TERMINAL: run 34262154550 (sha 25e60015) completed/FAILURE, 9 jobs, 0 non-terminal. I did NOT dispatch the authorized rate rerun. Holding for your triage. Full record: scratchpad r2/HOLD-attempt3-not-dispatched.md

THREE reds, not one.
1) test(Windows) 102216542171 — step 21 Phase B, the known webserve_attachment_e2e cell.
2) twohost-a 102229928746 — step 10 role A (requester) FAILED. two_host_web_helper_role_a panicked twohost_web.rs:968:9 after 900.40s: "no helper record on w2-helper-audience within 900s". 4 passed, 1 failed.
3) twohost-b 102229928689 — step 10 role B (owner) FAILED. two_host_web_role_b panicked twohost_web.rs:559:13 after 910.20s: "role B never got the user's message ADMITTED by A within 900s: broker QUIC op exceeded the 10s bound". 4 passed, 1 failed.

DISCHARGED: starvation is GONE — a and b both created AND started 20:41:23Z on their own runners, and ladder step 9 is GREEN ON BOTH SIDES. Not pairing, not the queue. Disk stays repaired: start+end floors success on all three jobs.

NEW, and this is why I held: B logged 75 x "A not ready … 10s bound" from 20:49:49Z to 21:04:37Z — the whole 900s, never ADMITTED. The CONFIRMED face is SIX bounds then ADMITTED at 63s. 75-and-never is a third face. AND its fix is already in this sha: Rig::id_a_for(offset), per-cell A identity, twohost_web.rs:186 used at :731 (git show at 25e60015). Per-cell identity LANDED and the stall PERSISTED unbounded — by hertz's own one-variable discriminator that is the PERSISTS arm, i.e. product candidate (fresh-endpoint QUIC reachability), not rig timing.

WHY NO RERUN: your authorization was ONE rate rerun on the ttl-race signature. These two are a different signature, they ARE the r2 acceptance cells (twohost-a helper green + role_b from B's own SERVED count) so a run that reds them cannot satisfy acceptance, and they survived their own landed fix. Spending the single attempt on them buys nothing. Your call, not mine to widen.

ALSO, structural: the Windows docs-drift gate SKIPPED A THIRD TIME. Step 38 windows docs gate = skipped, step 37 DISK docs floor (Windows) = skipped, both behind step 21 Phase B which failed. That gate CANNOT run until Phase B is green, so acceptance is blocked by construction, not merely unmet.

Gates I measured (script conditional, NOT fired): foreign non-completed runs 0; free 287604736000 B = 267.85 GiB; new gate 5 parent-chain attribution works — 1 builder, cargo.exe rooted at Runner.Worker.exe = CI, user-rooted 0.

Preserved r2/red/: twohost-a-102229928746.log 319061 B sha256 e2b0297dcbd87870489aa99f12cbbfa1f2f70a330446b5528999292bde28107a · twohost-b-102229928689.log 527394 B sha256 b72748e365e754b28bd48811e521b331142582212e2065918d70eaf871ee85df — hash them your side.

No third attempt at this sha on my initiative. Cascade untouched.