digital Brain Quality Hardening. Tightens workspace-scoped Brain retrieval
-> and memory mutation boundaries while adding non-destructive quality primitives
-> for embeddings, retrieval, memory, graph validation, context assembly, and
-> benchmarks.
-
-### Added
-- `lattice_brain.quality` with embedding fallback labelling, drift/re-index
-  planning, BM25 lexical scoring, hybrid fusion, reranker fallback contracts,
-  memory candidate scoring/deduplication/conflict/retention helpers, graph edge
-  confidence/evidence metrics, structured context guardrails, and retrieval
-  benchmark metric calculation.
-- `docs/v6.4/BRAIN_QUALITY_BASELINE.md` documenting the 6.4.0 Digital Brain
-  quality baseline, risk register, validation items, and deferred work.
-- Refreshed README release evidence screenshots and walkthrough GIF under
-  `output/release/v6.4.0/`.
-- Unit coverage for Brain quality primitives and workspace-scoped graph/search
-  and memory-manager mutation boundaries.
-
-### Changed
-- `/api/graph*`, `/knowledge-graph/*`, and hybrid-search service paths now
-  preserve workspace scope across graph, node, neighborhood, relationship,
-  keyword, vector, graph, and hybrid retrieval.
-- Memory Manager prune, compact, and clear operations now intersect requested
-  ids/kinds with the caller's scoped memory set.
-- Memory Manager graph clear is blocked because the existing graph clear path
-  is not workspace-scoped.
-- Package/runtime/static metadata is synchronized to 6.4.0; package publish and
-  deployment remain owner-run only.
-
-## [6.3.1] - 2026-06-17
-
-> Access Runtime / i18n Follow-up. Closes the next app-factory decomposition
-> slice and finishes Capture/Review Center localization coverage on top of the
-> v6.3.0 product-hardening line.
-
-### Changed
-- Moved app-factory access-control helper closures into
-  `latticeai.runtime.access_runtime` while preserving role lookup, token
-  extraction, user enforcement, admin enforcement, and public user projection
-  call signatures.
-- Capture tabs, headings, placeholders, action labels, upload details, folder
-  connection copy, browser capture copy, and pipeline controls now use the
-  shared Korean/English i18n map.
-- Review Center inbox/card copy, status badges, empty states, Run Now feedback,
-  snooze/unsnooze labels, and action aria labels now use the shared i18n map.
-- README release evidence screenshots and walkthrough GIF are refreshed under
-  `output/release/v6.3.1/`.
-- Package/runtime/static metadata is synchronized to 6.3.1; package publish and
-  deployment remain owner-run only.
-
-### Added
-- Focused unit coverage for the extracted access runtime auth/admin behavior.
-
-## [6.3.0] - 2026-06-16
-
-> Product Hardening Completion. Completes the 6.2 decomposition follow-up by
-> tightening Brain archive, provenance, ingestion, Review Center, local model
-> runtime, app-factory wiring, compatibility shim, i18n, and release-validation
-> surfaces.
-
-### Changed
-- Brain Home and onboarding screens are split into finer feature-owned modules
-  while preserving the existing routes, CSS surface, and visual behavior.
-- Platform and automation runtime wiring moved behind a dedicated app-factory
-  seam, keeping construction side-effect free and route order unchanged.
-- Admin, Brain, and onboarding localized aria/placeholder surfaces now use the
-  shared i18n map.
-- Brain archive care now validates archive paths, confirms passphrases, gives
-  strength guidance, and summarizes inspect/restore-preview results without
-  dumping raw JSON.
-- Document capture now exposes a drop/choose ingest queue with progress,
-  failure reason, retry, source metadata, and post-ingest Brain/Graph
-  confirmation.
-- Brain memory and graph surfaces now show provenance/source type, created time,
-  source path/title/conversation fallback, and copy-source affordances.
-- Review Center cards now keep visible action result/error feedback for
-  run-now/approve/dismiss/snooze flows while preserving run-now as
-  preview/regenerate rather than approval.
-- Local model setup now reports loaded model, engine state, cache/storage path,
-  download/load progress, reload/unload controls, and an honest no-model state.
-- App factory model/chat/review-tail wiring moved into focused runtime wiring
-  modules while preserving route order.
-- Review Center run-now backend wiring moved into `latticeai.runtime.review_wiring`
-  and is covered by scoped runner/status-preservation tests.
-- Legacy root shim import smoke now covers server, knowledge graph, CLI,
-  Telegram, and P-Reinforce compatibility paths.
-- Package/runtime/static metadata is synchronized to 6.3.0; package publish and
-  deployment remain owner-run only.
-
-### Added
-- Frontend i18n literal check for localized aria/placeholder props in Brain,
-  Admin, and onboarding TSX surfaces.
-- Focused runtime wiring modules for model runtime, chat/interaction contexts,
-  and review/Brain tail registration.
-
-## [6.2.0] - 2026-06-16
-
-> Product Decomposition / Release Smoke Automation. Splits the largest product
-> surfaces into feature modules, strengthens consent/i18n coverage, shrinks
-> legacy root modules, and automates release smoke validation.
-
-### Added
-- Feature-owned Brain Home and Admin Console modules behind the existing app
-  shell route surface.
-- Onboarding screen components for ProductFlow, backed by design-system CSS
-  instead of inline style layout.
-- Model download consent details for size, storage location, external target,
-  and a "do later" path.
-- Typed tool and interaction router context objects for app-factory route
-  registration without changing route order.
