deployah — ARM A COMPLETE, 15/15, terminal 15:28:29Z. Full report preserved at
.spt/preserved/deployah-294-armA-20260910/ARM-A-REPORT.md. Diagnostic tree ed8ecd94, NOT the
candidate. Box is now free; no preflight probe started yet.

OUTCOMES: 15/15 PASS, 15/15 exit 0, 15/15 exactly one Summary, 0 expiries, 0 LEAK.

FIRST-PULL POLL — the margin, budget 400 observations at 25 ms:
  all 15 sorted: 45, 48, 81, 88, 91, 92, 151, 153, 162, 184, 186, 231, 246, 281, 300
  min 45 (11.2%, 1.116s) · q1 88 · median 153 (38.2%, 3.862s) · q3 231 · MAX 300 (75.0%, 7.599s)
  hertz's attempt2 draw of 175 sits inside this range. Spread across draws is 6.7x.
  Contributing stages on the first pull: store-init 0.196-3.497s, four tips combined 0.522-5.545s.
Later pulls kept visible, not pooled: second pull 4-86 iterations, third 5-31.
Whole-cell wall 43.942-70.864s, median 55.957s (hertz's attempt2: 90.946s).

I am NOT classifying that max. What I will say and nothing more: with NO COMPETING TEST WORKLOAD BY
DESIGN — todlando's wording, and it is not the same as "unloaded" — a single draw reached 75.0% of
the acceptance budget on a run that PASSED. doyle's third pre-registered reading is the one this
touches; whether 300/400 is an "excursion" is his call, not a word I will coin at report time, and
his ruling that the budget does not move this milestone stands either way.

NOT REPRODUCED: zero expiries under this condition. That is a CENSORED observation about the golden
condition, not evidence the defect is absent, and I file nothing from it.

NON-COMPARABILITY, kept explicit: the CI red's 55.685s is a TRUNCATED path and every figure above is
a completed pass. Not the same quantity; no conclusion drawn from their ordering, in either
direction. Written into the report so nobody reconstructs the comparison later.

HARVEST INTEGRITY: the extraction defect was repaired post-hoc from preserved raw logs with a
whitespace-tolerant pattern (found independently by me at 15:16Z and doyle at 15:17Z). The check that
matters is POSITIVE COVERAGE, not non-empty output: all 15 runs match the expected 20-stage /
65-event profile EXACTLY, and grep -c IR294 = 65 on every raw log. Nothing was lost because the raw
output was preserved beside the extraction.

A1 IS FIRST-IN-SERIES ONLY, no mechanism attached — the binary was built and executed during hertz's
attempt2 (mtime 15:00:02Z vs his 14:59:45Z start) and my pre-loop build recompiled nothing.

NEXT, and I start it now unless doyle says otherwise: the bounded three-measurement preflight, in
order, stopping at the first success — (1) Win32_ProcessStartTrace/StopTrace probe, unelevated, WITH
A POSITIVE CONTROL (spawn a known short-lived process and require ITS pid's start and stop events to
arrive; a subscription that registers but delivers nothing is recorded as a FAILURE, not a success);
(2) if that fails, nextest stamp semantics at both endpoints; (3) if that fails, the per-run bracket.
No fourth construction — UNVERIFIED and move on, per the ruling. Numbers to doyle before B1.