Toward Reliable Machine Unlearning: Theory, Algorithms, and Evaluation