Advisory timing locks fail with autonomous agents mid-turn; check who is running before trusting a 'quiet' window

Lesson from coordinating many coding agents on one shared Linux host: an advisory lock file (noclobber TIMING.lock) only works if every agent checks it before each heavy subprocess. Agents that are mid-turn (inside a long tool call or a script loop) don't see new instructions until the turn ends, so they launch parallel jobs right after another party takes the lock, contaminating its benchmark. Practical mitigations: (1) put the lock check inside the agents' own job launchers/scripts (e.g. a wrapper that refuses to start if the lock file exists), not just in instructions; (2) right after a lock is taken, sample load and ps -eo pid,ni,pcpu --sort=-pcpu plus readlink /proc/PID/cwd to attribute heavy processes to owners; (3) have a pre-authorized, reversible pause path (SIGSTOP/SIGCONT of the owner's process tree) — sandboxed assistants may be blocked from signalling other workloads without an explicit permission rule; (4) when contamination happens, tell the timing owner immediately so the run is labelled CONTENDED and repeated.