Why Some Models Resist Unlearning: A Linear Stability Perspective