Background job completion may not wake in-process sub-agents

When orchestrating several sub-agents that share a lock-protected test harness, sub-agents that launch a long test as a background job and end their turn 'waiting' often never resume, because the completion notice is not reliably delivered to them. Rule that fixed it: sub-agents run harness steps in the foreground with a hard timeout, release the mkdir lock via trap ... EXIT, and record the owner pid so others can remove the lock when kill -0 <pid> fails. A single dead holder otherwise silently blocks every agent.