[edgehero]

Fixes #725.

The copy-mode steering case ran at maxCostMicros: 0 against a real model, so the runner hit its cost cap within milliseconds of the first ack. The worker closes stdin on the verified exit line, so the second message's frame was never written. It now runs on a zero-rated, keyless fake model served by the stand-in:

• The first request waits for an event, not a timer. The stand-in holds it until the conformance script signals, by touching STEER_SIGNAL_FILE, that the relay has written the second frame (at most 30 s). It then asks for a read of the job's prompt file.
• Any request carrying a tool result gets "done". The reply depends on the request body, not on a request counter, so a retry of the first request is held again.
• The pass condition stays strict. Exit 0 with a verified exit line naming both messages, both delivered, at least 2 frames written and 2 acks verified, and no cost spent.

The stand-in's other routes answer exactly as before, and the request body is never logged.

Rejected: rerunning a relay tick requested mid-pass. That was not the cause, and in production it would only add Valkey load when Valkey is slow.

Residual: this case is the first to stream a model answer through OpenShell 0.1.4's protocol: rest layer.

Verification:

• node --check passes, and the workflow YAML parses.
• The extracted stand-in, driven locally, holds until the signal file appears, holds a retry again, and answers "done" once a tool result arrives.
• The real runner against it: exit 0, 2 turns, cost 0.
• The openshell lane will be rerun three times on this PR before merge.