hertz — WINDOW 3 predictions, filed 2026-09-11T10:15Z, BEFORE any build.
Lane tip 71caedd4 (= 61bfd85c + 4ec34d1c repin + 71caedd4 pin bump).
Producers, in order: mock-session-fixture, attach-push-run (producer 4),
webserve-e2e-run (producer 3). One attempt each, no retry, --no-fail-fast.

== P0. mock-session-fixture ==
  cargo build --locked -p mock-adapter --bin mock-session

14. EXIT 0, and target\debug\mock-session.exe EXISTS afterwards. -> PREDICTED
    Rests on: common/mod.rs:253-258 maps the fixture name "mock-session" to
    package mock-adapter, and sibling_bin (:264) looks for it beside
    CARGO_BIN_EXE_spt — i.e. target\debug\. A plain `cargo build -p` puts it
    exactly there. This is the step whose ABSENCE, not whose failure, cost
    producer 4 its previous attempt.
15. It is the ONLY cross-package fixture this case needs. -> PREDICTED
    Checked rather than assumed, because failing twice for one class of reason
    is the whole thing I am trying not to do: attach_link_push_e2e.rs has
    exactly ONE sibling_bin call (:188, "mock-session"). The psyche binary at
    :202 is a COPY of CARGO_BIN_EXE_spt — same package, which cargo already
    guarantees to its own integration tests. So no second fixture is hiding.
    IF A SECOND FIXTURE IS NAMED ANYWAY: my grep was the wrong population and
    that is the finding, not the missing binary.

== P1. producer 4 — attachment_frames_reach_a_linked_shell_through_the_real_daemon ==

16. GREEN. -> PREDICTED, and I am RAISING confidence from my original filing,
    with the reason, because the doubt I named has been answered from source.
    ORIGINAL DOUBT (prediction 6, 61bfd85c): the clause ends a REAL `spt rc`
    viewer by piping ordinary keys plus the ctrl-b d chord into its stdin, and
    I did not know whether a piped stdin reaches rc's BYTE reader or its
    Windows console-key path. I called it "lowest confidence of the set".
    NOW SETTLED, rc.rs:1405-1417: spawn_stdin_reader branches on
    std::io::stdin().is_terminal() — terminal goes to the EVENT reader, and
    everything else falls through to spawn_stdin_reader_bytes. A pipe is not a
    terminal, so the real binary takes the BYTE path deterministically, by an
    explicit branch, not by luck. That path is parse_stdin_chunk, which is the
    exact state machine viewer_byte_detach_keeps_controller_and_child proved
    GREEN in-process at 61bfd85c 1.058 s ago. The channel question is closed.
    WHAT IS STILL UNPROVEN, named so this is not a clean guess: that the real
    daemon/session lifecycle around it behaves — viewer exits 0 AND the row is
    removed. That is the part no in-process case covers.
    IF IT REDS: the discriminator is WHERE. A non-zero rc exit with the viewer
    GONE is a teardown/exit-code defect. A zero exit with the viewer STILL
    PRESENT is a removal defect. A TIMEOUT with neither is the channel
    question re-opening, which would mean is_terminal() answered true for a
    pipe and my reading of :1409 is wrong.

== P2. producer 3 re-run — webserve_e2e, 5 cases, at 71caedd4 ==

17. It COMPILES. -> PREDICTED, and it is a real prediction, not a formality.
    4ec34d1c was written outside a window and has never been near a compiler.
    A compile error is a result about MY EDIT and I will report it as that,
    not as a result about the case — the same shape todlando declared for his
    first build. The edit adds one let-binding, two asserts and one http call,
    all copied from the idiom two lines above in the same function.
18. subnet_node_labels_shadow_only_the_docs_compat_alias_while_members -> GREEN
    Rests on: /llms.txt/ is a node root with an EMPTY remainder, which is the
    #301 arm at webserve.rs:778-781, so it answers 302 with Location
    /llms.txt/docs/ — and the repin now asserts exactly that, then follows it
    to 200. The five http calls the old panic MASKED were read at
    :267/:270-271/:277-280/:283-287 and none of them names a node root with an
    empty remainder, so there is no second stale assertion behind this one.
    IF IT REDS AT A LATER LINE THAN :266: my masked-region read was wrong and
    the case had two stale assertions, not one. That is the interesting
    outcome and I would rather find it here than at assembly.
19. The other four cases GREEN again, unchanged. -> PREDICTED
    They passed at 61bfd85c and 4ec34d1c touches one hunk in one case. If any
    of the four moves, the repin had a blast radius I did not predict and the
    fix is wrong even if case 2 goes green.

== NOT IN THIS WINDOW ==
producer 5 / webserve_cross_node_e2e: doyle ruled it PRODUCT
(REQ-WEB-PROXY-LOCATION-NAMESPACE, todlando's lane, fix at docshost.rs
proxy()). The cell at :592 is the int evidence and is NOT repinned; I add the
[int->REQ-WEB-PROXY-LOCATION-NAMESPACE] tag on that assertion after his fix
lands, and the cell runs in a LATER window against it. No prediction here.

twohost_web.rs: still a LABELLED HOLE, compile-gated only on one box, closed
by the #297 two-machine leg in W2. No prediction is possible and none offered.