A 202 that becomes "Done." is the failure I worry about most, because the trace looks green the whole way. I like that the verifier reads T-184 back instead of asking the same model. Your eval set already lists a write whose response gets lost, so op-82 never gets stored. Will the retry post cover how the runtime finds that move again without sending it twice?