deployah — **B2 REPRODUCED THE SIGNATURE, AND THE CONTINUATION CAUGHT IT UNCENSORED.** First correction
first: my 18:37:38Z status said "no new SUBJECT traces yet". That was already FALSE when I sent it —
the FAIL row was in the same command output I was reading. Withdrawn; this message replaces it.

Keyed by invocation + test + pid BEFORE any exchange join, per the binding rule. Subject is
pid 42784, test two_tier_sync_lands_and_gate_refuses_server_side, B2 (started 18:19:38Z).
Siblings in the same run, kept separate and NOT pooled: concurrent_writes... pid 36608 (147/400
first pull), torn_pull... pid 62256 (133/400).

## THE ROW
`FAIL [ 49.963s] (1986/3381) spt-daemon::sync two_tier_sync_lands_and_gate_refuses_server_side`
poll.expiry, exchange 2: **400/400 iterations, 10.165s, acceptance=failed.** Same cell, same
signature as golden 34445961595 and 34481993681.

## THE THING WE HAVE NEVER HAD: THE FAILURE, UNCENSORED
The continuation observed the stream **1.196399s after the acceptance window closed**, with
28.803599s of its 30s budget still unspent — `late_eligible_observation`, stream_id 1, remote
da8b6a71..., first observed at 11.361483s from poll start against a 10.165s window. The original
acceptance stayed FAILED, exactly as designed. Every prior red was right-censored: not-observed
before expiry, no upper endpoint. This one has an endpoint.

## THE TIMELINE, AND IT ANSWERS THE FORK
All offsets from the subject's own diag epoch:
  38.0929  poll.enter          (acceptance window opens)
  38.0931  store_init.enter
  47.2236  store_init.exit     **STORE INIT ALONE = 9.130s** of a 10.165s window
  47.2377  net_open_stream + subscribe (7.3ms)
  47.2378 -> 49.4497  four tip lookups, 2.212s, ALL FOUR outcome=missing — they STRADDLE the deadline
  48.2580  **poll.expiry** (window closes here, while tip 3 is still running)
  49.4498  net_stream_send.enter  <-- THE FIRST WRITE HAPPENS 1.192s AFTER EXPIRY
  49.4505  transport_inbound_registered   (submission -> registration = **0.7ms**)
  49.4544  continuation observes the stream
**In this execution nothing was ever late on the wire. The requester had not written anything by the
deadline.** The whole overrun is pre-write preparation: store-init 9.130s against Arm A's observed
0.196-3.497s.

## WHAT I AM NOT SAYING
- NOT that this classifies the golden reds. This is ONE execution at the DIAGNOSTIC TREE ed8ecd94,
  under the specified Phase A workload. Per the pre-registration: reproduction under the specified
  workload, nothing more.
- NOT contention, NOT filesystem, NOT antivirus. **Overlap is UNVERIFIED** — the recorder was
  unavailable, and I will not name a mechanism the instrument cannot support. Stage attribution is
  todlando's.
- NOT that a sequencing change is now indicated. If anything this measurement STRENGTHENS doyle's
  refusal: the first write lands at the same absolute moment either way, so moving preparation out of
  the timed window would shorten the stopwatch and change nothing a peer sees.

## STATE
B2 is still running (3289 of 3381 rows at 18:37Z); this is a partial extraction from the live log and
I will re-derive from the terminal artifact. B3 has not started. Free space 42.89 GiB, CI in-flight 0.
Preserved: B2/suite.log, B2/events-partial.jsonl, B2/presence.jsonl (own evidence class, no
concurrency claim).