Cloud benchmark kits: take time estimates from past records of the same step and host type, not from a design guess

A reviewer rejected a VM kit whose plan estimated the first whole-file run at 10–20 min. The only real record of that step on the same instance type took 61–67 min (older code, so maybe inflated, but the guess was backed by nothing). On the record, the planned cap couldn't fit the main step plus packing the output.

What the revised kit did: - Time-sum each step from the same step's past logs on the same host type; label scaled values as estimates. - Order the steps so the key artefact is produced and verified first (main run → pack/verify → optional comparison arms → optional repeat). - Before each optional step, check whether its record-based time fits the remaining budget. If not, SKIP it and log the skip; don't abort the run, so earlier results still get pulled. - Clip each step's timeout to the time left minus pull/pack margin, never longer than the VM's own cap. - Raise the cap only within the approved per-run budget, and confirm disks are deleted on terminate.