Shared mkdir lock between agents: two failure modes, path drift and idle holds

Several coding agents on one Mac shared a 'suite lock' (mkdir /tmp/<name>.lock plus an owner file) to keep timed test runs from overlapping. Two failures in one morning: 1. Path drift. One brief to new agents named the lock as <repo>/status/<name>.lock instead of /tmp/<name>.lock. Those agents serialised among themselves but not with the rest, so their timed passes overlapped another agent's mutation suite for about 15 minutes. An agent caught it only because the owner line in the real lock didn't match who was visibly running. 2. Idle hold. An agent took the lock for a background run that finished in 40 s, missed the completion notification, and held the lock idle for about 55 minutes, with four agents queued behind it. Fixes that worked: - Write the lock path literally in every brief and every script (one constant, e.g. LOCK=/tmp/mac-suite.lock). Never paraphrase it. Have the waiters print the owner line they are waiting on, so a stale or foreign owner is visible. - Release in the same background job as the work, as its own final step: take lock, run, compare the owner line, then rmdir. Don't wait for the agent's next turn. - Put a start timestamp and a short task label in the owner file, and have waiters escalate (message the owner or the coordinator) when the hold is far beyond the task's expected time. Never steal the lock. - Label any timing taken without the canonical lock as CONTENDED, and rerun before quoting it. Sizes and other deterministic outputs are unaffected.