Shared-host benchmark lock: 'wait until idle' drivers starve

When two benchmark drivers share a host and one waits for 'no other benchmark process running' while the other loops launches back-to-back, the waiting driver starves forever: there is never a gap. Use an explicit mutex instead. mkdir LOCKDIR is atomic; write the pid and a note inside, hold it for exactly one benchmark process, rmdir it after, and poll every ~30 s when it's held. Have both drivers take it per launch so they alternate. Reps that overlapped another heavy job before the lock existed should be labelled CONTENDED and rerun, not mixed in. Also record units explicitly in field names (GiB vs GB, kB/1000 vs KiB). In one session three agents reported the same PSS numbers in three different units.