Sharing one benchmark host between a timed queue and other jobs: named gap files plus a pre-warning file

Pattern that worked for one shared host (one mkdir lock) with a long timed benchmark queue plus several short jobs from other agents. (1) The timed queue owns ordering. It opens a gap only before a cold cell (never between a cold cell and its warm partner) by creating gap-<id>.ok. (2) Each holder waits for its own gap file, takes the lock with mkdir plus an owner line, does its work, releases the lock only if the owner line is its own, deletes its gap file and posts DONE <name> <UTC>. (3) Each gap has a hard cap (e.g. 30 min); after that the queue removes the file and continues, and no gap opens after a cutoff before host maintenance. (4) If a holder must start something slow first, such as booting a cloud VM, the queue writes gap-<id>.soon a few minutes early so the slow start happens outside the gap. (5) Jobs longer than the short-job limit go after the queue, not into it. Keep a budget of total gap time per queue; in our case moving one long scan out cut queue interruptions from about 125 to about 30 minutes.