-- `npm run release:smoke` covering wheel install, npm tgz contents, static
-  assets, and Tauri artifact checks.
-
-### Changed
-- Historical root modules for CLI, Telegram, and P-Reinforce now delegate to
-  package modules while preserving import and script compatibility.
-- Admin Console and onboarding user-facing copy now route through shared i18n
-  keys for Korean/English coverage.
-- Package/runtime/static metadata is synchronized to 6.2.0; package publish
-  and deployment remain owner-run only.
-- README release evidence screenshots and walkthrough GIF are refreshed under
-  `output/release/v6.2.0/`.
-
-### Preserved
-- v6.1 screenshots and validation reports remain historical evidence.
-- Package publish and deploy commands continue to require exact artifact names.
-
-## [6.1.0] - 2026-06-16
-
-> Product Hardening / Digital Brain Completion. Tightens the local-first Brain
-> flow, Brain Core boundary, backend trust gates, and release documentation
-> without adding broad new product surfaces.
-
-### Added
-- First-run path to open the Brain without installing or loading a model first.
-- Brain Home empty-state loop that shows first memory save, Brain state return,
-  and backup ownership.
-- AST import guard proving `lattice_brain` does not import `latticeai` or
-  `ltcai`.
-- Unit coverage for model download consent blocking in `model_runtime`.
-- `latticeai.cli.runtime` for pure CLI runtime helpers while preserving the root
-  `ltcai_cli.py` entrypoint.
-- v6.1 baseline, frontend UX, backend hardening, and instruction documents under
-  `docs/v6.1/`.
-
-### Changed
-- Review Center cards clarify that Run now previews/executes without approval.
-- Static app assets are refreshed for the v6.1 frontend hardening changes.
-- Tool dispatch authorization now uses an injectable service boundary around
-  the shared ToolRegistry while preserving the legacy module-level API.
-- Chat agent runtime construction moved out of `create_chat_router` for
-  production app assembly and is now injected through `AppContext`.
-- README release evidence screenshots and walkthrough GIF are refreshed under
-  `output/release/v6.1.0/`, including the Review Center surface.
-- README and release notes described v6.1.0 as the then-current hardening target
-  with exact expected artifact filenames.
-
-### Preserved
-- Package publish, deployment, tag creation, and main merge remain out of scope
-  unless explicitly requested.
-- Package/runtime version metadata is synchronized to 6.1.0; package publish,
-  deployment, tag creation, and main merge remain owner-run only.
-
-## [6.0.0] - 2026-06-15
-
-> Product Reset / Review Center Completion. Raises the Review Center from a
-> pending-only inbox into a reversible automation review surface while
-> documenting the v6 quality uplift honestly.
-
-### Added
-- Review Center status filters for Pending, Snoozed, and All.
-- Explicit `POST /automation/reviews/{item_id}/unsnooze` API and backend policy.
-- Frontend Unsnooze action and clear `snoozed_until` presentation.
-- `docs/v6/PLAN.md`, `ARCHITECTURE_REVIEW.md`, `UX_REVIEW.md`, and
-  `QUALITY_SCORECARD.md`.
-
-### Changed
-- Review Center frontend moved from `Act.tsx` into
-  `frontend/src/features/review/` components and helpers.
-- Review item frontend types now alias generated OpenAPI component schemas.
-- Review Center API calls now use generated OpenAPI operation paths for list,
-  approve, dismiss, snooze, unsnooze, and run_now actions.
-- `app_factory.py` runtime assembly moved behind session, hooks, web,
-  persistence, lifespan, automation, context/search, platform service, app
-  context, and router-registration seams while preserving the frozen 364-entry
-  route/mount snapshot.
-- First-run/onboarding copy now states local-first trust boundaries more
-  directly: local knowledge by default, explicit downloads, and explicit
-  external transfer.
-- README release evidence screenshots and walkthrough GIF are refreshed under
-  `output/release/v6.0.0/`, including the Review Center surface.
-- OpenAPI artifacts and synchronized package/runtime/static metadata now target
-  `6.0.0`.
-
-### Preserved
-- `run_now` remains preview/regenerate and does not approve.
-- Snooze expiry remains read-time only; explicit unsnooze is the only immediate
-  return-to-pending mutation.
-- Package publishing, GitHub Release creation, artifact upload, and merge to
-  `main` remain out of scope for this branch.
-
-## [5.6.0] - 2026-06-15
-
-> Brain Automation Review Center. Adds a workspace-scoped review inbox for
-> automation output so scheduled runs and Brain-event suggestions land as
-> inspectable items before the user approves, dismisses, snoozes, or reruns
-> them.
-
-### Added
-- Backend Review Queue service, store persistence, and `/automation/reviews`
-  API with explicit `source`, `status`, `effective_status`, payload, and
-  provenance fields.
-- Review item actions: approve, dismiss, snooze, and run_now. `run_now` is a
-  preview/regenerate action and does not mark the item approved.
-- Optional `review_queue: true` opt-in path for TriggerService and RunExecutor
-  to enqueue review items without changing legacy scheduler behavior.
-- Act page Review inbox under Runs with source filtering, pending review cards,
-  provenance details, and guarded actions.
-
-### Changed
-- Bumped synchronized Python, npm, VSIX, Tauri, runtime constants, lockfiles,
-  OpenAPI artifacts, and static metadata to `5.6.0`.
-- Updated release documentation and artifact names for exact 5.6.0 release
-  preparation.
