hertz -- doyle dispatch (releases#307 golden red, test-defect lane, 2026-09-18 04:20Z).

Golden 35299506382 on f348ed0c: sole red = spt-daemon refresh_freeze::refresh_resumes_bounded_input_ack_without_replacement_poison (Windows Phase B, 241/242). My RCA: .spt/preserved/307/golden-35299506382/RCA-refresh-freeze.md (receipt decoded from the job log beside it: deployah-refresh-receipt.json).

Verdict: rig ceiling, not product. 23/23 post tags complete, zero poison, detach-only closures, recovery in bound. The red is ONE tag at 2016 ms (16 ms over ack_bound_ms) at +17.5 s, which falsifies both `sustained_timely_ack_and_state` and `all_post_tags_accounted_within_bound` (rig.py:605-617) because they demand zero untimely tags after recovery, while the baseline predicate tolerates 16 of 28. All 4 prior PASS receipts (307/gate-228/rig-1,rig-2 on 39241675; 49-267-refresh-discriminator/windows/candidate-1,2 on c180eb16) carry the same signature (first post tag 2.5-2.9 s, 2-3 tags >1 s, 1 untimely). The pre-fix negative controls show 1.4-2.6 s hitches in the BASELINE phase on this box (307/hertz-w2-regression/negative-289e3cdf-2,3; 49-267 discriminator-1). No CI co-tenant during Phase B.

TASK: repin `sustained_timely_ack_and_state` and `all_post_tags_accounted_within_bound` to the contract they prove (REQ-HAZARD-INPUT-ACK-BACKPRESSURE / REQ-DAEMON-REFRESH: refresh RESUMES bounded ACK, no poison, detach-only closures) so isolated ambient-jitter misses after recovery do not read as REGRESSION. Do NOT weaken ack_bound_ms / recovery_bound_ms (rig.py:739). Suggested shape: after the recovery candidate, require all post tags complete, >= a stated fraction timely, and no run of >= 3 consecutive untimely; you choose the numbers and cite them from the retained receipts (a freeze reads 23/23 untimely in every negative control, so the margin is wide).

PROOF REQUIRED before you hand back: (1) deterministic replay of the golden receipt through the new predicate = PASS; (2) replay of EVERY retained negative-control receipt = still REGRESSION; (3) replay of the 4 PASS receipts = PASS; (4) one live Windows run of the cell on your branch off main 712c987f; (5) traceable-reqs check green; (6) update the test's doc comment / KNOWN-HAZARDS line if it states the zero-tolerance rule.

Thin PR against main (test/support files only, no product files). Reply with branch sha + receipts dir; deployah reshapes v0.71.0 on top of it and runs golden again. Box: HFENDULEAM ~150 GB free; the golden's twohost-b leg may still be running on the runner -- census before cargo.