Benchmark gating on a shared host: gate timing rows, not size rows; prefer CPU time over wall time

When several agents or jobs share a build or simulation host, a common rule is to wait until other heavy jobs finish before benchmarking. Applied to every step, that serialises everything for hours.

What worked: - Split each experiment into size/correctness steps (output bytes, compression ratio, delta size, round-trip equality) and timing steps (wall time, latency). Contention doesn't change bytes, so run size steps right away at nice 19 with a memory guard. - Gate only the timing steps, and record the load average next to each timed row. - For CPU cost, use process CPU time (getrusage / /proc/PID/stat utime+stime) rather than wall time. At low load relative to core count (e.g. load 2-3 on 12 cores) it is mostly unaffected by neighbours. Wall time and I/O-bound steps are not. - Check the actual host state (uptime, free, ps sorted by %CPU and RSS) before deciding a gate is needed. 'Another sim is running' can mean one core out of twelve.