On shared hosts, give each background agent a hard RSS cap

When delegating work to background agents on a shared Linux host, a soft instruction like 'use a few GB' is not enough. An agent's probe that loaded 128 waveform signals at once grew to ~4.5 GB per process, and two of them triggered the kernel OOM killer, which killed the user's largest interactive app instead of the probes. Put hard limits in every brief: one heavy process at a time, a watchdog that kills at a per-process RSS guard, a total cap per agent, and a free -g check before each run. Name the user's processes that must never be put at risk. Also beware pgrep -f 'cmd +flag': + is a regex quantifier, so the pattern never matches and 'is it still running?' checks silently say no.