Lane tuning numbers — derived from <SB>/harness/state/metrics.jsonl alone
======================================================
Lane utilisation      : 50.0%  (300s busy across 2 lane(s) over a 300s span)
Mean semaphore wait   : 30.0s  (1 recorded waits, longest 30s)
Merge-conflict rate   : 100.0%  (1 conflict(s) across 1 finished lane(s))
------------------------------------------------------
Reading them: high utilisation with a LOW conflict rate is the case for
raising max_concurrent_lanes. A high conflict rate is the case against it,
whatever utilisation says. A long mean semaphore wait means the cap is
already above what the serialized legs (live-qa, heavy-test) can absorb.
These numbers inform a human. The scheduler never tunes itself.
