deployah — B1 TERMINAL, measured 18:22:47Z. DIAGNOSTIC TREE ed8ecd94, not the candidate. B2 is
already running. Artifacts: .spt/preserved/deployah-294-armB-20260910/B1/.

## B1 OUTCOME
exit 100 · exactly ONE Summary · `Summary [ 973.295s] 3381 tests run: 3380 passed (9 slow, 5 leaky),
1 failed, 2 skipped`. Selected population 3381 reconciles with the frozen listing exactly.

## THE SUBJECT DID NOT EXPIRE
`PASS [ 99.454s] (2019/3381) spt-daemon::sync two_tier_sync_lands_and_gate_refuses_server_side`,
flagged SLOW > 60s. Its first pull: **182/400 iterations, 4.605s** — 45.5% of the acceptance budget.
Later pulls 20/400 and 21/400. Stage coverage complete: exactly 65 events for the subject pid.
That 182 sits INSIDE Arm A's observed range (45-300, median 153). One draw under load; I am not
comparing distributions off a single B run.

## THE ONE FAILURE IS NOT THIS SIGNATURE
`FAIL [ 9.473s] ( 27/3381) spt::daemon_stop_convoy_e2e
stop_under_an_api_storm_stays_down_then_start_brings_up_exactly_one`, plus 5 LEAK rows
(contract_e2e live_agent_lifecycle_e2e, brainproc stale_generation_minus_one, livehost
legacy_psyche_sweep_guard, legacy_resident_sweep_e2e, false_promote). Per hertz's protocol, other
failing cells do NOT constitute a reproduction of the sync signature, and I am filing nothing about
them here beyond their preservation.

## A CORRELATION TRAP I HIT AND CLEARED — worth knowing before you read the traces
The B1 log carries **202 IR294 events, not 65**: TWO SIBLING SYNC TESTS also run the instrumented
helper and emit on the same stream — `concurrent_writes_reconcile_on_elected_node_and_conv`
(pid 18092, 4 polls: 199, 61, 19, 49) and `torn_pull_recovers_by_repulling` (pid 60612, 3 polls:
116, 48, 33). A naive "first poll.success in the log" read would have attributed **116/400 to the
subject**, which is another test's number. Separated by the events' own `test` and `pid` fields;
the subject is pid 7784. Arm A never showed this because its filter selected one test.

## WHOLE-CELL WALL, both completed passes so comparable in the way doyle allowed
subject under this workload 99.454s vs Arm A's 43.942-70.864s across 15 runs. Above A's maximum.
No mechanism attached to that, and the first-pull margin is the quantity the arm is about.

## LABELS
OVERLAP UNVERIFIED. 1765 sampled presence records preserved with stamped query intervals and
(PID, birth, image) identity — its own evidence class, NO concurrency claim attached.
Confounders stated, not corrected: A-then-B ordering, B's prebuild and cache changes, sampler
overhead, not a randomized comparison.

## METER DEFECT FIXED (doyle caught it, not a blocker, B1 not restarted)
`07-selected-total` had recorded **0** because my count pattern expected indented lines and
`nextest list` emits `<binary-id> <test-name>` unindented. Corrected to 3381 from the listing, with
`07-selected-total.NOTE` recording the mechanism in place. Same class as the empty events files: a
count describing the QUERY, not the world. The listing was always the authority and the suite's own
"Starting 3381 tests across 138 binaries" corroborates it.

B2 running, B3 to follow. todlando: subject traces are in B1/events.jsonl, pairing preserved per run.