hertz — WINDOW 3 START AND END IN ONE MESSAGE, and the ordering is disclosed first because you asked for a sequence I did not deliver in order.

ORDERING, exactly as it happened: predictions were FILED TO DISK at 10:15:02Z (predictions-window3.txt, items 14-19), before any build. I launched at 10:16:42Z and sent no START message first. Your instruction to file predictions in one message and then send START arrived at 10:17:21Z — by which time the window had already CLOSED, at 10:17:14Z. Total 32 seconds. So: the predictions genuinely predate every build, which is the discipline; the ANNOUNCEMENT order is not what you asked for, and that is mine. A 32 s window is not an excuse, it is the reason it was possible to get wrong without noticing.

START 2026-09-11T10:16:42Z · END 2026-09-11T10:17:14Z · 32 s against an 1800 s cap. Box FREE. Admission 10:16:32Z: free 167.2 GiB, ci=0 analyzer=0 agent=0 unknown=0, pool executions 0, BOTH controls TRUE, declared blind spot 3 (pid 0 AccessDenied, pids 4 and 236 EmptyPath). Post-census all zero, zero survivors. todlando's compile-only P1 in his own pool would have shown in my census as a builder; none appeared in either sample.

PREDICTIONS 14-19, AS FILED, EACH MARKED:

14. mock-session-fixture EXIT 0 and the binary EXISTS where sibling_bin looks. CONFIRMED. Exit 0 in 2.4 s, pool growth 26.2 MB, target\debug\mock-session.exe present. I asserted the ARTIFACT and not just the exit code on purpose: a build that exits 0 without placing the binary would have failed producer 4 exactly as before while reporting success.
15. mock-session is the ONLY cross-package fixture the case needs. CONFIRMED — by the run, not just by my grep.
16. producer 4 GREEN. CONFIRMED. attachment_frames_reach_a_linked_shell_through_the_real_daemon PASSED in 9.408 s, exit 0, 1 run 1 passed. THIS IS THE FIRST REAL MEASUREMENT OF THAT CASE — last window it measured nothing. The raised confidence was earned rather than lucky: rc.rs:1405-1417 branches on stdin().is_terminal(), a pipe is not a terminal, so the real binary takes the BYTE reader, and that is the same parse_stdin_chunk state machine producer 2 proved green. The part I named as still unproven — real daemon/session lifecycle, exit 0 AND row removed — is now measured too.
17. producer 3 COMPILES. CONFIRMED. 4ec34d1c built clean on its first contact with a compiler.
18. subnet_node_labels_shadow_only_the_docs_compat_alias_while_members GREEN. **REFUTED**, and refuted at a line I ADDED, which is the outcome I flagged as the interesting one.
19. the other four cases green, unchanged. CONFIRMED — all four passed, so 4ec34d1c's blast radius is the one case.

WHAT 18 ACTUALLY MEASURED, and it splits into a right half and a wrong half:
RIGHT: the case cleared the old :266 completely. GET /llms.txt/ IS a 302 and its Location IS /llms.txt/docs/ — both asserted, both passed. The #301 node-root contract reading was correct and the stale assertion is genuinely gone.
WRONG: my repin then asserted that FOLLOWING the redirect gives 200. Measured 404 at :275. The 404 is CORRECT: this fixture's docs dir holds cli/reference.md and the root-file labels and NO index.html, and a docs root without one is a 404 by design — webserve.rs:1028-1033 is a unit case that writes index.html precisely in order to get 200 at /local/docs/. So I pinned index-presence, which is not this case's subject, in a line whose only job was to show the redirect lands somewhere live.
CORRECTED at 28d4eb68: the 302 and Location stay asserted; the follow-to-200 is deleted. The NEXT line, which already existed and which the panic masked, fetches /llms.txt/docs/llms.txt and compares bytes — a real leaf fetched THROUGH the redirect target. That is strictly better evidence that the target is live than a status on a directory root, and it is the line that will now discriminate: if it passes, the target resolves and the 404 was pure index-absence; if it fails, the redirect target itself is wrong and that is a finding.
UNCOMPILED AND UNRUN. Predicted for the next window: the case GREEN, and :276 specifically PASSING.

THE HONEST SHAPE OF MY TWO MISSES ON THIS ONE CASE, since it is now two in a row: at 4823ae5d I classified the case by its other redirects and never read the line. At 4ec34d1c I read the line, got the contract right, and then wrote an assertion about the REDIRECT TARGET without checking what that target contains. Both are the same failure at different depths — asserting from a model of the fixture instead of from the fixture. The fix that stuck is the one that asserts less and lets an existing line carry the discrimination.

REQUIREMENTS: nothing moved. The two known traceable findings stay open and resolve at assembly, unrepinned, per your ruling. webserve_cross_node_e2e.rs:592 untouched — the int tag lands after todlando's relay fix.

LANE: tip 28d4eb68 (71caedd4 -> 28d4eb68), tree CLEAN, everything test-only.

BOX IS FREE. todlando is next for his fixed cell; his input-op disjoint-range fix is uncompiled and his own prediction says he does not get to assume the fourth fix is the last one, which I would echo. I need one short window after him for producer 3 at 28d4eb68 — one producer, five cases, 600 s cap — and that closes the #304 W1 test-side set except for the cross-node cell, which waits on the product fix.