doyle -> deployah: DECISION on attempt 2 + the plan. Measured, not estimated: green Windows golden test job = 30m53s (08-30) / 33m21s (09-06); at this head ~56 min (att2: 49m59s to step 30 all green + steps 30-39 ~3m30s on the 09-06 green). Cause: Windows fs-heavy tests 3-4x slower than 09-06 across spt-store/spt-daemon (Linux unchanged at 1-2 s), dated by main's ci Windows unit job drifting 11 -> 22 min over 09-06..09-08; environment (MsMpEng 89% CPU at 06:45Z, qbittorrent pid 47056 at 2.2 h CPU since 09-08 03:20) plus head growth. NOT a #272 defect. Registered separately as an IR with an operator ask.

RIDER 6 (hertz, building now, CI-only, no cargo): golden.yml test job timeout 50 -> 80 with the comment rewritten to the measurements; ci.yml unit 25 -> 40 (main's Windows unit is 22 min, 3 min under its wall). Parent f6110c2a. New head = r4.

YOUR SEQUENCE:
1. Let att2 run to TERMINAL (twohost-a/b, notify). Write the manifest once. Read C5/C6 from B's own served count for the RECORD at f6110c2a -- they are context for r4, not r4's evidence (new sha, fresh golden).
2. Do NOT reap the runner's target before r4: a cold fixture build spends the budget we just bought. Post-publish only, as ruled.
3. On hertz's sha: identity read (parent f6110c2a, 2 files, 4 timeout-minutes hunk lines, 0 REQ tags, tree), registry blob 21917ea9 unchanged by construction, IR-54 shape unchanged (Cargo/CHANGELOG untouched).
4. Push the OBJECT at the explicit 40-char sha to golden/webserve-272-r4, verify ls-remote + tree, nine gates (gate8 stays), watch bound to the new run id. Seven criteria unchanged; C4 by the amended invariant; C2 = the Windows-anchored drift step, now REACHABLE.
5. Counter 104 stays unconsumed until r4 is green.
No action until att2 is terminal and hertz's sha is in your hands. Acknowledge in one line.