Resume interrupted sessions after graceful stops

[Aaronontheweb]

Change

• The session actor records a durable candidate after a graceful stop confirms model cancellation.
• The daemon restores an output subscriber before it asks the actor to resume the original turn.
• The actor uses the recorded input, requester, and trust authority. It follows the original turn with one ordered queue call.
• Slack, Discord, and Mattermost check every recorded requester against the current channel ACL.
• A ten minute absolute deadline blocks old candidates. A durable claim blocks duplicate model calls after a crash.
• The restart manifest retains blocked candidates until the output route returns or the deadline expires.

Evidence

• Actor suite: 4,037 passed, six platform skips. A later focused queue grace test passed.
• Daemon suite: 1,135 passed, one platform skip.
• Spark2 Skill Discovery eval: 8/8 passed with deepseek-v4-flash-dspark (run c34ad157-0e4f-41ac-80ac-11186e359b18).
• Isolated production image stop took 8,304 ms. The persistent home retained one candidate with mode 0600. Restart kept the candidate quiet without an output subscriber.
• Slopwatch found zero issues. Header verification and strict OpenSpec validation passed.
• All 15 OpenSpec tasks are complete. Main session and container specs are synced.

Stack

Base: #2204. Merge this PR after #2204.

The active daemon on this machine was not restarted. Live rollout remains unverified.

This PR replaces #2205 so GitHub can link the same-repository branches as a stack.