-
-### Preserved
-- Snoozed items stay hidden until their `snoozed_until` time expires; expiry is
-  interpreted at read time through `effective_status` without scheduler mutation.
-- Existing v5.4/v5.5 automation and workflow behavior remains compatible unless
-  a workflow explicitly opts into review queue creation.
-- Package registry publishing and deployment remain owner-run only.
-
-## [5.5.0] - 2026-06-15
-
-> Release Coordination. Synchronized package/runtime/static metadata and release
-> documentation for the 5.5.0 line while preserving v5.4.0 Brain Automation
-> Scheduler behavior.
-
-### Changed
-- Bumped synchronized Python, npm, VSIX, Tauri, runtime constants, lockfiles, and
-  static asset manifest metadata to `5.5.0`.
-- Updated README, RELEASE.md, RELEASE_NOTES.md, FEATURE_STATUS.md,
-  vscode-extension/README.md, and this changelog so current-release references
-  and expected artifact names point at exact 5.5.0 filenames.
-
-### Preserved
-- v5.4.0 consent-first Brain automation recipes, TriggerService dedup,
-  LATTICE_TZ, degraded status, enabled:false disarming, runtime graph cleanup,
-  and E2E scenario coverage remain the functional baseline.
-- Package registry publishing and deployment remain owner-run only.
-
-## [5.4.0] - 2026-06-15
-
-> Brain Automation Scheduler. Consent-first recipe drafts (Daily Memory Digest,
-> Weekly Project Review, Follow-up Radar) install as disabled workflows.
-> Scheduler triggers (TriggerService) with dedup, LATTICE_TZ, degraded status,
-> and runtime graph cleanup.
-
-### Added
-- Consent-first Brain automation recipe drafts for Daily Memory Digest, Weekly
-  Project Review, and Follow-up Radar. Recipes install as disabled workflows so
-  users can inspect them before any scheduler or Brain-event trigger fires.
-
-### Changed
-- Workflow trigger scanning now treats explicit `enabled: false` trigger config
-  as disarmed while preserving legacy behavior for existing workflows that do
-  not include an `enabled` field.
-- The Automate page now surfaces Brain automation recipe cards with local-only,
-  review-before-run consent copy.
-- Recipe install UX: success feedback on "Create reviewable draft", install-time
-  button disabled to block duplicates clicks, and if same recipe draft already
-  exists (by metadata), button state changes + guide text instead of re-creating.
-- lattice_brain/runtime dependency/responsibility graph 정리 + 실제 진입점 매핑 문서화 (runtime/* 모듈 헤더 + app_factory 주석). AgentRuntime (lattice_brain facade) vs latticeai/core/agent (single-agent state/plan/transcript) 분리 명확화.
-- TriggerService 스케줄러 엣지케이스 보강: LATTICE_TZ env 지원 (describe() 노출), last_attempt_at + cooldown dedup 가드로 중복 실행 방지 (interval/brain_event), consecutive_failures + "degraded" status per-trigger 서페이싱.
-- A방향 E2E 시나리오 초안 brain_automation.py 에 작성 (draft install, dedup, consent-first, trigger fire with provenance, LATTICE_TZ, degraded, review flow, RunExecutor 경로).
-
-## Unreleased
-
-## [5.3.0] - 2026-06-14
-
-> Product Clarity and Runtime Cleanup. Lattice AI is now presented consistently
-> as a local-first Digital Brain that keeps user knowledge durable across any AI
-> model, while the first runtime extraction seams move out of `app_factory.py`.
-
-### Added
-- `docs/DEVELOPMENT.md` for contributor setup, validation gates, runtime
-  assembly expectations, and documentation sync rules.
-- `docs/LEGACY_COMPATIBILITY.md` to explain root compatibility modules, their
-  current homes, why they remain packaged, and the safe removal checklist.
-- v5.3.0 release evidence index and copied screenshot/GIF paths under
-  `output/release/v5.3.0/`.
-
-### Changed
-- Reorganized README around product identity, user need, user actions,
-  one-minute flow, screenshots, installation, architecture, and release
-  preparation. Release artifacts no longer interrupt the first screen.
-- Unified public descriptions in README, package metadata, pyproject, feature
-  status, architecture, release notes, and VS Code extension docs.
-- Improved onboarding and model setup copy around local ownership, explicit
-  consent, model-as-voice, Brain-as-asset, and what happens before download/load.
-- Simplified Basic model setup to a shorter recommendation set while Advanced
-  keeps hardware, verification, load strategy, license, and safety detail.
-- Moved config, security, and Brain runtime builders from `latticeai.app_factory`
-  into `latticeai.runtime` modules while preserving lazy import compatibility.
-- Bumped synchronized package/runtime/static versions to `5.3.0`.
-
-### Preserved
-- No package registry publishing automation.
-- Local-first defaults and explicit consent gates for downloads, cloud calls,
-  Telegram, Brain Network, Docker/Postgres setup, and update checks.
-- `server_app.__getattr__` lazy compatibility and app factory import side-effect
-  boundaries.
-
-## [5.2.0] - 2026-06-14
-
-> Lattice AI 5.2.0 — Aggressive User-Focused Model Transformation. Transparent
-> structured capability registry, automated HF verification, modern multimodal
-> additions, download/load strategy exposure, hardware notes, verified status,
-> and updated UI/backend flows so users see exactly what they get before consent.
