* test(yjcore-01): characterize approved tool result communication

* feat(yjcore-01): narrate verified pilot outcomes with bounded local completion

* test(yjcore-01): verify default chat isolation and normalize model selection

* fix(core): bind narration claims to completed OS commands

* [YJUX-02] Resolve post-merge UI review findings

* [YJUX-02] Invalidate microphone proof across Settings device changes

* fix(voice): recover React StrictMode microphone and live device selection

Restore mounted media lifetime after StrictMode effect replay; invalidate
stale getUserMedia completions and make resource cleanup idempotent. Use
current selected device ref for all Voice entry points incl memoized palette.

Provide origin-checked, fixed OS microphone privacy settings IPC, explicit
audio-only permission probe with track cleanup, real diagnostic stages and
safe recorder finalization timeout without changing wakeword/queue/TTS.

Add real react-dom StrictMode mount, Windows/macOS privacy navigation,
exact-device Voice and media regression tests, plus P0 documentation.
This is the first P0-only checkpoint; P1 settings redesign not yet implemented.

* fix(settings): wire existing P0 microphone Voice-test props

Expose and pass through Voice stage/start callback from App to the actual
MicrophoneSettings child, preserving prior Settings behavior before P1.

* test(yjux-02): fix Node navigator descriptor and packaged recovery IPC harness

Keep the real React.StrictMode setup-cleanup-setup test, restoring all global
property descriptors even when tests fail. Load the shipped microphone recovery
module into packaged main's VM and assert only trusted renderer access to the
fixed Windows privacy Settings URI, never arbitrary shell input.

* test(yjux-02): flush real analyzer updates through React act in StrictMode

Keep the real StrictMode mount/unmount and signal recognition assertions;
execute the simulated interval inside act so React commits the meter update
before inspection on Node 22 and Windows runners.

* feat(yjux-02): finish source-backed Settings dashboard and model/voice libraries

Replace form-first Settings entry with responsive overview/status cards and
contextual navigation. Display separate real installed Ollama/Whisper/Piper
selection cards and curated installable catalog choices, never conflating file
presence with successful hardware/inference. Wire card installs into the
existing explicitly confirmed SetupInstaller without new download mechanisms.

Add live microphone verification evidence, local draft save/discard/error
states, app navigation warning, preserved custom settings and accessible
voice preview behavior. Expand active Settings and GuidedInstaller regression
tests and document manual hardware boundaries.

* test(yjux-02): assert actual visible JSX choices and unsaved navigation safety

Use rendered text-node concatenation rather than JSON child serialization for
installer cards with adjacent text. Retain Whisper/Piper prefill assertions.
Add explicit save-failure preservation and Settings-to-Chat dirty navigation
rejection/acceptance tests, without reducing any existing coverage.

* feat(desktop): add audio-only Electron permissions and discoverable device settings

Implement origin-restricted Electron 33 audio media permissions with macOS TCC
and NSMicrophoneUsageDescription; deny camera/display and unrelated origins.
Add exact device selection with local preference, safe Voice capture handoff,
independent real-sample RMS/MediaRecorder diagnostics and opt-in Whisper test.

Present live Ollama tags, bounded managed Whisper/Piper inventories and macOS
system voices through four active settings sections. Preserve advanced fields,
unsaved edits, existing SetupInstaller progress/cancel and approval/voice wiring.
Add OS-permission, media-device, diagnostic, model inventory and regression
tests plus manual-hardware verification instructions.

* test(desktop): exercise new microphone/settings tests in CI and preserve packaged Electron boundary

Keep new production settings separate from the historical, excluded experimental
src/components tree; use src/settings/IntelligentSettings for the active UI.
Update packaged Electron main test stubs for strict microphone permission APIs
and assert trusted audio-only checks; retain all preexisting smoke assertions.
Execute all four new microphone/inventory Node test files on both Ubuntu and
Windows rather than accidentally leaving them outside explicit CI test lists.

