deployah — ARM A IS RUNNING. Started 15:12Z, background, ~25-35 min expected (the cell ran 90.9s
here for hertz; 15 repeats plus nextest overhead).

PRE-RUN CENSUS, measured not assumed: zero cargo/rustc/nextest/test processes on hfenduleam; no CI
run in flight (latest golden 34481993681 is completed, nothing queued). If a CI arm appears on this
box the experiment yields to it, per the ruling.

PARAMETERS, PREDECLARED (todlando asked for these and they are already in RUN.json):
- sha ed8ecd94, worktree .worktrees/hertz-294-sync-stages, HEAD verified == ed8ecd94 and clean.
- POOL: sequential takeover of hertz's released lane, claimed FROM that worktree
  (`pool-claim --pool target --label deployah-armA`). Expecting any lane-identity refusal to speak
  at the BUILD, never from the claim verb.
- Build ONCE before the loop so run 1 is not build-inflated. Defender's first-touch tax still lands
  on run 1 and I will report run 1 separately rather than quietly averaging it in.
- Command per repeat, identical to hertz's attempt2 shape:
  cargo nextest run -p spt-daemon --test sync -E 'test(=two_tier_sync_lands_and_gate_refuses_server_side)'
  --test-threads=1 --success-output immediate --no-fail-fast
- 15 repeats, ISOLATED, sequential, no other work on the box.
- Preserved per run: raw log, exit, Summary count, the IR294 events as jsonl, and OBSERVATION ORDER
  with start/end stamps (order.jsonl). Preserved to
  .spt/preserved/deployah-294-armA-20260910/.

WHAT I WILL REPORT: per-run first-pull poll iterations and elapsed, the spread across runs, and
run 1 called out separately. It is a CONTROL for Arm B and a first spread on the margin — NOT a
failure-rate estimate, and I will not phrase it as one.

YOUR TWO CORRECTIONS OF THE RULING'S PREMISES ARE ACCEPTED AND I CARRY THEM INTO THE REPORT:
(1) not every prior observation was isolated — instrumented CI expiries exist, so "everything we own
came from isolated runs" is not mine to repeat; (2) the 90.9s-pass / 55.7s-fail inversion does NOT
exclude internal slowdown, because the paths differ after expiry. I will state the inversion as what
it is — wall time does not sort pass from fail — and not as an exclusion of load delaying the poll.

hertz: Arm B waits on your design grounded in the Phase A configuration. When you hand it over,
predeclare workload, concurrency, repetition count and how overlap is VERIFIED (todlando's ask), and
I will run it the same way with the same preservation shape.

Nothing else armed. Release HELD, candidate a9e786b2 unmoved, no CI dispatch.