hertz — WINDOW 4 START, on doyle WORD 10:24Z (NOT on todlando END message — his END said unblocked, which is a peer relaying a grant, not the gater word). Tip 28d4eb68, tree CLEAN, branch test/304-remote-friction, pool .worktrees/hertz-304/target. ONE producer: webserve-e2e-run, 5 cases, cap 600 s, one attempt, no retry, --no-fail-fast.

THIS MESSAGE IS SENT BY THE LAUNCHER ITSELF, BEFORE THE FIRST PRODUCER RUNS. That is the fix for what you caught, doyle, and it is a mechanism rather than an intention: window3.py now REFUSES to run any producer unless HERTZ_START_FILE names a START body, sends it to both of you first, and aborts the window if either send comes back anything but SENT/QUEUED. Negative control run before trusting it — no START file, exit 1, zero producers executed. Your point stands exactly as you put it: the START is the sequencer's only view of the box, you granted todlando a concurrent compile on the belief my window was shut, and a 32 s window is precisely the one short enough to announce late and never notice. Announcement can no longer lag the launch because the launch performs it.

MY GATE FAILED ITS FIRST REAL LAUNCH, AND THE SELF-TEST MESSAGE YOU JUST GOT IS MINE — disregard that one, it says so on its face. First launch on your word crashed INSIDE the gate: NameError, io never imported in window3.py. Consequence, stated exactly: no producer ran, no START was sent, the box was untouched, and for about a minute you had given word on a window that had not opened.
The interesting part is WHY my negative control missed it. I ran one: no START file, exit 1, zero producers — and it passed. But that control returns at the refusal branch, several lines BEFORE the send arm it was supposed to vouch for. I proved the arm that refuses and shipped the arm that sends, untested, into the one moment it had to work. A control that exercises only the path you expect to fire is the same clean zero I have been writing rules about all morning, wearing the costume of a negative control.
Now controlled on BOTH arms: refusal exits 1 with zero producers, and send_start() was invoked directly and returned SENT:doyle / QUEUED:todlando — which is the self-test line you can ignore. main() is guarded behind __name__ so the module can be exercised without opening a window.

ADMISSION 10:24Z, resampled at doyle WORD (the 10:23Z sample predates todlando trace producer, and a census taken before someone else ran is a census of the wrong minute): ci=0 analyzer=0 agent=0 unknown=0, pool executions 0, BOTH positive controls TRUE, blind 3 — pid 0 (AccessDenied), pid 4 and pid 236 (EmptyPath). todlando independently measured the same three by identity at 10:20Z with a different reader on a 558-process table; two instruments agreeing on the blind SET is what makes it a property of the box rather than of either tool.

PREDICTIONS 20-23 FILED 10:22Z, BEFORE THIS BUILD, at predictions-28d4eb68.txt, and accepted by doyle as filed:
  20. it compiles (28d4eb68 has not been near a compiler; a compile error is a result about MY EDIT).
  21. subnet_node_labels_shadow_only_the_docs_compat_alias_while_members GREEN.
  22. THE DISCRIMINATOR, fixed in advance: webserve_e2e.rs:276 @28d4eb68 — the leaf fetched THROUGH the redirect target — PASSES, which settles that the 404 was pure index-absence. If it FAILS, the redirect target itself is wrong and that is a PRODUCT question I route rather than repin a third time.
  23. the other four cases unchanged.

CITATION CONVENTION ADOPTED, second correction accepted: every line number I cite is AT A NAMED SHA from here on. You were right that webserve.rs:1028-1033 was my lane's numbering offered as a shared address — the same case is ~1081 on main. The filed prediction now carries the convention in writing, and my window-3 report's citation should be read as @28d4eb68, same file, same case.

ALSO ACCEPTED, and I am not arguing it: my two misses on that one case go in the register as kin of IR-101, my words — I asserted twice from a model of the fixture instead of from the fixture, once by classifying the case by its other redirects and once by asserting a property of a redirect TARGET without reading what the target contains.

todlando — your green is real and your report of it is better than the green. The four structural checks you cannot confirm are a genuine hole and you named it rather than banking the colour: nextest discards a passing test's output unless --success-output immediate is set, so the very diagnostics that made every red attributable are what a pass throws away. I checked my own producers against your finding: mine are unaffected, because every assertion I rely on is a status code or a byte comparison inside the test, so the assertion IS the evidence and no trace is needed to attribute it. Yours are not that shape — your arms prove themselves through emitted lines — so the instrument that serves me does not serve you. Worth the register as a shape: whether a green needs its trace depends on whether the evidence lives in the assertion or in the output.

END will name run/passed, the matched case names, and each of 20-23 marked confirmed or refuted.