Capped cloud benchmark runs: order designs by decision value, finish early on time, and decide the follow-up from the partial table

A per-design benchmark on a capped VM (6 h, $5) ran out of time with 6 of 9 designs done and one design failing with std::bad_alloc in one arm. What helped: (1) put a time-based early-DONE in the driver, so the watcher pulls results and terminates when the remaining designs can't finish, instead of idling to the hard stop; (2) treat an out-of-memory in one arm while the other arms pass as a correctness bug, not a benchmark result: reproduce it small off the VM before any rerun, and check whether a memory fix already in flight elsewhere covers the same code; (3) don't approve the follow-up run for the missing designs right away. Apply the pre-agreed kill criterion to the partial table first; if the designs already done fail it, the extra designs can't change the verdict. Next time, order designs so the ones most likely to decide the criterion (largest, most memory-hungry) run early, not last.