hertz 09:31Z — WINDOWS LEG LANDED. Both halves of what you asked, and the prediction held.

1) PREDICTION TEST — HIT, EXACTLY. Posted before the leg: 3114 run / 1 skipped. Measured:
`Summary [ 189.614s] 3114 tests run: 3114 passed (5 leaky), 1 skipped`, job 101688097059,
completed 09:29:42Z, conclusion success. SummaryCount = 1, correct for ci.
So both boxes now read as classified: Linux ci 3088 -> 3101 (+13), Windows ci 3101 -> 3114 (+13).
Equal deltas on two OSes = the platform-neutrality claim was not merely asserted, it was RISKED
and survived. The golden prediction (Linux 3300/220/3520, Windows 3325/235/3560) rests on the same
neutrality argument, which is now one measurement stronger than when I posted it.

2) THE LIST-PHASE READ — ALIVE THROUGH, and it is a RESULT.
   LIST window this run: `Finished test profile` 09:26:17.7802958Z -> `Starting 3114 tests across
   26 binaries (1 test and 200 binaries skipped)` 09:26:27.9013979Z. **Window = 10.121s, 226
   binaries touched.**
   Watcher, pid 48232: poll 3570 logged 09:26:05.912Z (11.87s BEFORE the window opened) and poll
   3600 logged 09:26:40.847Z (12.95s AFTER it closed). Exactly 30 polls spanned the window at
   1.16s/poll, so roughly NINE probes fell inside the 10.121s window itself.
   Whole-log census of non-heartbeat lines: ARM x2 and the CONTROL probe. **Zero DEATH, zero
   IDENTITY-CHANGED, zero WITNESS.** Watcher still running, poll 3780 at 09:30:06Z.

   WHY THAT IS A NEGATIVE AND NOT A SAMPLING GAP — I checked the script rather than assuming, and
   it is the load-bearing point of the whole read. `alive poll=N` is gated on `if ($n % 30 -eq 0)`
   (line 48), so the heartbeats ARE sampled. But DEATH (line 27) and IDENTITY-CHANGED (line 45)
   are logged UNCONDITIONALLY on any poll that sees them. So the nine probes inside the window
   each ran the full check and none of them had anything to report. Absence of a death line across
   that window is evidence, not silence. Identity is re-verified every poll by BOTH image path and
   creation time, so a reused pid would have logged IDENTITY-CHANGED rather than reading as
   survival — the specific way this observation could have lied is the one thing the instrument was
   built to close.

3) STRENGTH OF THE EXPERIMENT, stated honestly. The window that killed 14444 was **7.795s**
   (08:03:08.010 -> 08:03:15.805, same workflow, same runner, same 226 binaries). This one was
   **10.121s** — 30% MORE exposure to the same churn, same box, and the daemon held. That is the
   right direction for a survival observation to point.
   It does NOT refute pid reuse: reuse is probabilistic and a single survival cannot falsify a
   probabilistic mechanism. It converts the hypothesis from untested to tested-once, negative.
   **No killer has been shown. The incident stays OPEN** — nothing here licenses closing it, and I
   would resist a read of "alive through one LIST phase" as a clearance.

4) SIDE OBSERVATION, unasked, flag only. Windows ci leaky count moved 8 (e3bd53d4) -> 5 (401a19ad)
   while Linux held at 2. Matches the 5 todlando reported from his own battery, so it is the tree's
   number and not a runner artifact. Given ir79/ir80 are both about rigs leaking daemons, a drop of
   three is worth a glance from whoever takes those — I am NOT claiming the W2 commits caused it,
   only that the number moved and the direction is favourable.