hertz — WINDOW 4 CLOSED. END 2026-09-11T10:26:18Z. Open 10:26:05Z -> 10:26:18Z, 13 s against a 600 s cap. BOX IS FREE. One producer, one attempt, no retry. Post-census ci=0 analyzer=0 agent=0 unknown=0, zero survivors, pool growth 1.25 MB.

RESULT: webserve-e2e-run EXIT 0. `Summary [0.102s] 5 tests run: 5 passed, 0 skipped`. All five by name:
  PASS the_adapter_docs_facet_serves_over_the_production_listener
  PASS production_listener_exposes_only_the_adapter_web_root_through_both_urls
  PASS production_listener_routes_nodes_docs_and_live_registry_references
  PASS production_listener_redirects_directory_roots_before_serving_relative_links
  PASS subnet_node_labels_shadow_only_the_docs_compat_alias_while_members

PREDICTIONS 20-23, EACH MARKED:
20. It compiles. CONFIRMED — built in 5.99 s, first contact with a compiler for 28d4eb68.
21. subnet_node_labels_shadow_only_the_docs_compat_alias_while_members GREEN. CONFIRMED. Third measurement of this case, first pass.
22. THE DISCRIMINATOR, as fixed in advance: webserve_e2e.rs:276 @28d4eb68 — the leaf fetched THROUGH the redirect target — PASSES. CONFIRMED, and it is confirmed by construction rather than by inference: the case runs to completion, and :276 sits ahead of the assertions that follow it, so a failure there would have stopped the case exactly as the previous two did. This SETTLES the open question: the 404 I measured at 71caedd4 was pure index-absence, /llms.txt/docs/ IS where that node's docs live, and my correction at 28d4eb68 was right. The alternative branch I named — redirect target wrong, route as PRODUCT — is closed, not merely unobserved.
23. The other four cases unchanged. CONFIRMED — four passes, so 28d4eb68's blast radius is zero.

VERIFICATION NOTE, because I nearly reported the wrong thing: my first look after launch read a STALE webserve-e2e-run.raw — window 3's file, same path, same panic at :275, same pid 32704 — while this window's producer was still running. I caught it on the timestamp and the empty runs{} in the report, and re-read only after end_utc was written. The artifact path is reused across windows, so a read taken before the producer finishes hands you the LAST window's verdict wearing this window's filename. I am naming it because it would have been a confident, fully-sourced, completely wrong END.

THE START GATE FAILED ITS FIRST REAL LAUNCH AND THAT IS THE OTHER THING TO RECORD. Your word came at 10:24Z; my first launch crashed INSIDE the gate — NameError, `io` never imported — so no producer ran, no START was sent, the box was untouched, and for about a minute you had granted a window that had not opened. The negative control I ran and trusted (no START file -> exit 1, zero producers) 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. That is the same clean-zero shape I have written rules about all morning, wearing the costume of a negative control: a control that only exercises the path you expect to fire certifies nothing about the path that does the work.
Now controlled on BOTH arms — refusal exits 1 with zero producers; send_start() invoked directly returns SENT:doyle / QUEUED:todlando (that was the self-test line, labelled ignore-me on its face). main() is guarded behind __name__ so the module can be exercised without opening a window. This window's START went out as the launcher's first act, recorded in the report as start_dispatch.

WHERE #304 W1 STANDS ON MY SIDE, everything test-only, lane tip 28d4eb68, tree CLEAN:
  producer 1  spt-daemon --lib, 22/22 .......................... GREEN (9f09db4e)
  producer 2  spt --bins, 5 rc detach cases .................... GREEN (61bfd85c) — the three-sha red is closed
  producer 3  webserve_e2e, 5 cases ........................... GREEN (28d4eb68) — this window
  producer 4  attach_link_push_e2e .......................... GREEN (71caedd4) — first real measurement
  docs_server_e2e 1/1 ......................................... GREEN (96325c9b)
  producer 5  webserve_cross_node_e2e ......................... RED BY DESIGN, product, REQ-WEB-PROXY-LOCATION-NAMESPACE, todlando's rider. :592 UNTOUCHED; the int tag is mine to add after his relay fix lands.
  twohost_web.rs .............................................. LABELLED HOLE, compile-gated only here, closed by the #297 two-machine leg in W2.
That is the whole W1 test-side set except the one cell that is waiting on a product fix.

OWED, unchanged: the two traceable findings resolve at assembly, unrepinned. My two misses on the dotted-label case go in the register at W1 close as kin of IR-101, my words, as you ruled.

Box free. todlando, the rider is yours — register REQ-WEB-PROXY-LOCATION-NAMESPACE first, then build in your own pool. When your relay fix lands I need one short window for the cross-node cell against it, and I will file its prediction before the build: the case GREEN with Location /rig-b/docs/, and the one-box hazard I named — same hostname on both homes, so a wrong Location silently serves the WRONG NODE'S BOOK at 200 — becomes unreachable rather than merely unobserved.