deployah — B2 TERMINAL, re-derived from the terminal artifact (not the live log I quoted earlier).
B3 started 18:38:27Z. DIAGNOSTIC TREE ed8ecd94.

## B2 TERMINAL FACTS
exit 100 · exactly ONE Summary ·
`Summary [1047.917s] 3381 tests run: 3379 passed (11 slow, 4 leaky), 2 failed, 2 skipped`
Population 3381 reconciles with the frozen listing. 165 IR294 lines in the log, 165 extracted,
28 of them the subject's (pid 42784) — a truncated event count BY CONSTRUCTION: the failing path
stops after the continuation, so it never emits the second and third exchanges' stages.
Presence samples 1916, own evidence class, overlap UNVERIFIED. Free space 42.88 GiB.

## THE JOIN HOLDS — hertz's check, run rather than assumed
`transport_request_identity` for subject pid 42784 carries **`"sync_id": "pull-2", "correlated": true`**,
with remote_endpoint_id `da8b6a71...`. The continuation's late observation carries remote_id
`da8b6a71...` — the same endpoint, on the exchange decoded as pull-2. So the late row IS the
subject's pull-2 row, established through the DECODED sync id and not through stream_id 1, which
your own ID0 finding showed is not a unique cross-pull key. Had the join failed I would have
reported the observation as uncorrelated.

## TWO FAILURES, ONE OF THEM OURS
`FAIL [ 49.963s] (1986/3381) spt-daemon::sync two_tier_sync_lands_and_gate_refuses_server_side`
`FAIL [ 10.973s] (  26/3381) spt::daemon_stop_convoy_e2e stop_under_an_api_storm...`
The convoy failure recurs from B1 (9.473s there, 10.973s here) — TWICE now, same cell, different
signature from ours. Still doyle's post-B3 read; I am not chasing it and nothing is filed.

## CORRECTED LABELS, both withdrawals applied
- "nothing was ever late on the wire" — WITHDRAWN (hertz, todlando). The hook is SEND-SUBMISSION
  ENTRY, not a measured wire write. Supported statement: in this execution the acceptance window
  closed while the requester was still in pre-submission preparation; submission entry came
  1.191840s after the expiry event and correlated inbound registration 764us after submission.
- My 18:37:38Z "no new subject traces yet" — WITHDRAWN; the FAIL row was in the same output.

## THE NUMBERS, todlando's to attribute
store-init 9.130515s of a 10.165050s window · four tips 2.212s, straddling the deadline, all four
outcome=missing · expiry 400/400 at 10.165050s, acceptance failed · continuation observed the
matching row 1.196399s after expiry with 28.803599s of 30s unspent · original acceptance preserved
as FAILED throughout.

## READING, per the pre-registration and nothing beyond it
B reproduced an expiry under the SPECIFIED WORKLOAD. That is the whole claim. It does not identify
contention (overlap UNVERIFIED), does not explain the golden reds, and does not unblock the release
— doyle pre-registered that in advance precisely so it could not be argued now. What is new and was
never available before: the failure has an ENDPOINT instead of being right-censored, and in this
execution the overrun sits upstream of submission.

B3 running, 72 rows in. Release HELD, candidate a9e786b2 unmoved.