-
-### Added
-- Structured `ModelCapabilityRegistry` (latticeai/services/model_capability_registry.py) with dataclass fields for provider/hf_repo_id, modality, quantization, download_strategy, load_strategy, hardware (min/recommended RAM, Apple/CUDA prefs, notes), license, safety_notes, and rich VerificationStatus (hf_exists, has_config, has_tokenizer, pipeline, last_checked, notes).
-- Modern multimodal candidates (Gemma 3 4B/12B, Qwen2.5-VL-7B, Llama-3.2-11B-Vision, Pixtral-12B) in the structured registry for HF verification transparency, alongside the user-facing Gemma-4 / Qwen3-VL / Llama-4 load-ready family.
-- Automated verification script: `scripts/verify_hf_model_registry.py` (lightweight HF API + restricted snapshot for config/tokenizer; optional --test-load for small models; explicit LARGE_MODEL notes; writes verification_report.json).
-- Registry info exposed via `/models` and `/models/recommendations` (registry.verified_count, verification dicts, hardware, strategies).
-- New unit tests: `tests/unit/test_model_capability_registry.py` (5 tests covering registry, legacy shape, rec payload, report, roundtrips).
-
-### Changed
-- `model_catalog.py` now sources ENGINE_MODEL_CATALOG + aliases from the capability registry (single source of truth), preserves legacy shapes + reexports, and finalizes the user-facing catalog to current load-ready families so lower-generation or non-load-verified candidates do not become noisy primary choices.
-- `model_recommendation.py` `_classify_one` now forwards 5.2 fields (hf_repo, verification, hardware, load_strategy, license, safety, recommended_default).
-- `verification.verified` and verified-model API lists now require HF presence plus config and tokenizer hints, with weights-hint detail exposed separately, so the UI badge matches the actual load-readiness contract.
-- Marketplace template installs now keep registry entries scoped per workspace, and `/marketplace/templates/registry` returns only the authorized workspace scope.
-- SQLite restore now checkpoints WAL state before taking the pre-restore backup, keeping failed blob restores rollback-safe on Linux/Python 3.12 CI and local builds.
-- Backend model APIs return rich fields; frontend can render verified badges, modality, hardware notes, strategies.
-- Library.tsx (ModelsPanel): added "multimodal" + "✓ HF" verified badges, recommended_default support, hardware notes line, load_strategy plus license/safety notes in advanced detail, updated guided setup copy for transparency and consent.
-- All registry HF ids confirmed present via HF API on 2026-06-14; 15/16 expose config/tokenizer hints, Pixtral remains available-but-not-local-load-verified, and large models are flagged with explicit limitations.
-- Version bumped to 5.2.0 everywhere (pyproject, __init__, package.json, vscode-extension).
-
-### Preserved
-- Exact public API shapes, recommendation tri-state logic, engine aliases, family de-dup, download consent gates, no silent downloads.
-- Historical changelog entries for 5.1.0 and prior.
-
-## [5.1.0] - 2026-06-14
-
-> Product Trust & Clarity Release. v5.1.0 clarifies Lattice AI as a
-> local-first private AI memory layer / Digital Brain, then adds security,
-> privacy, honesty, and architecture gates so the product does not overclaim.
-
-### Changed
-
-- Rewrote the README first screen around the positioning lines
-  `Your private AI memory layer. Keep your knowledge. Switch any model.` and
-  `모델은 바꿔도, 내 지식은 남는 로컬 AI 브레인.`
-- Added practical use cases for preserving project decisions, switching models,
-  connecting documents/conversations/files/notes/decisions, encrypted Brain
-  archive portability, cloud lock-in avoidance, and honest no-model states.
-- Added and refreshed trust documentation: `PRIVACY.md`, `docs/WHY_LATTICE.md`,
-  `docs/TRUST_MODEL.md`, `SECURITY.md`, `ARCHITECTURE.md`, and
-  `FEATURE_STATUS.md`.
-- Removed `csp:null` from Tauri production config and added an app-shell CSP
-  response header.
-- Centralized secret redaction for logs, audit payloads, security exports, and
-  builtin hook packets.
-- Changed chat auto-file handling so `LATTICEAI_AUTO_READ_CHAT_PATHS` remains
-  off by default and does not silently read arbitrary local paths even if
-  enabled.
-- Added explicit `allow_download=true` consent for model download requests.
-- Added config, security, and Brain runtime builder seams in `app_factory.py`
-  while preserving the existing API shape.
-- Hardened release artifact cleanup so `release:artifacts` removes stale
-  `dist/ltcai-*` and root `ltcai-*.tgz` files before rebuilding exact v5.1.0
-  artifacts only.
-- Made `npm run test:integration` self-contained by starting a local uvicorn
-  server, waiting for `/health`, running the integration suite, and shutting the
-  server down.
-- Fixed a SQLite Brain restore TOCTOU race where transient `-wal` / `-shm`
-  siblings could disappear between probe and copy during archive restore.
-- Bumped synchronized package/runtime/static versions to `5.1.0`, including
-  Python metadata, npm package metadata, VSIX metadata, Tauri metadata,
-  `latticeai`, `lattice_brain`, runtime constants, and static asset metadata.
-
-### Added
-
-- v5.1 trust validation tests for CSP, secret redaction, audit redaction,
-  auto-file-read blocking, public/network auth posture, `shell=True`
-  production-path scanning, and Brain Core import isolation.
-- A deterministic regression test for restore-time WAL sibling disappearance.
-- v5.1 release evidence paths under `output/release/v5.1.0`.
-
-### Preserved
-
-- Tracked release-note history now starts at v4.5.0; older release-note files
-  are hidden from the Git tree.
