deployah — CORRECTION, mine, inside 10 minutes of making the claim, and it needs a ruling from you.

I told you "twohost-a and twohost-b are NOT re-run and stand green from a3." THAT IS WRONG. Measured at 01:49:03Z, run_attempt = 4:
  traceability            completed success  startedAt 2026-09-08T18:17:32Z
  test (Linux, kitsubito) completed success  startedAt 2026-09-08T18:17:59Z
  n1-gate (Windows)       completed success  startedAt 2026-09-08T18:42:07Z
  n1-gate (Linux)         completed success  startedAt 2026-09-08T18:47:21Z
  changes                 completed success  startedAt 2026-09-08T18:17:49Z
  test (Windows)          in_progress        startedAt 2026-09-09T01:47:33Z
  twohost-a / twohost-b / notify — ABSENT from the job list entirely
The jobs that are NOT re-run keep their ORIGINAL records and DO still appear, with their old
timestamps. twohost-a/b are not appearing as retained a3 successes; they are gone from the list, which
is the same "count is 6, expect 9, twohost materializes after test" shape the r2 gate notes recorded
from earlier attempts. They gate on `needs: test`, so rerun-of-failed-jobs pulls them in as DEPENDENTS
of the job it re-ran. They WILL re-run and their a3 green will be re-decided.

I should have read the list before asserting; the warning was in my own gate notes.

WHAT THIS NEEDS FROM YOU: your stop condition is worded "a4 red on ANY Phase B cell = STOP, no a5."
A twohost red is not a Phase B cell. So if test(Windows) comes back green and a twohost half reds —
plausibly on the same contention, with seeding live by the operator's choice — the stop condition as
written does not cover it and I will not improvise a ruling on the last attempt. Tell me which:
  (a) any red at all on a4 = STOP, respin (the reading I would default to, since a4 was the last
      authorized attempt at this sha and a twohost red re-opens criteria 5 and 6), or
  (b) a twohost red is separately triageable because the ACL fix is now proven by traffic shape and a
      fresh contention red on a re-run half is not the same class as a Phase B victim, or
  (c) something else you want stated before the evidence lands, so it is not shaped by it.

Acceptance is otherwise unchanged and I hold criteria 5 and 6 to B's own SERVED count and A's helper
cell wall time again, freshly, at a4 — not carried over from a3.

Watch bwy99mxii live and emitting through the real path (its first event carried status, non-terminal
count and the completed set). test(Windows) started 01:47:33Z, so ~45 min, then twohost materializes.