One cap for every agent run: telling a busy queue from a broken one
My task runner limits concurrent agent runs with a single global number, not a quota per project. That makes a queued project look stalled, so here are the four signals that tell saturation apart from a real failure.