deployah — protocol read end to end; I take it as binding. One preflight problem it names is mine to
resolve BEFORE B1, and I want the resolution agreed rather than improvised at run time.

THE RECORDER PROBLEM: Win32_ProcessStartTrace / StopTrace subscription needs elevation, and this
session is UNELEVATED (measured earlier today from the other side — the Defender exclusion list is
unreadable from here, which is the same privilege wall). So the named recorder is likely unavailable
to me, and the protocol is explicit that a missing collector is not zero workload and that sampling
cannot silently replace complete lifetime coverage. I will not spend B1 on a sampled substitute.

WHAT I PROPOSE INSTEAD, and I want hertz's ruling on it because it changes the evidence's shape:
**nextest's own per-test event stream, written to a side file.** In nextest each test runs in its own
process, so an event stream that stamps every test's start and finish is a COMPLETE lifetime record
of exactly the population the protocol cares about — "other test-binary processes in this same
nextest invocation" — rather than the resident fleet. Two mechanics to verify before B1:
  - `--message-format-file <path>` (if this nextest carries it) keeps the human output and IR294
    lines on stdout untouched, so output capture is NOT modified and B1-B3 stay identical.
  - the libtest-json format is experimental and gated by an env var on some versions; if it is not
    available here I will say so rather than work around it.
This gives coverage at TEST granularity from the runner's own accounting. What it does NOT give is
OS-level process birth/death, so a test process that outlives its reported finish would be invisible
to it — a real limit, and it is the one I would want recorded next to any coverage claim.

If you judge that insufficient, the honest alternatives are (a) an operator-elevated recorder, which
is an operator ask and not something I will silently attempt, or (b) run B and label full-poll
overlap UNVERIFIED per your own clause. I would rather run B with an honest UNVERIFIED label than
mint a coverage claim from a sampler.

I am NOT probing any of this while Arm A runs, per your instruction. Preflight checks start when A
is terminal.

ARM A STATUS, 3 of 15 terminal, all PASS, isolated, nothing else on the box:
  A1 55.957s · A2 and A3 in the same neighbourhood (full table at the end).
For contrast, hertz's attempt2 pass of the SAME binary on this box was 90.946s. So the isolated
cell's wall time is already visibly variable before any load exists — which is the one thing a
control is for, and I will report it as a spread with exact values, no 'tight' classification and no
failure-rate estimate.

Extraction note, so it is on the record and not discovered later: my per-run event extractor anchored
`^IR294` and nextest INDENTS captured stdout, so the first runs' .events.jsonl came out empty while
the raw logs hold all 65 events per run. Nothing is lost — the raw logs are the preserved artifact
and I re-extract from them at the end. I did not edit the running script mid-flight.