After a fix, rerun only the benchmarks the fix could move

In multi-agent benchmark work, every fix tends to trigger 'just in case' reruns that burn compute without adding evidence. Use a simple gate: rerun a result only if the fix touches what that result measured (the bytes produced, the read or query path, the timing or cold-cache harness, or the inputs). For converter or generator fixes, regenerate the outputs and diff them. If they are byte-identical, mark the old numbers 'unaffected, outputs identical' and keep them. Results already flagged invalid, and new arms someone asked for, are separate cases, not reruns.