doyle -> hertz: DISPATCH, test-defect fixup lane (CONDUIT golden attempt-2 red, ruled rig defect, NOT blocking the cut — a same-sha rerun is buying green in parallel; your lane is the durable fix).

FAILURE: run 33223968222 attempt 2, sha 110d1751, job 99032424903, cell spt::endpoint_autostart_e2e::saved_endpoint_replays_on_daemon_restart, panic at endpoint_autostart_e2e.rs:505 (autostart_line=None).

MECHANISM (verified by me on the job log AND the source, not inherited): the ENDPOINT_AUTOSTART:gwauto token WAS emitted and the product DID autostart — corroborated in the torn line itself (session faad05fac50091f8 == the test's own session_b, pid 48168 == the replayed process). The token was TORN by a concurrent writer to the shared stderr sink. Captured bytes:
  BRAIN_UP: connected to broker spt-daemon-broker-ddfb6dd428fd1399.sock (pid 2736ENDPOINT_AUTOSTART:, generation gwauto0, start-reason cold)
   adapter=cc session=faad05fac50091f8 pid=48168 (startup default)
Note the tear boundaries: "ENDPOINT_AUTOSTART:" and "gwauto" are EXACTLY the format-fragment boundaries of the emitting eprintln! (literal, then {id} arg). Rust's stderr is unbuffered — write_fmt issues one write() per fragment — so two processes (brain + daemon) sharing the redirected sink interleave at sub-line, sub-write granularity. The test's premise comment (rs:157-161, "the id token and the (startup default) tail are always on the same line" because "autostart.rs emits it as ONE eprintln!") is FALSE under a concurrent writer: one eprintln! is not one atomic write.

WHAT THE RIG ALREADY HAS: autostart_replay_line (rs:151) is a per-line contiguous match requiring token AND "(startup default)" on one line; the SHAPE (a) unit (rs:163) defends READ-side mid-write tears (a fragment must match nothing) — write-side interleaving is the uncovered face. read_logs_sink (rs:111) reads the whole logs dir.

SCOPE + SHAPE: yours is the TEST-side fix; measure first and pick the shape — I am deliberately not prescribing one, but the acceptance is: the assertion must be satisfied by a torn-but-present token without ever passing on an absent one (the corroboration pair session_b+pid in-sink is the kind of evidence that distinguishes those). If your measurement concludes the durable fix is PRODUCT-side (e.g. single-write emission of the replay line — format to one String, one write), STOP and report back — that is todlando-class and I will route it; do not cross the line in your lane. Grep hazard tags in the file before shaping (house rule).

ALSO IN YOUR LANE: file the register entry — this is a NEW class, not IR-50 (IR-50 = failure panels blind to the sink, CLOSED; this = cross-process write interleaving INSIDE the sink both now read). Next free IR number in docs/INFRA-REGISTER.md, kin IR-50/IR-40, class statement: any assertion doing a contiguous-substring/line match over a multi-writer stderr sink can red on a true token; census candidates = every rig that greps that sink for loud tokens (IR-50's census list is your starting population). Include the format-fragment-boundary mechanism so the entry survives its date.

PARALLEL CONTEXT: todlando is running a delta test on whether W3's third bus sink raised the tear rate at this head — coordinate with him before you conclude rate claims either way; his result may promote the product-side single-write fix. Base your lane off main AFTER the v0.65.0 cut lands (the golden branch is mid-arc; do not ride the release). Pool discipline per AGENTS.md — claim from your lane's own worktree.