[codex] Add advisor watchdog for stale task-output waits

[morganmcg1]

Summary

Adds a narrowly scoped advisor watchdog path for stale Claude task-output waiters. This covers both observed DL24 stall variants without matching one literal sentinel string and without adding new launch/config surface.

What changed

• Detects Claude task-output wait loops scoped to /tmp/claude-.../tasks/*.output that use until, grep -q, sleep, and done.
• Intervenes only when the referenced output file has stopped changing for the existing script-level threshold and the waiter has no real worker descendants left.
• Kills just the stale waiter subtree so the outer advisor loop can recover and re-poll.
• Adds shell regression coverage for the original ERROR|=== waiter, the later custom grep waiter, and the safety case where a live Python worker is still running.

Why

The DL24 advisor pod was kube-healthy but the advisor loop wedged behind stale Bash waiters on Claude task output files. The first stall waited for the ERROR|=== sentinel; the later recurrence waited for a custom grep sentinel (no val yet|... state|... error). In both cases the task output file was stale and no real worker process remained below the waiter.

The fix keeps the boundary narrow: it still requires a Claude task-output path, a polling grep/sleep waiter, stale file mtime, and no real work below the waiter. It just stops depending on one exact sentinel string.

Validation

• bash -n k8s/advisor-claude-watchdog.sh tests/test_advisor_claude_watchdog.sh
• bash tests/test_advisor_claude_watchdog.sh

[github-actions]

Thank you for your submission, we really appreciate it. Like many open-source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution. You can sign the CLA by just posting a Pull Request Comment same as the below format.

---

I have read the CLA Document and I hereby sign the CLA

---

_{You can retrigger this bot by commenting recheck in this Pull Request. Posted by the CLA Assistant Lite bot.}