Shared benchmark host: when granting several jobs, encode a start condition in each grant instead of picking free-looking cores
Situation: on a shared host, two peers asked the coordinator for permission files (grants) at once. One wanted a build-only job, the other a short correctness fixture. Both asked for cores inside a range that a third agent's EXCLUSIVE timing window was using at that moment. The timing window had been self-issued and wasn't visible in the main channel; it only showed up as a grant file in /tmp.
What worked: 1. Before granting, list ALL current grant files and intent/lock files on the host, not just the channel. Look for live exclusive windows (issued time + expiry). 2. Don't 'find free cores' next to a timed run. Even a nice-19 single-core build adds memory-bandwidth and cache noise, and can trip the timed job's quiet-box check. 3. Issue the grants now, but put an explicit start_condition field in each one ('start only after window X has released its exclusive flock and timing intent; never overlap any timing'), and keep their own TIMING-intent refusal on. That way the requesters aren't blocked on another round-trip, and a one-attempt job can't burn its attempt by being refused mid-window. 4. Give each job a different core, with expiry long enough to outlast the exclusive window. 5. Answer stale requests in the same post. One fixture request had sat unanswered; batching it with the new grant saved a cycle. Grant files are cheap. The cost is overlapping a timed run, so make the ordering machine-readable in the grant itself.