test(tasks): lock down handoff lifecycle + hook injection

[NagyVikt]

Summary

Puts a foundation under the Task Threads primitive shipped in #8 before more features get built on top. Two test files target the layers most likely to break silently — MCP-level state mutation and SessionStart context injection — plus three small source fixes the tests required.

What's new

apps/mcp-server/test/task-threads.test.ts — 4 MCP integration tests:

1. Atomic claim transfer — claim → handoff → accept migrates the claim to the receiver. Verifies the claim slot is empty between handoff and accept so a third agent can't grab the file in-flight.
2. Decline cancels + records a reason — declined handoffs MUST NOT transfer claims (the ugliest possible failure mode). Decline observation is visible in task_timeline so the sender sees refusal next turn.
3. Expiry rejection — accepting past expires_at throws and flips status from pending to expired so the sender's next turn sees the outcome instead of a handoff that looks live forever.
4. task_updates_since symmetry — each session only sees other sessions' posts. Tests both directions because a one-sided filter would pass a one-sided assertion.

packages/hooks/test/task-injection.test.ts — end-to-end proof that a fresh codex session landing on a branch where claude just left a handoff sees the handoff in additionalContext with a copy-paste-ready accept call (including session_id) and a decline hint.

Source fixes the tests required

• decline is now its own CoordinationKind (not a generic note) so task_timeline consumers can filter for it.
• buildTaskPreface is exported from session-start.ts so the integration test can drive it directly without the full runner/transport stack.
• SessionStart + UserPromptSubmit prefaces now include session_id in suggested accept/decline tool calls — agents were dropping session_id from ad-hoc calls and hitting validation errors.
• UserPromptSubmit caps injection at 6 most-recent messages with an (N older messages omitted) footer so one chatty agent cannot drown out another's context.

Test plan

• MCP server tests: 13/13
• Hooks tests: 11/11 (includes 1 new integration test)
• Core tests: 15/15 (existing decline test still green under the kind change)
• Typecheck 12/12 projects
• pnpm build green
• Biome clean on touched files

🤖 Generated with Claude Code