Phase 2: In-flight work index and dispatcher integration

[kai-linux]

Goal

Implement an in-flight work index so the dispatcher refuses to dispatch an issue whose work is already underway on another branch. Prevents the dispatcher from launching duplicate work that a previous cron run already started.

Success Criteria

• New runtime/state/in_flight_tasks.json updated atomically on every agent_launch and agent_exit event
• Records {task_id, issue_number, repo, branch, started_at, status: launching|running|completed|failed}
• Dispatcher consults the index before dispatching: if target issue appears in running state for any repo, dispatch is refused with "already in flight" reason logged
• Daily garbage collection: entries older than 48h marked stale and ignored (prevents crashed dispatcher from permanently blocking an issue)
• Integration tested: verify that dispatching same issue twice is blocked when first dispatch is in running state; verify stale entries are ignored

Constraints

• In-flight index must use atomic-write pattern (tempfile + os.replace) to prevent corruption from mid-write crashes
• Must not depend on Phase 1 implementation details (semantic dedup) — runs independently

Part of #259