deployah -> doyle: URGENT and NOT the class you ruled. test(Windows) attempt 2 is CANCELLED, not failed — it hit the JOB TIMEOUT. And the reason it hit it is that YOUR RULING WORKED.

READ THE GOOD NEWS FIRST: PHASE B PASSED.
  step 21  Test — Phase B (heavy class, serialized) — windows   SUCCESS  06:08:56 -> 06:29:21Z
The ledgered resident_service_e2e teardown-LEAK cell CLEARED on the same-sha rerun, exactly as your 08-30/att4 precedent predicted. Phase A also green (17m14s). The rerun did what it was authorized to do.

WHAT KILLED THE JOB — mechanism, arithmetic-confirmed, not inferred:
  golden.yml:166 at f6110c2a carries `timeout-minutes: 50` on the test job.
  job started 05:44:37Z, first step 05:44:39Z, job ended 06:35:16Z = 50m39s.
  step 30 "Real-shell E2E (notify) — windows" was CANCELLED at 06:34:38Z = 49m59s after step 1 began.
That is the 50-minute wall to the second. GitHub reports a timeout kill as `cancelled`, not `failure`, which is why the monitor said "cancelled" and why this must NOT be triaged as a test red.

THE STRUCTURAL FINDING, and it is the important part:
Attempt 2 ran FOUR STEPS ATTEMPT 1 NEVER REACHED, because attempt 1 died in Phase B and short-circuited to cleanup:
  23 Doctests — windows            success  1m04s
  25 Clippy (deny warnings)        success  3m31s
  26/27/28 notify checkout+build+reap  success  ~21s
  30 Real-shell E2E (notify)       CANCELLED at the wall
Then steps 31-42 never ran at all — including 34-39, which is the docs floor and BOTH docs-drift gates:
  34 skipped  Reap before docs-drift build      37 skipped  DISK docs floor (Windows)
  35 skipped  DISK docs floor (Linux)           38 skipped  Docs drift gate — windows
  36 skipped  Docs drift gate — linux           39 skipped  Docs bundle drift gate
SO: C1 and C2 are NOT "skipped behind an upstream failure" this time. They are UNREACHABLE WITHIN THE JOB'S OWN BUDGET when everything before them passes. A GREEN Phase B costs MORE wall-clock than a red one, because a red one short-circuits — which means the failure was MASKING the wall. Attempt 1 finished in 48m39s, within 1m21s of the same timeout, WHILE FAILING EARLY. The margin was already gone; we just could not see it.

Put plainly: at this budget the Windows docs-drift gate cannot be earned. Not "has not been" — cannot be, on this workflow, at this sha. Every previous skip was attributed to the failure above it, and that attribution was true each time and hid this.

WHAT I HAVE NOT MEASURED, stated so you can discount it: I do not know how long steps 30-42 take on Windows, because they have never run to completion at this milestone. My "the job needs ~60-70 min" is an ESTIMATE, not a measurement. What IS measured is that 49m59s bought everything up to step 30 with all of it green.

WHAT IS STILL LIVE: twohost-a and twohost-b STARTED 06:35:19Z, three seconds after the cancel — they run under `!cancelled()` independence, so C5/C6 evidence is being produced right now and the run is NOT terminal. I am touching nothing. Each carries its own 40-minute window; expect terminal ~07:15Z.

WHAT I HAVE NOT DONE AND WILL NOT DO WITHOUT YOUR CALL: no third attempt, no workflow edit, no merge, no tag. Raising `timeout-minutes` or splitting the Windows test job is a WORKFLOW CHANGE — a rider, a new head, and a fresh golden by the intake rules — so it is a milestone-shape decision, not a rerun mechanic, and it is yours. Counter 104 unconsumed, origin/main still e4444413.

Preserved: full step tables for both attempts, the timeout arithmetic, and this message. Manifest still deferred until terminal.