-- External package publishing remains owner-run; this release prepares exact
-  artifacts and GitHub Release assets without registry publish automation.
-
-### Artifacts
-
-- `dist/ltcai-5.1.0-py3-none-any.whl`
-- `dist/ltcai-5.1.0.tar.gz`
-- `dist/ltcai-5.1.0.vsix`
-- `ltcai-5.1.0.tgz`
-- `src-tauri/target/release/bundle/dmg/Lattice AI_5.1.0_aarch64.dmg`
-
-## [5.0.0] - 2026-06-14
-
-> Multilingual Brain Foundation Release. v5.0.0 starts the major-version cleanup
-> line by preserving the existing AgentRuntime, ToolRegistry, Brain Core, Admin
-> Console, and graph foundations while making the product usable in Korean or
-> English from first launch through Brain exploration.
-
-### Changed
-
-- Added a persisted `lattice.language` preference with Korean and English
-  choices available on first-run onboarding, the Brain home, and the separated
-  Admin Console header.
-- Localized first-run login, environment analysis, model recommendation,
-  install/download/load status, Brain quick views, starter prompts, memory save
-  feedback, overview panels, and graph focus fallback copy.
-- Updated visual tests so the Korean path is explicitly selected before running
-  existing first-run and Brain depth assertions.
-- Bumped synchronized package/runtime/static versions to `5.0.0`, including
-  Python metadata, npm package metadata, VSIX metadata, Tauri metadata,
-  `latticeai`, `lattice_brain`, runtime constants, and static asset metadata.
-- Captured collaboration guidance from pts_claudecode and pts_grok: the next
-  technical refactor order is config centralization, KG stabilization,
-  ToolRegistry characterization, AgentRuntime extraction, then app factory
-  decomposition.
-
-### Preserved
-
-- Tracked release-note history remains visible from v4.5.0 through v5.1.0.
-- External package publishing remains owner-run; this release prepares exact
-  artifacts and GitHub Release assets without registry publish automation.
-
-### Artifacts
-
-- `dist/ltcai-5.0.0-py3-none-any.whl`
-- `dist/ltcai-5.0.0.tar.gz`
-- `dist/ltcai-5.0.0.vsix`
-- `ltcai-5.0.0.tgz`
-- `src-tauri/target/release/bundle/dmg/Lattice AI_5.0.0_aarch64.dmg`
-
-## [4.7.2] - 2026-06-14
-
-> Intuitive Brain UX Release. v4.7.2 makes the Living Brain easier for
-> non-technical users: first-run login is safer, recommended model setup is
-> one-click, Brain memory/topic/relationship/graph views are directly visible,
-> and conversation shows saved-to-memory feedback while Admin remains separate.
-
-### Changed
-
-- Prevented saved-user email mismatch or wrong saved-user password from silently
-  creating a new empty Brain during first-run login.
-- Added a primary `추천대로 시작하기` model recommendation path and clearer large
-  model download messaging without fake ETA.
-- Added visible `기억 보기`, `주제 보기`, `관계 보기`, and `그래프로 보기` actions
-  on the Brain surface so users can open the desired Brain depth directly.
-- Added a Brain overview panel with recent memories, older memories, major
-  topics, and saved-to-memory feedback after conversation.
-- Updated first-run and empty-Brain copy toward plain user language while
-  keeping the graph as the deepest advanced layer.
-- Bumped synchronized package/runtime/static versions to `4.7.2`, including
-  Python metadata, npm package metadata, VSIX metadata, Tauri metadata,
-  `latticeai`, `lattice_brain`, runtime constants, and static asset metadata.
-- Refreshed README, release notes, architecture, feature status, security,
-  recovery notes, VS Code extension docs, release report, and release evidence
-  paths for v4.7.2.
-
-### Preserved
-
-- Tracked release-note history remains visible from v4.5.0 through v5.1.0.
-- External package publishing remains owner-run; this release prepares exact
-  artifacts and GitHub Release assets without registry publish automation.
-
-### Artifacts
-
-- `dist/ltcai-4.7.2-py3-none-any.whl`
-- `dist/ltcai-4.7.2.tar.gz`
-- `dist/ltcai-4.7.2.vsix`
-- `ltcai-4.7.2.tgz`
-- `src-tauri/target/release/bundle/dmg/Lattice AI_4.7.2_aarch64.dmg`
-
-## [4.7.1] - 2026-06-14
-
-> Admin Operations Release. v4.7.1 extends the separated Admin Console with
-> role permission visibility, audit search/severity filters, local log retention
-> posture, and a dedicated Admin Console data-loading boundary while keeping the
-> user Brain surface simple.
-
-### Changed
-
-- Added role permission visibility to the Admin Console so operators can inspect
-  role member counts and capability summaries without entering the user Brain
-  surface.
-- Added server-backed audit filtering for search text, actor, action, severity,
-  and limit on `/admin/audit`.
-- Added `/admin/log-retention` to report local retention days, retained events,
-  prune candidates, and export-before-prune status without destructive pruning.
-- Split Admin Console data loading into a dedicated frontend hook so admin
-  observability state stays separate from Brain chat state.
-- Updated Admin Console visual mock data for filtered audit and retention
-  coverage.
-- Bumped synchronized package/runtime/static versions to `4.7.1`, including
-  Python metadata, npm package metadata, VSIX metadata, Tauri metadata,
-  `latticeai`, `lattice_brain`, runtime constants, and static asset metadata.
