Shared-daemon benchmarks: taking the max over clients' process-tree totals drops N-1 clients' own cost

When N benchmark clients share one daemon and each client measures CPU/PSS over a process tree rooted at its own launcher plus the daemon, each per-client total = daemon + that client's own subtree. Taking the cell figure as max over clients (to avoid counting the daemon N times) quietly drops the other N-1 clients' own processes (their launcher/client shims). If the competing arm sums N standalone processes, the comparison is biased towards the shared-daemon arm by (N-1) shims.

Fix: have each client report its own subtree with the daemon excluded. Compute the cell as the daemon tree counted once (sampled when the last client finishes) plus the sum of the clients' own subtrees. Write down whether the harness process itself is counted, and treat both arms the same way.

Related pitfalls in the same kind of driver: - a cleanup comment says the daemon is killed but no code does it (the daemon has its own session, so killing the clients' process groups misses it); - a 'wait until all N clients are ready' loop with no deadline hangs while holding the bench lock if one client dies early. Add a trap that releases the lock on EXIT and a deadline on every barrier.