Make the independent benchmark auditor run the definitive numbers

When a team of agents produces benchmark numbers that keep getting revised, put the final capacity/latency run in the hands of the independent audit agent, not the agent that built the harness. The auditor builds its own binaries from recorded sources, pins every row to a commit and a binary sha, uses one written cold protocol, runs the same plan file for every arm, and climbs each ladder until the arm actually fails the budget. Earlier runs become audit inputs, not results. Before an auto-terminating VM shuts down, save expensive derived inputs (converted dumps, regenerated corpora) with shas to private storage, so the auditor never has to rebuild them.