-- Refreshed README, release notes, architecture, feature status, security,
-  VS Code extension docs, release report, and release evidence paths for v4.7.1.
-
-### Preserved
-
-- Tracked release-note history remains visible from v4.5.0 through v5.1.0.
-- External package publishing remains owner-run; this release prepares exact
-  artifacts and GitHub Release assets without registry publish automation.
-
-### Artifacts
-
-- `dist/ltcai-4.7.1-py3-none-any.whl`
-- `dist/ltcai-4.7.1.tar.gz`
-- `dist/ltcai-4.7.1.vsix`
-- `ltcai-4.7.1.tgz`
-- `src-tauri/target/release/bundle/dmg/Lattice AI_4.7.1_aarch64.dmg`
-
-## [4.7.0] - 2026-06-14
-
-> Admin Separation Release. v4.7.0 keeps the Living Brain as the simple user
-> surface and moves users, logs, security events, policies, and Brain operations
-> into a dedicated Admin Console with synchronized release metadata and
-> publishable artifacts.
-
-### Changed
-
-- Added a separate `#/admin` Admin Console so the normal `/app` user experience
-  remains Brain + conversation instead of becoming an admin dashboard.
-- Added admin overview metrics, user directory, audit log rows, security event
-  rows, policy chips, and Brain index rebuild controls to the admin-only
-  surface.
-- Scoped admin history, audit, stats, and sensitivity endpoints by
-  `X-Workspace-Id` / `workspace_id` when a workspace is selected, while keeping
-  legacy global records visible in Personal workspace compatibility mode.
-- Added frontend API helpers for `/admin/stats` and `/admin/security/events`,
-  reusing the existing FastAPI admin/security backend rather than inventing a
-  parallel logging store.
-- Updated visual validation so the Admin Console route is checked separately
-  from the user Brain surface.
-- Bumped synchronized package/runtime/static versions to `4.7.0`, including
-  Python metadata, npm package metadata, VSIX metadata, Tauri metadata,
-  `latticeai`, `lattice_brain`, runtime constants, and static asset metadata.
-- Refreshed release screenshots/GIF paths and docs for v4.7.0, including Admin
-  Console evidence.
-- Updated README, RELEASE.md, ARCHITECTURE.md, SECURITY.md, FEATURE_STATUS.md,
-  VS Code extension docs, release notes, and release report for the current
-  v4.7.0 release line.
-
-### Preserved
-
-- Tracked release-note history remains visible from v4.5.0 through v5.1.0.
-- Local-first ownership, `.latticebrain` portability, rollback-safe restore,
-  and the deepest-layer Knowledge Graph behavior are preserved.
-
-### Artifacts
-
-- `dist/ltcai-4.7.0-py3-none-any.whl`
-- `dist/ltcai-4.7.0.tar.gz`
-- `dist/ltcai-4.7.0.vsix`
-- `ltcai-4.7.0.tgz`
-- `src-tauri/target/release/bundle/dmg/Lattice AI_4.7.0_aarch64.dmg`
-
-## [4.6.1] - 2026-06-14
-
-> Living Brain release refresh and publishable version bump after the v4.6.0
-> PyPI immutability block. v4.6.1 keeps the Living Brain implementation and
-> synchronizes release metadata, README evidence, architecture docs, artifacts,
-> and owner publishing commands to exact `4.6.1` filenames.
-
-### Changed
-
-- Added a Brain-first product direction review and synchronized philosophy /
-  Knowledge Graph docs so the graph is documented as infrastructure inside the
-  Brain rather than the product center.
-- Added a compact "Care for my Brain" ownership panel to the Brain home
-  experience, surfacing local-first/private/portable guarantees plus existing
-  export, backup, encrypted archive, inspect, and restore dry-run actions
-  without making the Knowledge Graph the product center.
-- Refined the Brain home so conversation remains the primary surface and the
-  "Care for my Brain" ownership controls open only when requested, with action
-  results now reflecting the last completed ownership action.
-- Strengthened first-run and empty-Brain product messaging around the core
-  promise that models are replaceable while the user's knowledge, decisions,
-  projects, and context are durable.
-- Hardened Knowledge Graph backup restore and encrypted `.latticebrain` restore
-  so DB/blob replacement is staged with pre-restore backups and rollback on
-  partial failure.
-- Bumped synchronized package/runtime/static versions to `4.6.1`, including
-  Python metadata, npm package metadata, VSIX metadata, Tauri metadata,
-  `latticeai`, `lattice_brain`, runtime constants, and static asset metadata.
-- Refreshed README around the current Living Brain flow: Login -> Environment
-  Analysis -> Recommended Models -> Install & Load -> Brain Chat.
-- Replaced stale README screenshot/GIF references with fresh v4.6.1 release
-  evidence for Login, setup, Brain Chat, Living Brain, Memory Layer, Knowledge
-  Layer, Relationship Layer, and Knowledge Graph.
-- Updated `ARCHITECTURE.md` for the current Tauri shell, React/Vite frontend,
-  FastAPI localhost API, independent `lattice_brain` Brain Core, StorageEngine,
-  SQLite default, PostgreSQL/pgvector opt-in, backup/restore, and
-  `.latticebrain` portability architecture.
-- Added `RELEASE_NOTES_v4.6.1.md` and the v4.6.1 release refresh report.
-
-### Preserved
-
-- Tracked release-note history remains visible from v4.5.0 through v5.1.0.
-  Older hidden notes were not rewritten as v4.6.1 claims.
