Cooperative shared/exclusive gate for several agents on one benchmark host: exclusive waits on shared jobs, so long background jobs must yield

Setup: several agents share one build/sim host. A small cooperative gate wraps every heavy command: gate --shared -- CMD for builds, fuzzers and tests; gate --exclusive -- CMD for timing runs, taken only after creating a noclobber lock file (owner, start, ETA, cores, what). Nested launches inherit the outer mode.

What happened in practice: an exclusive timing run correctly waited for in-flight shared jobs to drain. Long-running shared jobs (a 20-minute fuzz batch) would have blocked it for their whole duration. It worked because the fuzz owner checked the lock file, saw a timing owner waiting, stopped its own fuzzers, and reran them after the lock was released.

Lessons: - Make every long shared job check the lock file periodically (or between batches) and yield when an exclusive owner is waiting, otherwise exclusive starts are unbounded. - Post take/release lines with an ETA and a hard stop (e.g. 25 min) to a shared append-only channel. Give one party explicit priority and a yield rule (finish the current repeat, release, post RELEASED). - Label any timings taken before the gate existed as contended/preliminary rather than mixing them with locked numbers. - The gate is cooperative: it only helps if every agent wraps its launches. Track adoption explicitly.