H8 activates at DAA 79,251,000 on mainnet.
Every node must be on v1.5.0 before the gate.
Fixed — the inference reward vault was rejected at relay
Past the H8 gate an AiRequest no longer names a miner: it locks its reward in a keyless vault output, and the chain mints that reward to the first miner whose response is accepted.
Consensus accepts that vault output. The mempool did not. Standardness rejected it twice over — the vault script matches none of the recognised script classes, and the dust rule counts every unspendable output as dust. A v1.4.9 node therefore refuses every inference request it is offered:
transaction output #1: non-standard script form
The request never propagates, never reaches a miner, and never gets answered. Blocks containing one stay valid, so this is not a chain split — but in practice inference is dead past the gate on an un-upgraded network.
The vault output is now accepted as standard. The exception is an exact script match: a look-alike OP_RETURN remains non-standard.
Added — test coverage for the H8 escrow branch
Six tests over the routed branch, which had none. They pin the cutover in both directions — the vault accepted past the gate and refused before it, a designated CSV escrow refused past it — plus the script-version match, an underfunded vault, and a missing escrow output. One further test covers the mempool exception and its look-alike rejection.
Changed — peer version floor raised to 1.5.0 at the H8 gate
Past the gate, peers below v1.5.0 are refused. They derive request identity, reward locking and audit arming differently, and they cannot relay an inference request at all. The floor is keyed on the local virtual DAA, so a node still catching up to the gate keeps peering with the builds it needs to get there.