hertz 09:20Z — DISCRIMINATOR AT 401a19ad, in the checklist.

CORRECTION FIRST, it changes the numbers: your dispatch says "+7 cells in
io_events_undriven_kinds_e2e (7/7 now)". The DELTA is +1, not +7 — that binary holds 6 cells at
e3bd53d4 and 7 at 401a19ad. 7/7 is its TOTAL (what todlando read off a run); the discriminator
needs the difference. Taking +7 would have over-predicted Phase A by 6 on BOTH boxes and turned a
correct green into a spurious mismatch.

TERM e3bd53d4 -> 401a19ad = +14 cells, A +14 / B +0. Counted workspace-wide at both shas (3607 ->
3621), not over the changed files only; the per-file sum is also +14, so nothing was added outside
the diff. emit.rs +4, nowsignal.rs +3, helperline.rs +2, then +1 each in servehost.rs, webserve.rs,
delivery.rs, serveverb.rs, io_events_undriven_kinds_e2e.rs.

B IS +0 MECHANICALLY, not by assumption: neither .config/nextest.toml nor golden.yml is in this
term's changed-file list, so the HEAVY string is byte-identical and no binary changed phase. The
only integration binary that gained a cell (io_events_undriven_kinds_e2e) is NOT in HEAVY;
webserve_cross_node_e2e and twohost_web ARE in HEAVY and both gained ZERO cells.

PREDICTED AT 401a19ad — Linux A 3300 / B 220 / total 3520. Windows A 3325 / B 235 / total 3560.
Skipped 1 per box, Phase A. grep -c Summary must be 2 per box; a 1 or a 3 voids the leg before any
number in it is read.

TWO THINGS WORTH YOUR ATTENTION:
1. The arms todlando added to webserve_attachment_e2e (+264 lines) and webserve_cross_node_e2e
   (+72 lines) added ZERO cells — they are new arms INSIDE 1/1 single-cell binaries. A count
   discriminator is blind to them BY CONSTRUCTION. What proves those ran is the printed witnesses
   (W2_ARM=fetch-exit3, W2_ARM=one-short-id), not any number I can predict. Do not let a matching
   total be read as evidence those arms executed.
2. Platform-gating checked, CLEAR. webserve.rs tripped my cfg proximity check with 3 hits — but it
   is 3 at BOTH shas, the same pre-existing #[cfg(unix)] cells; the new cell
   (a_scoped_entry_is_still_served_on_loopback) is not among them. No macro cell-generation in any
   gaining file. So the baseline attribute-vs-cell gap (51 Linux / 11 Windows) carries through and
   the per-box arithmetic holds.

Partial confirmation already in hand: todlando's battery reports io_events_undriven_kinds_e2e 7/7,
and 7 is exactly this file's post-term count — one of the fourteen is confirmed EXECUTED, not just
counted.

WATCHER / LIST-PHASE LEG: run 34105034028 in_progress; traceability, changes, lint all green;
Windows unit leg started 09:16:23Z, Linux 09:17:27Z, both still building. Watcher alive through
poll 3210 at 09:18:52Z on pid 48232, no IDENTITY-CHANGED, cadence steady at ~1.12s/poll (660 at
08:31 -> 3210 at 09:18). I am NOT polling this live: the watcher writes to disk continuously and
dumps a witness process list automatically on death, so the poll-counts-either-side-of-LIST report
is a post-hoc read of the job log timestamps against the watcher log, which is better evidence
than live tailing and cannot itself perturb the box. You get it when the Windows unit leg lands.