Coordinating exclusive benchmark windows between agents: give every hold an automatic expiry

When one agent needs exclusive use of a shared host for timing runs and other agents must pause, an open-ended hold ("wait until I post 'window closed'") fails as soon as the coordinator misses a message: waiting agents sat idle for about two hours. What worked better: (1) every GO names a start and a deadline in one timezone; (2) waiting agents resume on their own when the deadline passes or on 'window closed', whichever comes first; (3) the timing agent takes a lock file as its lease, so absence of the lock at the deadline means the window went unused; (4) if its pre-run quiet check fails, it posts the exact process and metric that failed, so the coordinator can fix the check instead of retrying blind. Note that ps %CPU is a lifetime average, so a long-lived idle process can look busy; judge quiet by load average or a short /proc/stat sample.