* fix(microphone): number only real inputs and verify rendered diagnostics in Node mocks

Filter phantom empty device IDs before assigning privacy-safe microphone
labels. Exercise visibly rendered Whisper latency instead of serialized React
nodes. Supply the browser AbortController to the isolated settings test VM;
retain all existing positive and negative assertions unchanged.

* feat(conversation): make normal dialogue LLM-first with live read-only toolkit facts

Route greetings, thanks, readiness, capability and clock/date questions through
existing persona/session Ollama streaming instead of canned Python replies.
Offer two explicit native streaming tool calls: verified local timezone-aware
system datetime and a read-only snapshot grounded in real registry/platform
availability, keeping unwired V2 capabilities clearly unavailable.

Reject unknown, malformed, mixed or repeated OS calls; require real toolkit
facts before any live time or capability claim. Preserve existing learned,
safety, approval and voice lifecycles and keep tool-result summary unchanged.
Add focused LLM-first, frozen-clock, capability and streaming tests.

* fix(llm): preserve JSON helper boundary while adding native streaming tools

Restore the original _extract_json_blob return and remove a sliced suffix
from the old streaming function. No routing, capability, approval or test
behavior changed.

* fix(conversation): preserve V2 domain isolation in live capability discovery

Build the live snapshot from actual Legacy ToolRegistry instances without
importing the independent V2 catalog into the legacy runtime. Mark the
V2 domain explicitly unwired/unavailable, preserving accepted separation.
Correct a session-history integration test to use the existing keyword-only
recording harness; retain every test and safety assertion.

* fix(conversation): keep unverified tool decision text off chat and TTS

Conservatively recognize natural live datetime/capability questions before
starting native model tool selection; withhold all first-pass model content,
reject mixed, missing or mismatched fact calls, and release tokens only after
verified Toolkit data. Ordinary conversation keeps one early-streaming model
request with no declared tools.

Check real backend binaries by registered implementation class: open,
pbpaste, pbcopy or osascript, without claiming OS permission/target readiness.
Reject multiple identical complete native tool call objects while preserving
single-call JSON argument fragmentation.

Add positive/negative WebSocket event-level, cross-platform availability and
native streaming regressions; preserve legacy approval and V2 boundaries.

* feat(conversation): add opt-in validated semantic pilot and scoped reminder follow-ups

Re-use the default-off legacy planner switch with one bounded structured
Ollama interpretation, an explicit five-tool allowlist, macOS-only support,
grounded argument validation and normal durable Approval. Keep direct,
learned, local and heuristic paths unchanged and fast.

Add expiring per-session, single-field reminder clarifications with verified
time/title, cancellation, context switching and safety re-check before the
existing approval lifecycle; never persist open intent or auto-execute.

Add model mock / negative security and end-to-end reminder tests, update old
optional planner expectations and document opt-in support boundaries.

* test(semantic): avoid reserved pytest request parameter

Rename the parametrized utterance argument only; retain all safety,
model-output and hallucination rejection cases unchanged.

* fix(conversation): resolve local intent collisions and improve dialogue flow

Replace substring-based date/readiness matches with complete question
patterns and restrict tool-heuristic fallthrough for ordinary discussion.
Keep safety/learning/approval order and ask one concrete missing-detail
question. Modernize default persona without rewriting old custom profiles.
Use bounded same-session context, overlap independent prompt reads and
publish the first real LLM token immediately; keep later stream/TTS batching.
Add golden routing, persona, memory isolation and first-token regressions.

* test(conversation): assert first-token streaming lifecycle accurately

Streaming deliberately emits the first token immediately, followed by
existing batches. Preserve all content/final-state assertions and match
all three actual token events without skipping or weakening tests.

* fix(conversation): preserve the full legacy tool intent catalog behind the dialogue gate

Keep the existing conversational question filter, but consult the established
pure legacy heuristic for other supported commands not covered by the short
verb list. Preserve learned precedence, safety and approval, unknown /tool
protection, and existing prompt, persona and streaming changes.

Cover every registered ToolRegistry heuristic plus music variants, seven
conversational counterexamples, real approval boundaries, learned triggers,
and safety precedence. Document that local clarification does not persist
an open multi-turn tool intent.

