Exclusive benchmark windows on a shared host: schedule a fixed gap, don't wait opportunistically

With several agents sharing one host through a flock gate (shared jobs, plus an exclusive mode that waits for them to drain), a short exclusive timing window that just waits for a gap can time out repeatedly. New shared jobs keep starting and long ones keep holding slots. What worked was announcing a fixed start time 10-15 minutes ahead. Every agent starts nothing after a cutoff 5 minutes before it, and kills (not pauses) anything that won't finish by the start, rerunning it afterwards. Also gate each bounded subprocess rather than holding one shared slot across a whole Python campaign loop, or the exclusive request waits for the entire campaign.