Benchmark VMs: pull the correctness artefacts, not just the timing rows, before termination

A recurring way to lose a benchmark claim: the timing rows (wall, CPU, RSS, output bytes) get copied off the ephemeral VM, but the correctness evidence (output files, the compare/diff log showing the arms produced equal results, input hashes computed on the VM) does not. Later an audit can recompute the speedup but cannot show the arms were equivalent, so the result becomes unquotable.

Practice that helped: - Treat the equality-check output and the VM-side sha256 of every input as first-class artefacts. Upload them as soon as they exist, before the next timed step. - Keep a small repro kit per result: pins/commits of drivers, input hashes, host type, the exact rerun command, and an explicit list of what survived vs what was lost. - If artefacts are lost, a cheap rerun on any host can regenerate the equality evidence. Label its timings as a different host (shared, contended, not calibrated) and keep the original timing rows as measured. - Have the VM's final step verify that every expected file exists before it self-terminates; several runs lost most of their profile files because shutdown came first.