deployah — ARM B COMPLETE, all three terminal, re-derived from terminal artifacts. Full report:
.spt/preserved/deployah-294-armB-20260910/ARM-B-REPORT.md. Attribution is named per author below,
since this body goes to doyle, hertz and todlando and "you" resolves differently for each reader.

## OUTCOMES (each exit 100, exactly ONE Summary, population 3381 reconciling with the frozen listing)
B1 18:02:28-18:19:25Z  Summary  973.295s  3380 passed (9 slow, 5 leaky), 1 failed   subject PASS 99.454s
B2 18:19:38-18:38:10Z  Summary 1047.917s  3379 passed (11 slow, 4 leaky), 2 failed  subject FAIL 49.963s
B3 18:38:27-18:58:00Z  Summary 1108.559s  3380 passed (10 slow, 5 leaky), 1 failed  subject PASS 92.136s

## SUBJECT, keyed invocation + test + pid BEFORE any exchange join (doyle's binding rule)
B1 pid  7784, 65 events: 182/400 @4.605s · 20/400 · 21/400 · store-init 1.869609s
B2 pid 42784, 28 events: EXPIRY 400/400 @10.165050s          · store-init 9.130515s
B3 pid  2228, 65 events: 306/400 @7.748s · 22/400 · 19/400   · store-init 5.739629s
B2's 28 is a truncated count on a truncated path, not data loss — doyle's point, and the failing
path never emits exchanges 4 and 6.

## B2 REMAINS THE ONE REPRODUCTION, WITH THE CORRELATION PROVEN
sync_id "pull-2", correlated:true, remote da8b6a71... matches the continuation's late row. Not
stream_id, which hertz measured as non-unique across pulls (QUIC wire-stream id 0 reuse).
Expiry 48.2580; SEND-SUBMISSION ENTRY 49.4498 (1.191840s after expiry); correlated registration
764us later (todlando's figure); continuation observed at 1.196399s past expiry with 28.803599s of
30s unspent; acceptance preserved FAILED. Overrun localized UPSTREAM OF SUBMISSION.

## THE CONVOY CELL FAILED IN ALL THREE — AND THE DURATION GREW EACH TIME
9.473s (B1) · 10.973s (B2) · 14.553s (B3). doyle asked for the B3 row either way; it is a third
occurrence, not a first absence. Different signature, nothing filed against #294, disposition is
doyle's.

## SUITE WALL GREW MONOTONICALLY: 973.295s, 1047.917s, 1108.559s
Recorded as an observation with NO mechanism attached. It is consistent with the A-then-B drift
hertz named as a confounder in advance, and I am not claiming that is what it is.

## SIBLINGS, kept separate and NOT pooled with Arm A (todlando's ruling)
concurrent_writes_reconcile... first pulls 199 / 147 / 166 of 400
torn_pull_recovers_by_repulling   first pulls 116 / 133 / 266 of 400

## THE ONE BOUNDED COMPARISON
Subject store-init B: 1.869609s / 5.739629s / 9.130515s. Arm A's 15 isolated draws: 0.196-3.497s.
Two of three B values lie outside A's entire observed range — observed values only. Overlap
UNVERIFIED, confounders stated not corrected, no load attribution, no rate, no pooling.

## STANDING, unchanged and pre-registered before any of this ran
Reproduction under the SPECIFIED WORKLOAD is the whole claim. Finding 1 UNCLASSIFIED as to cause.
The golden reds are unexplained. Sequencing stays refused. No budget change. THE RELEASE DOES NOT
UNBLOCK. Candidate a9e786b2 unmoved, core#218 unmerged, v0.69.0 not cleared, no CI dispatch.

The arm is now doyle's for the convoy read and the verdict; todlando has all three runs' traces for
the paired attribution with per-run pairing preserved.