Bound graceful daemon stop to thirty seconds

[Aaronontheweb]

Change

• Limit session drain to 20 seconds inside a 30 second shutdown phase.
• Keep the CLI force limit at 45 seconds and the generated systemd stop limit at 60 seconds.
• Update daemon and container guidance. A pod should allow at least 60 seconds for the full graceful stop budget.

Reason

The prior 190 second drain bound waited for work that could not finish during shutdown. The preceding PRs persist accepted input and confirm model cancellation before they grant a resume candidate. The shorter limit caps a stalled stop. A tool with an uncertain external effect remains quiet after the limit; Netclaw does not replay it automatically.

Evidence

• Configuration budget tests: 4/4 passed.
• CLI daemon tests: 16/16 passed.
• Daemon shutdown and coordinator tests: 10/10 passed.
• Spark2 Skill Discovery eval: 8/8 passed with deepseek-v4-flash-dspark (run 096fdfca-31ac-4dc3-8cbd-47e39b6ce9ac).
• One disposable home and an active model turn: baseline stop 7.47 seconds, listener after 10.05 seconds; candidate stop 7.47 seconds, listener after 9.96 seconds.
• Both isolated stops retained a mode 0600 manifest. The second stop retained the earlier candidate and added the new one.
• Slopwatch found zero issues. Header verification and diff checks passed.

The normal stop path showed no meaningful speed change. The new limit reduces the maximum wait for a stalled drain. The active daemon on this machine was not restarted.

Stack

Base: #2205. Merge this PR after #2205.

This PR replaces #2206 so GitHub can link the same-repository branches as a stack.

[Aaronontheweb]

LGTM