docs(benchmarks): verify a quiet box rather than assuming it

[justinchuby]

A leaked benchmark process spun a full core on the dev box for eight hours — profile_native pid 9276, 01:00:52 to 09:05, ~5.6 CPU-hours — while casual checks reported the machine free. Every wall-clock figure taken in that window was measured a core short.

Both obvious liveness probes fail, in opposite directions, and the guidance circulating during the session named only the harmless one:

probe	failure	consequence
nvidia-smi, WMI	retains post-mortem rows → false positive	looks busy when idle; wastes time
Get-Process -Id <pid>	returned absent for that live pid; Stop-Process -Id also failed → false negative	looks idle when busy; invites starting a timing run on a contended box

The false negative is the dangerous one, and it is the one we were telling each other to rely on.

What actually worked: a repeated Get-Process -Name <name> snapshot, confirmed by a CPU-time delta over a wall interval — that process gained 5.55 s of CPU in 6 s of wall time. A process whose CPU time advances is alive whatever any single probe says. Terminating it required Get-CimInstance Win32_Process -Filter "ProcessId = <pid>" piped to Invoke-CimMethod -MethodName Terminate.

Also recorded:

• 0 % GPU utilisation is not evidence of an idle box. A live process holding a CUDA context sits at 0 % between kernels, and a spinning host-side loop consumes a core while showing no GPU activity at all. Two agents independently read 0 % as "free" during this window.
• Prefer contention-invariant counters where the question allows — CUDA-graph segment counts, htod_bytes, page-in and eviction counts, crash/no-crash ratios, token byte-identity. Several conclusions in this corpus rest on them precisely because a shared box cannot be trusted, and those conclusions are unaffected by this episode.

Sibling to §40 (a threshold calibrated in one regime misfiring in another) and §41 (a shared primitive's contract change breaking distant callers): all three are failure modes that produce confident, wrong results rather than obvious errors.

Docs-only. The measurements quoted are from this box today and the affected throughput figures are being re-taken now that it is genuinely quiet.