deployah — PREFLIGHT MEASUREMENTS 1 AND 2 ARE IN, BOTH NEGATIVE, AND 3 IS UNAVAILABLE AS PROOF.
Outcome: OVERLAP UNVERIFIED for Arm B. Preserved at .spt/preserved/deployah-294-armB-preflight/.

MEASUREMENT 1 — the protocol's named recorder. RESULT: **RECORDER UNAVAILABLE**, and it failed at
SUBSCRIPTION, before the positive control could even run:
  Win32_ProcessStartTrace : FAILED — "Call cancelled"
  Win32_ProcessStopTrace  : FAILED — "Access denied"
  elevated: false · controlPid: null · start events 0 · stop events 0 · control matched: false
The two errors differ in wording and I am reporting both verbatim rather than collapsing them to
"denied". The control process was never spawned because the subscriptions did not register, so this
is a subscription failure and NOT a delivery failure — a distinction that matters if anyone retries
this elevated later. My earlier guess that this would be denied is NOT what established it; this
probe is.

MEASUREMENT 2 — the non-invasive nextest side file. RESULT: **THE FLAG DOES NOT EXIST HERE.**
cargo-nextest 0.9.137 (75ddba7e9, 2026-05-26) offers `--message-format <human|libtest-json|
libtest-json-plus>` and `--message-format-version`, both marked experimental. An exhaustive census of
every file-directed flag in `nextest run --help` finds `--archive-file`, `--config-file`,
`--user-config-file`, `--tool-config-file` — and NO message-format file sink. So structured output
would have to REPLACE the human stream on stdout, which is the same stream carrying the IR294 lines,
i.e. it would modify output capture — the one thing the protocol forbids between repetitions. I did
not assume the flag was missing; I read the installed binary's own help.

MEASUREMENT 3 — the per-run bracket. NOT ATTEMPTED AS PROOF, and deliberately so. Its inner window
needs a JUSTIFIED FINITE cross-clock error bound, and both of you ruled that paired samples give
empirical calibration only, never a guaranteed bound between samples or during a later run. There is
no fourth construction to try, per the ruling, so I stop here rather than iterate.

**PREFLIGHT VERDICT: overlap UNVERIFIED. Arm B runs under doyle's amendment — permission to execute
is separate from permission to claim concurrency.** What I will still collect during B, labelled as
its own evidence class with NO concurrency claim attached: sampled process presence with stamped
query intervals and (PID, birth time, image) identity per KH 7.51 / REQ-HAZARD-RESUME-CUSTODY-ABA
and 7.58, because a sampler keyed on a bare pid could observe a recycled pid and manufacture the very
concurrency sampling was supposed to be incapable of.

TWO CORRECTIONS ACCEPTED AND ALREADY APPLIED TO THE ARM A REPORT ON DISK:
1. todlando: Arm A's 15 waits all COMPLETED, so they are UNCENSORED — they simply did not reproduce
   the golden failure. "Censored" belongs to the EXPIRED polls. My report said "a censored
   observation about the golden condition"; that sentence is replaced, with the correction and its
   author named in place rather than silently reworded.
2. doyle: "excursion" withdrawn as a classifier — nobody defines a threshold with the sorted list in
   front of them. The report states the distribution and no verdict word.
Also carried: zero expiries in 15 draws does NOT bound the failure rate; the probability of exceeding
400 under this condition is UNMEASURED, not small, and the report says so.

NEXT unless doyle says otherwise: Arm B PREPARATION only — extract the HEAVY expression verbatim from
golden.yml at ed8ecd94, freeze the expanded selection, prebuild fixtures and test binaries, check the
32 GiB floor, and freeze the selected listing with exactly one subject identity and zero
continuation-control identities. **I will not start B1 until doyle has these numbers and says go.**