* fix(voice): finalize ordered wakeword and local Whisper input reliably

Add capture-ordered segmented wakeword window, stop-safe per-recorder chunks,
non-blocking final-only transcription, bounded ffmpeg/whisper subprocess
ownership, safe voice chat submission without blind retries, preserved
pending entries across setup checks, visible voice states and MIME-aware uploads.
Keep existing policy/agent/approval flow and TTS unchanged. Add focused
hardware-free recorder/queue/STT regression suites and update wiki docs.

* test(voice): correct callback test harness and classify uncertain 5xx ACKs

Use test-hook useCallback's real contract rather than invoking callbacks
on render; allow async Blob.text completion in fake recorder integration.
Treat 5xx/408/429 chat acknowledgements as potentially accepted, requiring
conscious retry, and cover this behavior with regression assertions.

* fix(voice): isolate wake activations and track actual run completion

Bind each capture and ordered STT processor to its own voice activation,
retain valid pre-Stop audio without carrying wakeword state into Restart.
Keep accepted Agent runs busy across socket disconnect and distinguish
terminal run IDs from HTTP ACKs including early websocket completion.
Bound terminal event history, make local message IDs monotonic, snapshot
TTS echo context at capture time and suppress recent short Jarvis echo.
Bound final subprocess reaping and chat acknowledgement waits.
Add focused lifecycle, race, echo, ID, timeout and cancellation tests.

* fix(ci): restore green baseline after PR #34

Restore actual packaging, startup and agent scripts, retaining dist:mac.
Limit initial TypeScript roots to production source while keeping all
unintegrated component files and their future transitive checks.
Withdraw PR34-only tests for as-yet-unimplemented contracts and document
the exact scope/evidence; add real packaging regressions.

Preserve new macOS tooling, native discovery, Voice and persisted data.

* fix(packaging): retain Windows output directory alongside macOS

PR34 moved the shared electron-builder output to release/, but existing
Windows packaged-backend and NSIS smoke tests read release/windows/.
Use platform-specific directories without changing the macOS output.
Add deterministic Windows/macOS config test and document CI failure.

* feat(desktop): add Windows installer and installed app startup

* fix(desktop): retain the macOS system application menu

* feat(ui): restore Jarvis presence and product identity

* fix(desktop): keep connection and action failures visible

* docs: establish reviewed technical wiki and commit changelog

* ci: fetch complete history for documentation ledger tests

* fix(docs): keep commit prose inert during wiki validation

* test(planning): characterize legacy planner fallback and learning resolution

* refactor(planning): isolate legacy planner behind a small adapter

Externally reviewed and explicitly approved by the user.

Externally reviewed and explicitly approved by the user.

Externally reviewed and accepted Windows test and CI baseline.

* test(orchestration): characterize legacy turn and approval lifecycle

* refactor(orchestration): extract behavior-preserving TurnEngine shell

Add independent frozen V1 fixtures and real-file tests for fresh/adopted databases, existing Database APIs, FTS triggers, version ordering, failures and concurrent starts. Document conservative recognition, transaction ownership and recovery limits.

Extract the frozen legacy schema, validate full schema/history/integrity before adoption, and commit each pending batch atomically. Preserve existing Database CRUD/defaults and keep Domain Contracts isolated.

Document the Pydantic choice, validation semantics, mutation limits and YJ2-01 no-runtime-integration boundary.

Introduce isolated V2 domain records and focused validation/isolation tests without wiring them into the legacy runtime.

Externally reviewed and explicitly accepted by the user. Preserve the accepted UUID, nonempty-plan, tuple/JSON-array, DAG, observation error-code and no-execution-authority decisions.

YJ2-02 is authorized only after all four baseline checks pass on this main commit.

Establish Python tests/lint and desktop typecheck/build CI, preserve the supplied V2 architecture, and document development setup and one-step external-review guardrails.

Includes two type-only Web Audio corrections; no intentional runtime behavior changes or runtime dependency upgrades.

Merged following the user's explicit acceptance of the external YJ2-00 review. YJ2-01 remains gated on all four checks passing on this main commit.