fix: close PI core chat review gaps

.project-docs/30-worklog/tasks/20260823-pi-chat-timeline-c8f31a62.md

@@ -130,24 +130,60 @@
   buffers during target recovery, trims Snapshot-covered items and replays only
   a strict continuous tail.
 - Fixed a supported concurrency defect exposed by the full suite: the Pi
-  Session Registry now reuses one Conversation store per project, serializing
-  simultaneous first bindings instead of racing writes to the same metadata
-  file.
+  Session Registry and Coding Project Service now receive the same project-
+  scoped Conversation store provider, so runtime bindings and metadata writes
+  serialize through one mutation queue instead of racing the same file.
+- Closed the first planner review findings after candidate `612135f`: the
+  authenticated Host server now admits only the exact image-binary upload
+  exception; empty/invalid/oversize/not-found attachments have stable typed
+  errors and supported image signatures are checked before disk writes.
+- Made active target ownership atomic across Agent switches. An old or archived
+  Conversation cannot supply the visible draft/timeline/prompt target, and a
+  slow automatic or explicit create completion only selects when its captured
+  project and Agent are still active. Submission flights and errors are scoped
+  per Conversation, so one pending/rejected target does not disable another.
+- Bounded each message to 16 local images and four concurrent uploads. A real
+  Electron image fixture now traverses the Renderer facade, IPC channel and
+  authenticated Host server, reads the stored bytes back for the optimistic
+  timeline preview, observes one upload, and verifies the prompt contains only
+  attachment ids rather than base64.
+- Extended target-only gap recovery: when the first recovery Snapshot remains
+  behind a buffered continuous tail, Renderer requests one fresh target
+  Snapshot; a repeated identical stale response becomes a retryable error
+  instead of remaining in an infinite recovering state.
 - Synchronized `README.md` with the current Pi core Chat, first-Conversation,
   batch and attachment behavior. PI-130 feature controls and PI-140 legacy
   removal remain outside this task.
 
 ## Verification
 
-- `pnpm exec vitest run tests/unit/coding-workspace-store.test.ts tests/unit/coding-projects-facade.test.ts tests/unit/coding-attachments-facade.test.ts tests/unit/coding-attachments-routes.test.ts tests/unit/coding-chat-panel.test.tsx tests/unit/coding-conversation-timeline.test.tsx tests/unit/coding-conversations-store.test.tsx tests/unit/coding-core-routes.test.ts tests/unit/host-api.test.ts tests/unit/host-api-proxy.test.ts`: passed, 10 files / 68 tests before the final malformed-batch and Session Registry regression additions.
-- `pnpm exec vitest run tests/unit/pi-session-registry.test.ts tests/unit/pi-conversation-runtime.test.ts tests/unit/coding-conversations-store.test.tsx tests/unit/coding-chat-panel.test.tsx`: passed, 4 files / 25 tests.
-- `pnpm test`: passed, 217 files / 2303 passed / 2 skipped. The final run is green after fixing the same-project parallel Session Registry write race and stabilizing the AI Hardware async query at its actual loading boundary.
+- First implementation candidate `612135f` received planner review result
+  **Standards Needs Fix / Spec Needs Fix**. PI-120 remained the unique Ready
+  Frontier; the findings above were treated as blocking rather than advancing
+  PI-130.
+- Post-review focused suites passed: attachment Host routes, stale recovery,
+  shared ProjectService/Registry mutation queue, slow Agent switch, cross-
+  Conversation submission isolation, attachment count/concurrency and pressure
+  tests all pass. The consolidated focused run covered 3 files / 30 tests, and
+  `coding-chat-panel.test.tsx` covered 6/6 behavior tests.
+- `REN-008` actual-path fixture passed with 100 runtime patch items, 20 public
+  patch batches, 21 SSE frames including Snapshot, 20 Renderer transactions,
+  20 React commits, 131,148 wire bytes, and measured Main-to-React p95
+  `35.995 ms` against the `<=50 ms` budget. It includes mixed message blocks,
+  4 KiB tool output and more than 100 KiB of cumulative thinking output.
+- `pnpm test`: passed, 218 files / 2316 passed / 2 skipped.
 - `pnpm run typecheck`: passed.
-- `pnpm run lint:check`: passed with 0 errors and six unchanged warnings in `ExecutionGraphCard`, `Home`, and `Makelore`.
-- `pnpm run build:vite`: passed; existing Vite dynamic-import and chunk-size warnings remain.
-- `pnpm exec playwright test tests/e2e/pi-coding-first-chat.spec.ts`: passed, 1/1.
-- `pnpm exec playwright test tests/e2e/pi-coding-first-chat.spec.ts --repeat-each=20 --workers=1`: passed, 20/20 fresh-userData runs. Every sample asserted Composer editability below 500 ms while runtime Snapshot was held, so the measured sample p95 is below the locked 500 ms budget; the same fixture verifies optimistic prompt visibility and HTTP 202 acceptance without a real Provider.
-- `git diff --check`: passed; only repository line-ending conversion notices were printed.
+- `pnpm run lint:check`: passed with 0 errors and six unchanged warnings in
+  `ExecutionGraphCard`, `Home`, and `Makelore`.
+- `pnpm run build:vite`: passed; existing Vite dynamic-import and chunk-size
+  warnings remain.
+- `pnpm exec playwright test tests/e2e/pi-coding-first-chat.spec.ts`: passed,
+  1/1. The post-review real-image fixture also passed 5/5 fresh-userData repeats
+  with one upload, Host readback preview, optimistic message, HTTP 202 and no
+  base64 in later payloads. The earlier text-only first-Conversation fixture
+  passed 20/20 fresh-userData runs below the 500 ms editability budget.
+- `git diff --check`: passed; only repository line-ending conversion notices
+  were printed.
 - Real Provider remains **Explicitly Waived / Accepted Risk** with
   `realTurnVerified=false`; no Provider Account, credential isolation,
   protocol-specific endpoint, rate-limit or real concurrency behavior is