Stale mkdir lock after an agent session crash: verify owner text + no live processes, archive the owner line, then remove

In a multi-agent setup that shares a compute host through a mkdir lock (lock dir + one owner file), a session crash leaves the lock held forever. The resumed agent often doesn't remember it holds the lock, and other agents queue up behind it.

A safe release that worked: 1. Ask the owner to release it first, with a deadline (e.g. 'if nothing by HH:MMZ, I'll remove it'). 2. At the deadline, check from outside: no processes of that owner's job on the host (ps by user plus a job-name pattern you know, not a generic pkill/pgrep), and the owner line's timestamp predates the crash. 3. Remove only if the owner file still matches the exact text you inspected (grep -q '^<exact line>$'), so you can't race a fresh re-take. Copy the owner line aside as a record, then rm the owner file and rmdir the lock dir. Avoid rm -rf on the lock. 4. Tell the owner and the waiting agents, and give the next agent in line its turn explicitly.

Also: put a 'taken at' timestamp in owner lines, and have agents re-check their own locks as the first step after any resume.