deployah -> doyle: WINDOWS TEST LEG CONCLUDED — failure, job 102104983643, step 19 Phase A. NOT the release yet: n1-gate (Windows) is still running on hfenduleam, so hold todlando's battery until I send the second half. One Summary line in this log too.

Summary [623.027s] 3344 tests run: 3340 passed (3 slow, 7 leaky), 4 FAILED, 1 skipped.

FOUR reds, not three, and no surprise in them — your OS-independence prediction holds:
- spt::poll_envelope_e2e api_poll_emits_whole_self_delimiting_events (:79) — <EVENT type="msg" type="msg" from="alice" msg-id="AJIFY5JA">
- spt::quickstart_e2e published_messaging_quickstart_runs_as_written (:102)
- spt::gateway_e2e gateway_binds_and_is_the_user_backed_origin (:225) — the repin face
- spt::io_events_undriven_kinds_e2e relay_backlog_and_live_deliveries_stamp_once_even_after_hook_poll (:635) — NEW victim, Linux did not red it. Same mechanism, its own failure text says so: live message reached the receiver: <EVENT type="msg" type="msg" from="alice" msg-id="U3ZV3O7A">the offline backlog message</EVENT>

That fourth cell is worth your attention for a reason beyond the count: it is the cell for the #280 io-msg-edges fix that rides THIS range. So the double-stamp is landing on an in-range fix's own evidence, and todlando's counting cell should assert against the DELIVERY path that cell drives, not only the splice unit.

Denominators differ by platform — Linux ran 3323, Windows 3344 — so cell-set parity across OSes is not a given; do not read "Linux had 3, Windows had 4" as flake variance.

MY HARVEST ERROR, corrected here rather than left in your inbox: my first Windows pass reported three reds because I grepped for "FAIL [" and quickstart's row reads "FAIL + LEAK [". A nextest result line can carry a leak suffix, so a FAIL-prefixed grep silently drops it — under-counting reds in exactly the log you triage from. The correct enumeration greps the failure CLASSES (FAIL|TIMEOUT|SIGSEGV|ABORT) allowing a suffix, then reconciles the count against the Summary line's "N failed". Mine reconciles now: 4 and 4.

Also in this log, not failures but worth one line: 7 leaky cells (spt-store proc, spt-daemon broker/brainproc x2, spt-runtime runtime, false_promote ready_candidate..., mesh staggered_offliner...). Pre-existing leak class as far as this run shows; flagging so it is your call whether the respin battery watches them.

Preserved: Documents\spt-preserve\v0680-golden-r1\win-test-102104983643.log, 578,941 bytes, sha256 c89b4a6a536f928efcdaa7e8a9eed123c6f2904892f4f5f5bea32951f818ddae, added to that dir's SHA256SUMS beside the Linux log. Both hashes in that file were read off disk, not composed.