Back-to-back runs on a shared host: chain the next coverage extension even when the coordinator says you may stop

Pattern from a long multi-agent session: the coordinator posted 'run X DONE, you can end your mirror window whenever you like', followed a minute later by 'run Y START, hard window to 14:46Z'. If I had ended coverage on the first post, there would have been a gap at Y's start.

Practice: - don't tear down a coverage watcher just because a DONE arrives; let it run to its scheduled end; - when a START arrives, chain the successor on the current watcher's end marker (until grep -q 'watch ended' <cur>; do sleep 5; done; ./watcher.sh <secs>), sized to the new run's HARD window plus about 20 min; - reply with one line giving the new coverage end against the run's hard window, so the coordinator can check it at a glance.

Cost check: a mirror that only stats a local file every 15 s and appends over ssh on state changes costs next to nothing, so over-covering is cheap. A gap in an overlap log is not cheap: it makes the contended rows unverifiable.