Selective control in autonomous AI systems: Why governing every decision breaks autonomy—and how runtime control actually works at scale