hertz 10:10Z — WINDOWS LEG CONCLUDED 10:08:49Z, success. Three things.

1) PREDICTION: 3116 posted, **3116 measured** (`Summary [ 199.918s] 3116 tests run: 3116 passed
(9 leaky), 1 skipped`), SummaryCount 1. Third consecutive exact hit, all three posted BEFORE their
leg landed: 3114, 3103, 3116. The classification (lib vs integration, A vs B, platform-neutral) has
now been risked four times across two OSes and has not missed.

2) LIST-PHASE READ #2 — ALIVE THROUGH, but WEAKER evidence than #1, and I want that said plainly
rather than buried.
   Window: `Finished test profile` 10:05:14.5450725Z -> `Starting 3116 tests across 26 binaries
   (1 test and 200 binaries skipped)` 10:05:20.5134063Z. **5.968s, 226 binaries.**
   Watcher pid 48232: poll 5580 at 10:05:09.176Z (5.37s before open), poll 5610 at 10:05:43.184Z
   (22.67s after close) — 30 polls spanned it, ~5 probes inside the window. Whole-log non-heartbeat
   census still exactly 3 (ARM x2 + CONTROL). Zero DEATH, zero IDENTITY-CHANGED. Still running,
   poll 5820 at 10:09:43Z.
   **WHY IT IS WEAKER:** this window is 5.968s. The window that killed 14444 was 7.795s. So unlike
   observation #1 (10.121s, which EXCEEDED the killing exposure by 30%), this one is SHORTER than
   the killing exposure and therefore does not even reach the condition that has been shown to kill
   once. "Two survivals" is not twice the evidence — it is one survival above the bar and one
   below it. Cumulative LIST exposure survived is 16.089s across two runs against a single 7.795s
   death. **Still not a refutation, and I would not present it as trending toward one.**

3) THE LEAKY RETRACTION IS NOW DECISIVELY CONFIRMED, and in the direction that embarrasses the
   original flag hardest. Windows ci leaky across four shas: **8, 8, 5, 9.** Nine at f3c8495b — the
   HIGHEST of the four, on the tip carrying the MOST fix work (F1-F16 all in). Anyone reading the
   count as a quality signal would now have to conclude W2 made leaking WORSE, which is exactly as
   wrong as my "direction favourable" read was at 8->5, and wrong for the same reason. The count is
   a threshold artifact in both directions. Arc rule holds: name-diff plus causality check, or it
   is noise.
   Union is now TEN cells; the newcomer is `spt-daemon shellhost::tests::kill_shell_at_still_kills_a_matching_pair`,
   never seen at the previous three shas. **And it is ANOTHER kill cell.** So the structural finding
   survives the population growing by a quarter: 10 of 10 union members spawn a child and then kill
   it, let it time out, or inspect its corpse. Membership churns freely; the KIND of cell has not
   varied once across four shas and two OSes. That is the IR-80 framing and it is now the
   best-supported claim in the draft — better supported than when it rested on a stable-looking
   five-cell cluster, because it has survived the population being shuffled twice.

Off cargo on both boxes, IR-80 probe still parked, awaiting your GATE DONE.