Shared benchmark host: label overlaps per run type (timed vs untimed) rather than blocking every user session

A coordination pattern that worked across a long multi-agent session on one shared Linux benchmark host: - TIMED runs take an exclusive lock or intent file. Interactive tools (a demo viewer's backend) check it and refuse or stop while it exists, so the timed numbers stay clean. - UNTIMED or contended runs (correctness smokes, builds, a long run the user explicitly accepted as CONTENDED) take NO lock. Interactive use continues, and overlaps are only recorded: an append-only activity log on the host, written by a watcher on the user's machine ('<UTC ISO> viewer-active|viewer-idle <pid|none>'). The run's records mark each overlap. - The coordinator posts START/DONE for every run in the channel, with cores and expected end. The watcher's coverage is extended to each announced deadline, with chained handovers and verified marker lines.

Why: blocking the human's session for every untimed job is costly and buys nothing, since untimed results don't depend on a quiet machine. Labelling keeps the record honest for the few rows that are timed under contention.

Pitfall: a run that starts while a user-hold is in force (a stale schedule) should be asked to withdraw or to finish within a short bound. Then confirm whether it touched the user's processes; here it hadn't.