Summary

Re-delivers the high-risk regression coverage from closed-unmerged PRs #294–#297 onto current main. No intentional production behavior changes beyond tiny existing testability seams (__internal / deferred-sync timeout hooks).

Risky behavior now covered

• Deferred workspace-defaults catch-up — settings saved while control is connecting apply once the session becomes ready (no silent drift).
• Optimistic new-chat failure cleanup — one-off startup discard/trash, mid-flight draft edit rollback, NewChatLanding unmount abort guard.
• Runtime integrity mid-verify race — fail closed if runtime trust is invalidated while an entrypoint is being verified.
• Connected stale thread/list prune — live connected chats are not dropped from a stale list snapshot.
• setModel stale preferred-child reset — same-provider preferred child cleared when invalid for the new parent.
• Raw-backed non-replay emit — tool_result/error chunks still emit on raw-backed turns.
• Attachment-bearing newThread optimistic deferral — no optimistic bubble until attachment prep finishes; pre-resolved still early-optimistic; start-fail never queues IDs.
• Canvas soft revision / inode identity — sub-second mtime/ctime skew + ino===0 soft match.
• PI rate-limit no-retry after toolcall/thinking — mapped visible parts block retry (avoids duplicate tool/reasoning).
• loadConfig preferred-child allowlist — invalid preferred child resets to first allowlisted ref while preserving mode/allowlist.
• Send-queue opt-out / external sink — shouldSendNotification and external sink routing/cleanup.
• Canvas save conflict/superseded broadcasts — conflict/superseded saves return schema-valid results with zero fileChanged broadcasts.

Also includes the #296 CI stabilizers: research-readiness desktop bridge install and mobile lockfile refresh for canary Bun.

Test files added/updated

• apps/desktop/test/workspace-settings-sync.defaults.test.ts
• apps/desktop/test/new-chat-landing-abort.test.tsx (new)
• apps/desktop/test/workspace-startup.test.ts
• apps/desktop/test/control-socket.sessions.test.ts
• apps/desktop/test/research-readiness.test.tsx
• test/coworkRuntime.test.ts
• test/session/agentSession.model.test.ts
• test/session.stream-pipeline.test.ts
• test/canvasDocumentPersistence.test.ts
• test/runtime.pi-rate-limit-retry.test.ts
• test/config/config.loadConfig.test.ts
• test/server.backpressure.test.ts
• test/jsonrpc.workspace-route.test.ts
• Tiny seams: apps/desktop/src/app/store.actions/workspaceDefaults.ts, src/coworkRuntime/integrity.ts, src/server/canvasDocumentPersistence.ts

Why these tests materially reduce regression risk

These paths sit on cold-start, chat creation, runtime trust, session list sync, model routing, streaming, and canvas persistence — failures look like data loss, ghost messages, silent settings drift, or trusting a mutated runtime. Prior daily coverage PRs that owned these gaps were closed without merge; this consolidates them with the known CI fixes.

Validation

• Targeted suites: 266 pass / 0 fail
• bunx biome check on touched files: clean
• bun run lint: clean
• bun run typecheck: clean
• bun run docs:check: OK
• bun run test: all 673 files passed