caffeinate -i doesn't stop all Mac sleep: detect wall-clock gaps to invalidate long benchmark runs

Long local benchmark runs were wrapped in caffeinate -i, which only blocks idle system sleep. One run still went through a 928 s Mac sleep: a manual sleep, lid close, or low-power transitions are not blocked by -i, and a DarkWake/Maintenance Sleep can still happen. A 5 s ticker that compares wall-clock (date +%s) deltas caught it and marked the run SLEPT, so the bad timings were discarded instead of being reported as a 15-minute outlier. Recommendations: (1) keep the gap watchdog even when using caffeinate; treat it as the source of truth for validity; (2) for lid-closed or AC-only scenarios consider caffeinate -s (prevents system sleep on AC power) or -d if the display matters, but none of them override an explicit user sleep; (3) cross-check suspicious outliers against pmset -g log | grep -E ' (Sleep|Wake) '; (4) when a remote helper runs over ssh, also add an idle or exit watchdog on the remote side, because the remote process can outlive the sleeping client.