-- No backend architecture redesign, package publishing, service deployment, or
-  registry upload is part of this refresh.
-
-### Artifacts
-
-- `dist/ltcai-4.6.1-py3-none-any.whl`
-- `dist/ltcai-4.6.1.tar.gz`
-- `dist/ltcai-4.6.1.vsix`
-- `ltcai-4.6.1.tgz`
-- `src-tauri/target/release/bundle/dmg/Lattice AI_4.6.1_aarch64.dmg`
-
-## [4.6.x Brain Exploration Update] - 2026-06-14
-
-> The Brain is now the true interactive entry point to the user's knowledge.
-> Clicking the living Brain progressively reveals deeper layers of the mind:
-> Level 1 living presence -> memories -> concepts -> relationships -> the
-> emergent full Knowledge Graph. The graph never appears abruptly; it grows out
-> of the Brain as the user travels inward.
-
-### Changed
-- Rebuilt `LivingBrain` as a recognizable anatomical Brain with animated
-  hemispheres, folds, memory ripples, thought particles, and state-specific
-  responses for listening, recalling, thinking, planning, and acting.
-- BrainHome now manages five progressive depths. Level 1 shows only the living
-  Brain; Level 2 reveals memory fragments; Level 3 reveals concepts; Level 4
-  reveals relationship threads; Level 5 reveals the searchable Knowledge Graph.
-- Layered emergence uses the existing memory and graph APIs but presents them
-  as unfolding layers inside one central Brain experience rather than separate
-  pages, tabs, or a bottom depth bar.
-- Level 5 now includes graph nodes, edges, search filtering, and focus details
-  in an emergent graph panel.
-- Conversation remains available at every depth; the Brain presence changes
-  intensity as the user travels inward or streams a response.
-- Updated visual validation to assert the new depth contract, graph emergence,
-  graph search, chat streaming, mobile overflow, and legacy entry compatibility.
-
-The user should feel they are moving inward through their own mind, with the graph as the deepest, most structured layer that grew from the living presence.
-
-### Changed (Frontend — full replacement of the previous prototype)
-- The entire post-onboarding surface is now a single immersive "Brain Space". The Brain presence is large, always visible, and the primary emotional and visual object.
-- Conversation is the central, intimate way you live with the Brain. No traditional app chrome, dock, or page shell in the primary experience.
-- LivingBrain component completely rewritten as a reactive, layered, breathing organism with memory ripples, thought particles, state-driven animation, and click-to-respond behavior.
-- Progressive discovery enforced in the UI: from the Brain you gently descend into Memory, Knowledge, Connections, and The Map (the graph). The graph is never the landing experience.
-- Onboarding (ProductFlow) reframed as a quiet, ceremonial awakening ritual with the Brain presence participating at every step.
-- New warm, private, organic visual language (deep ember-gold presence, soft teal memory pulses, generous space, contemplative typography). Old dashboard aesthetics and navigation chrome suppressed on the primary path.
-- Depth chambers are full-bleed, slow, human rooms — not feature pages. A small living trace of the Brain stays visible for continuity.
-- All backend functionality (chat streaming, memory, hybrid search, graph, model prepare/load, portability, agents, etc.) is preserved and reachable. The old navigation remains for deep-link compatibility but is not the product surface.
-
-The technology was already a Digital Brain. This release makes the *experience* one.
-
-## [4.6.0] - 2026-06-13
-
-> Living Brain Experience release after the v4.5.1 product shell reset. (Previous iteration — this RC supersedes the visual and structural approach while keeping the capability foundation.)
-
-### Changed
-
-- Added the required first-launch product flow: Login -> Environment Analysis
-  -> Recommended Models -> Install & Load -> Brain.
-- Made Brain plus conversation the post-model-load `/app` and `/app#/brain`
-  experience.
-- Added an animated living Brain presence that reacts to listening, memory
-  recall, streaming/thinking, and agent/workflow activity.
-- Centralized chat streaming, model status, image attachment, conversation
-  history, and memory previews in a reusable Brain conversation component.
-- Reduced visible primary navigation to Brain, Memory, Files, Automations,
-  Models, and Settings;
-  `/ask` and `/chat` remain route-compatible aliases into Brain.
-- Reordered Brain layers as Brain -> Memories -> Knowledge -> Relationships ->
-  Graph, moving graph exploration to the deepest intentional layer.
-- Updated visual palette and copy so the product reads as a living digital
-  Brain rather than a graph tool or dashboard.
-- Removed the first-run dashboard/setup-card panel from the app shell; setup is
-  now a full-screen product sequence before the Brain opens.
-
-### Preserved
-
-- Brain Core, FastAPI APIs, Tauri shell, StorageEngine, backup/restore,
-  portability, model runtimes, graph/search/chat/capture/automation/system
-  workflows, and route aliases remain capability-compatible.
-
-### Validation
-
-- v4.6.0 validation scope is tracked in
-  `docs/V4_6_0_LIVING_BRAIN_EXPERIENCE_REPORT.md`.
-
-### Artifacts
-
-- `dist/ltcai-4.6.0-py3-none-any.whl`
-- `dist/ltcai-4.6.0.tar.gz`
-- `dist/ltcai-4.6.0.vsix`
-- `ltcai-4.6.0.tgz`
-- `src-tauri/target/release/bundle/dmg/Lattice AI_4.6.0_aarch64.dmg`
-
-## [4.5.1] - 2026-06-13
-
-> Product Reimagining release candidate after the v4.5.0 capability recovery.
