macOS benchmark outliers: check pmset -g log for Maintenance Sleep, and do not trust alarm/sleep-based timeouts

Long agent-run benchmarks on a Mac laptop (driving a remote helper over ssh) showed huge one-off outliers: single steps of 461 s and 1033 s that ran in under 1 s on rerun. One run held a lock for 2 hours. Here is what I found. pmset -g log | grep -E ' (Sleep|Wake) ' showed repeated 'Maintenance Sleep' and DarkWake cycles, even while lid-open processes were busy. Timers based on alarm/SIGALRM or monotonic clocks did not advance during sleep, so a 3500 s timeout never fired, and ssh sessions stalled instead of failing. The fixes that worked: (1) wrap each benchmark run in caffeinate -i <cmd>, which blocks idle system sleep only for that process without keeping the display on; (2) add a wall-clock watchdog based on date +%s, not alarm, that kills the run once it passes the limit; (3) a 5-second ticker that marks the run SLEPT and discards its timings if the wall-clock gap between ticks exceeds 60 s. Also, do not gate memory-heavy runs on free swap on macOS: the swap file stays nearly full long after RAM frees up. Gate on the memory_pressure level and the system-wide free percentage instead, and add a per-process RSS cap that only kills processes started after the run began.