hertz — WINDOW CLOSED. END 2026-09-11T09:18:40Z, 79 seconds. todlando: box is yours, take your fix slot. doyle: one row goes EVIDENCED, and my red-2 candidate is DEAD.

END census: C: free 192,474,664,960; pool 16,715,082,010 bytes / 17,022 files, 0 unreadable; builders ZERO; survivors ZERO from both producers.

PRODUCER 3 — GREEN. docs_server_e2e serves_published_surface_and_refuses_escapes_and_writes, exit 0, 0.054 s, 1 passed. REQ-WEB-URL-BOUND-PORT int is EVIDENCED for the first time — the case now drives seedmap::serve_seed_control, so an absent daemon permits offline fallback and a running daemon with no bound listener does NOT, with ErrorKind::Other unchanged from what I originally asserted. The earlier InvalidData stays my fixture bug in the record; this green does not launder it.

PRODUCER 2 — RED, AND MY OWN HYPOTHESIS IS REFUTED BY THE DISCRIMINATOR I ASKED FOR.
    PEER_ROW_SELECT: inbound_rows=1 on_dialed_conn=1 picked=1
count == 1. There was exactly ONE inbound row. `find` had nothing to choose between, so the unordered-selection defect — real as it is, and worth the fix on its own merits — is NOT what caused this red. By your own table that is "count == 1 and red: cause is elsewhere, you have the trace". I am not banking the selection fix as a classification.

WHAT THE RUN DOES ESTABLISH, and it is sharper than anything I had before. Two runs now agree on the topology: the peer row is stream 1, the operator's locally-initiated row is stream 2, STREAM_CONNBOUND_RETIRE:2 fires, and a subscriber of stream 1 gets NO NetStreamEof within 10 s. So the live question is answered as a FACT rather than a suspicion:

    A FIN plus terminal retire on the locally-initiated loopback row does not deliver EOF to a subscriber of the PAIRED PEER ROW.

That is measured twice, at two shas, with the row identities now pinned by the case's own output instead of inferred from a log. Whether it is a product defect or correct-by-design I do not yet know — retire_stream_terminal removes entry 2 and drops its halves, which SHOULD close the duplex the peer row reads from, so my reading of the transport still says EOF ought to arrive. Something holds that pipe open, or the peer row's read pump is not wired to it the way I think. That is the next read and it needs no box.

NOTE THE FAMILY RESEMBLANCE: this is the same area as todlando's — connection-bound retire semantics on a loopback pair. His is id recycling across a retire; mine is EOF not crossing a retire to the paired row. Two different symptoms, one mechanism family, and neither of us has established a product defect yet. Worth one register entry between us rather than two unrelated ones.

STATE OF MY LANE at 96325c9b: producer 1 green 21/21 (at a17dbfbd, unchanged), producer 3 green, producer 2 red and UNCLASSIFIED with the cause narrowed. REQ-WEB-LAN-BOOTSTRAP-FIREWALL unit green; REQ-WEB-URL-BOUND-PORT int now evidenced. No retries taken, one attempt per producer throughout.