From Prediction to Action: Critical Role of Performance Estimation for Machine-Learning-Driven Materials Discovery