Summary

Adds focused regression coverage for residual high-risk seams on main that are still untested and not claimed by open coverage PRs #298–#307 (main tip still at fa5c0f68). Also refreshes the canary mobile lockfile and unflakes the calendar-aged composer draft prune assertion.

Risky behavior now covered

• Research empty-input fail-closed — whitespace-only start / followUp / refinePlan throw before persisting a row or opening a Google stream; refinePlan/approvePlan return null when the research is missing or not awaiting a plan.
• Refine-plan execution — a pending plan with trimmed user input starts a collaborative stream (input, previousInteractionId, collaborativePlanning: true) and clears planPending.
• Research JSON-RPC schema — whitespace-only research/start, research/followup, and research/refinePlan prompts are rejected at the protocol boundary.
• Transcript inbox hard capacity — when a failed projection leaves a pending batch occupying every slot, further appends throw Transcript inbox capacity reached (503); a rejected empty batch (413) does not consume a slot.
• Reconnect feedback — reconnectThreadWithFeedback reports failure when reconnect cannot start, and when the chat is removed while waiting for connected.

Test files added/updated

• test/research/research.service.validation.test.ts (added)
• test/transcriptInbox.capacity.test.ts (added)
• test/research.schemas.test.ts
• apps/desktop/test/thread-reconnect.test.ts
• apps/desktop/test/composer-draft-thread-switch.test.ts (Date.now()-relative prune fixtures)
• apps/mobile/bun.lock (CI stabilizer)

Why these tests materially reduce regression risk

These paths are stuck-state and data-validation hazards: empty research briefs that would otherwise persist/stream, inbox overflow when projection fails closed, and reconnect UI that can report success after the chat is gone. None duplicate open #298–#307 claims. No production behavior changes.

Validation

• Targeted new suites green
• Touched files Biome-clean
• bun run lint / bun run typecheck / bun run docs:check / bun run check clean
• Full bun run test: all 674 files passed (Bun 1.4.0-canary)