Human-on-the-Loop: A Theory of Oversight Capacity, Its Failure Modes, and the Non-Delegable Residual in the Age of AI Agents