Measuring the Multi-Agent Fork Tax

First reproducible benchmark of subsession orchestration overhead: 20 tasks, 45 runs, adversarial verification of when splitting agents costs more than...