fix(kernel): checkpoint TTL evict race and wake loss on inject error

[cursor]

Summary

Fixes two critical correctness bugs found in recent kernel changes.

1. Checkpoint corruption during slow TTL evict (data loss)

Impact: Daemon restart could drop a live session's conversation/goals.

Root cause: CorePool.evict() pops the session, then awaits slow work (summarize). If the user messages during that window, a new Core is acquired and writes a fresh checkpoint. The stale evict then called mark_expired() on the shared checkpoint file, setting expired=True on the new state.

Fix: Skip mark_expired() when the session is already back in _pool.

Trigger: Idle session past Core TTL → TTL evict starts (summarize takes seconds) → user sends a message → turn completes → stale evict marks checkpoint expired → restart loses session.

2. Scheduled wakes dropped after inject_turn errors

Impact: Goal auto-continue / schedule_wake silently stops after any processing error (e.g. LLM 401).

Root cause: On inject_turn failure (no waiter), _finalize_agent_wake_delivery(confirmed=True) permanently removed the wake instead of aborting for retry.

Fix: Use confirmed=False on the error path, matching cancel/skip semantics.

Validation

• test_evict_skips_mark_expired_when_session_reacquired
• test_inject_turn_error_aborts_agent_wake_delivery

Notes

• Existing open dashboard security PRs (fix(dashboard): restrict tasks/queue leak and gate config to admin #13, fix(dashboard): gate jobs/user admin ops and redact users in snapshot #14) unchanged.
• Remote TTL silent local fallback (separate issue) not addressed in this PR.