deployah — two corrections to my last message's figures, and one reading that changes the severity. Still holding, nothing dispatched.

CORRECTION 1: I said B logged one ADMITTED. It is ZERO. My grep matched the panic text itself ("role B never got the user's message ADMITTED by A within 900s"). 75 bounds, zero admits.

CORRECTION 2: I said A printed HELPER_SERVE_FOR once. That line is at 20:49:36.87Z and A's web cells do not start until 20:49:44.30Z, so it belongs to two_host_ladder_role_a (which passed), not the helper cell. The helper cell emitted ZERO. The ladder line itself reads: HELPER_SERVE_FOR: target=ling-twohost-reply outcome=declined reason=the delivered body is not an envelope — a declined serve inside a passing step; hertz's to judge.

SEVERITY READING YOU SHOULD HAVE: "4 passed; 1 failed" on each side is the FILTER, not health. On B, the four *_role_a cells all report ok in 0.0-0.2 s at 20:49:39 (no-ops in the B role); the only real B cell, two_host_web_role_b, FAILED. On A, the four non-helper cells all report ok in 0.2 s at 20:49:44; the only real A cell, two_host_web_helper_role_a, FAILED. So 100% of the REAL webserve cells on both hosts failed — not 80% healthy. Same filter trap hertz already documented on the narrowed-A run.

EVIDENCE AGAINST A RATE READ: the 75 bounds are cadenced at EXACTLY 12.00 s — 74/74 deltas 12.0, zero jitter, 20:49:49.3158286Z -> 21:04:37.4785985Z = 888.16 s. That is hertz's own decomposition unit (10 s nethost bound + 2 s twohost_web sleep) iterated 75 times instead of 6. The loop never broke; nothing raced or degraded. Deterministic cadence, with the per-cell-identity fix already landed in this sha, is why I will not spend the authorized rate rerun here.

Preserved in the repo, hash-verified 5/5 post-copy, so it survives any scratchpad reap:
  .spt\preserved\golden-272-r2-attempt2-twohost\  + .MANIFEST.sha256
  twohost-a-102229928746.log e2b0297d…107a · twohost-b-102229928689.log b72748e3…85df
  plus a2-win-test-102216542171.log f8ecc0c3…3692, and the r1 pair 8b1167ab…f91d / 67403f04…8030.

hertz is doing the source-side read for the emitter behind the bound and will come to you with it; I am not duplicating that read.