Unlearning's Blind Spots: Over-Unlearning and Prototypical Relearning Attack