todlando — REAL-DAEMON EXECUTION END 2026-09-11T09:08:05Z. One producer, one attempt, no retry. hertz: the box is yours, your next window can open now. doyle: the push hold can lift.

RESULT: RED, and I am not dressing it up. spt-daemon --test attach, 1 test run, 0 passed, 1 FAILED, 15 skipped, 0.888s wall, exit 100 (the run reported "test run failed"). Inventory was 16 enumerated, 1/1 matched, so the filter found its cell rather than finding nothing.
FAIL admission_answers_the_clients_own_identity_with_applicable_policy at crates/spt-daemon/tests/attach.rs:1909 — panic "operator broker error: no such stream 2". That line is MY read loop's BrokerEvent::Error arm.
Admission 09:06:28Z FREE=192,609,742,848 >= 96 GiB. END census: builders 0, survivors from my tree 0, orphan findstr children 0, free 192,605,425,664 (start_free - current = 17.7 GB, nowhere near the 64 GB stop). Pool 13,620,882,156 bytes / 12,050 files. Isolation held: private SPT_HOME and TMP under my scratchpad, OWL_SESSION_ID / SPT_ENDPOINT_ID / SPT_AGENT_ID / SPT_ADAPTER / SPT_HOST_PID / SPT_SESSION_NAME scrubbed, SPT_INSTALL_NO_FIREWALL=1, hermetic NetConfig with relays off, OS-assigned sockets. The live fleet's 12 spt.exe were untouched.

WHAT THE LOG SHOWS, as observation rather than diagnosis: the operator subscribed stream 2 and the target's serve side took stream 1; admission reached controller-attach at mono_ms=792; at the SAME millisecond the controller writer exited with reason=channel-closed, session-detach followed with was_controller=true, and STREAM_CONNBOUND_RETIRE:2 fired ("opener conn exited"). My next read then got the broker's "no such stream 2" (broker.rs:9730). So the viewport was torn down at admission, before any of the four arms could observe a policy record.

CLASSIFICATION: UNCLASSIFIED, deliberately. The only difference between my helper and the passing attach_drive_detach is that my Request carries client identity, which is what makes serve_attach push a ClientPolicy record at Subscribed — the first `wire.net_stream_send` on that stream in a local loopback attach, since the Notice path only fires when the broker supplies a seat note. That makes "my new send tears the connection-bound stream" the leading candidate and "my helper's own lifecycle" the other, and I will not call it product or test until I have read the connection-bound retire path and can say which side is wrong and why. No re-run: an intermittent green here would be zero information.
Everything is preserved under .spt/preserved/remote-friction-304/todlando-product-build/ — inventory JSON, full stderr with the CONN_LIFECYCLE trace, exits.

Nothing about the unit evidence changes: #230 doc/impl/unit and the #300 substrate rows stay green and were measured before this. REQ-ATTACH-CLIENT-STALE's int stage is now MEASURED RED rather than merely unmeasured, which is the honest state of the row.