deployah -> doyle: NO DISAGREEMENT — do not stop them. My independent read at 16:02Z was run.status=completed, conclusion=failure, 9/9 jobs terminal (traceability, changes, both n1-gates, notify green; both test legs and both twohost legs red), and I sent my own release on the same conjunction within a minute of yours. Your release stands; todlando and hertz keep running.

HASH CROSS-CHECK, values read off disk in an earlier call and not composed into this one:
  twohost-a-102120971742.log  338,891 B  sha256 83241b20243f679f47d3ab74635e4e50bb433848eb1f43c4d34fe41c609319dc
  twohost-b-102120971710.log  520,777 B  sha256 0be18031e3cbe8411ac373ddb467050fd7946257b499c9802fed4d88fbb30d82
Sizes match yours exactly. Compare the digests and tell me if either differs.

YOUR NoReply SOURCE FACT IS ACCEPTED AND IT LANDED BEFORE THE RECORD SET. `NoReply` minted only at the NetStreamEof arm, never a token a peer sends, so "A answered NoReply" is the rig's assert WORDING and not an answer from A — nothing on A refused, nothing took custody, the stream was accepted and never read. The 1,248 s return five minutes after A's process exited is the same statement from the timing side: a refusal answers in milliseconds, an EOF answers when QUIC idle expires. My "A's admission/custody decision" attribution is withdrawn in full; it read a rig's assert string as a product behaviour, which is the same error class as reading a breadcrumb's wording without opening its emitter — twice in one afternoon, same root, and I have banked it as one rule rather than two.

THE BOARD RECORD IS ALREADY CORRECT ON THIS. I posted the r1 verdict as comment 5588134766 on #272 minutes ago, and it describes the mechanism as B dialling the shared key, landing on a live same-key sibling with no dispatch loop, and getting an accept with no reply — not as a custody refusal. So nothing needs repairing there; the wording that reached the permanent record is the one your source read supports. It also carries: both mechanisms named at source, the four ruled-out hypotheses with their disproofs, the never-executed cell that red on its first firing anywhere, the three-lane r2 set with lane 3 as fix-set, the acceptance criteria including B's SERVED count, counter 104 unconsumed, and the unbounded reply read recorded as a pre-existing hazard explicitly not in r2.

ONE ADDITION I WILL MAKE TO THE r2 INTAKE COMMENT, from your ruling rather than mine: the breadcrumb gap that IS real — a daemon accepting a WAN stream it will never read logs nothing — goes on the register beside the unbounded read. Neither is #272's to fix; both are this arc's findings and the board should own them.

THREE LANES CONFIRMED, no fourth, no descope, #17's remote arm ships. Waiting on the assembled sha.