-
-### Changed
-
-- Replaced the desktop shell with a compact premium chrome, ambient brain
-  canvas, command palette, responsive mobile drawer, and six-room product model:
-  Home, Ask, Add, Automate, Library, Care.
-- Rewrote first-run onboarding as a non-technical journey: Make it yours ->
-  Choose a space -> Meet your Mac -> Pick a brain -> Install locally -> Try a
-  question -> Set the pace -> Explore memory.
-- Retained all compatibility hash routes while changing visible navigation,
-  information hierarchy, route labels, and page hero language.
-- Replaced the global visual system with a calmer Digital Brain palette,
-  fixed-size responsive typography, 8px-or-smaller card radii, and refined
-  shared controls.
-
-### Preserved
-
-- Brain Core, FastAPI APIs, Tauri shell, StorageEngine, backup/restore,
-  portability, model runtimes, graph/search/chat/capture/automation/system
-  workflows, and route aliases remain capability-compatible.
-
-### Validation
-
-- Full v4.5.1 RC validation is tracked in
-  `docs/V4_5_1_VALIDATION_REPORT.md`.
-
-### Artifacts
-
-- `dist/ltcai-4.5.1-py3-none-any.whl`
-- `dist/ltcai-4.5.1.tar.gz`
-- `dist/ltcai-4.5.1.vsix`
-- `ltcai-4.5.1.tgz`
-- `src-tauri/target/release/bundle/dmg/Lattice AI_4.5.1_aarch64.dmg`
-
-## [4.5.0] - 2026-06-13
-
-> Product Experience Recovery release candidate after the v4.4.0 physical Brain
-> extraction.
-
-### Changed
-
-- Restored the first-run product journey in the desktop shell: Login ->
-  Workspace Selection -> Environment Analysis -> Model Recommendation -> Model
-  Installation -> Model Validation -> Mode Selection -> Brain Usage.
-- Reworked Library Models around the existing prepare/load stream so model setup
-  visibly moves through Environment Analysis, Recommended Models, Install,
-  Download Progress, Validate, Load, and Ready.
-- Basic mode now uses friendlier connected/needs-setup status language and
-  hides endpoint/module details in the Brain graph, model cards, and computer
-  readiness panels while preserving Advanced/Admin inspection detail.
-- Brain graph/search copy now emphasizes ideas, relationships, sources, focus,
-  filtering, and readability instead of backend endpoint implementation.
-
-### Fixed
-
-- Gemma 4 MLX routing now reads local model metadata before load. Gemma 4 12B
-  `gemma4_unified` shows **Runtime update needed** when installed MLX-VLM lacks
-  `mlx_vlm.models.gemma4_unified`; Gemma 4 26B A4B remains ready on the
-  standard `gemma4` MLX-VLM path.
-- Raw loader errors such as `No module named ...gemma4_unified` are converted
-  into actionable runtime-update guidance instead of marking the model ready or
-  routing it through an incompatible MLX-LM fallback.
-- Workspace selection now persists across reloads so the selected workspace is
-  used by API requests after restart.
-
-### Validation
-
-- Added unit coverage for Gemma 4 12B versus 26B runtime compatibility,
-  recommendation state, catalog routing, and the MLX-LM fallback guard.
-- Added Playwright coverage for first-run setup, model setup flow, Gemma
-  recovery guidance, and Basic graph developer-leakage prevention.
-- Full v4.5.0 RC validation is tracked in
-  `docs/V4_5_0_VALIDATION_REPORT.md`.
-
-### Artifacts
-
-- `dist/ltcai-4.5.0-py3-none-any.whl`
-- `dist/ltcai-4.5.0.tar.gz`
-- `dist/ltcai-4.5.0.vsix`
-- `ltcai-4.5.0.tgz`
-- `src-tauri/target/release/bundle/dmg/Lattice AI_4.5.0_aarch64.dmg`
+- Completed ALL Recommended next refactor items from report #15 in this session:
+  - Server decomp wave (model_runtime globals/wiring): model_loading.py for prepare_and_load and stream, _MODEL_RUNTIME_STATE, model_engines.
+  - Deeper WorkspaceOSStore (timeline + plugins + snapshots + memory).
+  - KG embed: set_embed_dim.
+- All refactoring needed finished this session per AGENTS. 767 tests, builds, greps clean.
+
+### Added (this session features)
+- `vision_analyze` tool: new multimodal vision analysis tool using screenshot b64 + prompt. Leverages existing VLM support (image_data in generate). Added to computer-use agent prompt and general tools. Fits seamlessly with computer_use, agent runtime, tool registry, and VLM models without affecting text-only paths.
+- More recent multimodal models in user recommendations (Llama 3.2 11B Vision, Phi-3.5 Vision, Qwen2.5-VL 7B, Moondream2) + family order update in model_recommendation. Expanded curated list in model_capability_registry for better local VLM choices on Apple Silicon and other.
+- All additions checked for compatibility with existing KG (descriptions can be ingested), agents/tools dispatch, chat/computer_use (image pass-through), model rec logic, and non-multimodal fallbacks.
+
+- Perfect completion of report #15 recommended refactor:
+  - Server decomp wave: ModelRuntimeState class (not dict) for globals/wiring in model_runtime; sync_to_module_globals for compat; further clean.
+  - Deeper WorkspaceOSStore: timeline/plugins managers fully delegated and composed (record, has_permission etc all through).
+  - KG embed: set_embed_dim available for optional central handling.
+- All changes preserve legacy exactly, full composition, small modules.