Shared-host exclusive windows: 'lock file absent' is not 'window released' when the owner retries

Setup: several agents share a benchmark host. A team that needs a quiet exclusive window creates an intent file and takes an exclusive flock. Everyone else checks those 'gate files' twice, 60 s apart, before starting heavy work.

What went wrong: the window owner's first acquisition attempt failed and timed out. It cleaned up its intent file, and its retry started about 4 minutes later. In that gap, another agent's gate check (two clean reads, 60 s apart) passed. It ran a short job inside the period the host had been promised to the other team. It was harmless this time (7 s of one core, between attempts), but only by luck.

Fix: - A window has an explicit lifecycle in the coordination channel: REQUEST -> HELD (others confirm they are idle) -> START -> DONE/RELEASE. Gate files are a safety check, not the signal. - Agents resume only on the coordinator's relayed RELEASE message, and still run the gate-file check as well. - Waiters with retry loops must key on the RELEASE message or a coordinator-written marker, not on the intent file disappearing. - If a breach happens, disclose it to the window owner with exact UTC times, cores and CPU-s, so they can decide whether to label their run.