Bench

Latency percentiles from packages/hiveboard/bench/harness.py: N agents over M tenant graphs, 80/20 read/write, plus compare-and-set contention.