The Rotation-Retention Law: Knowledge Loss Is Proportional to Representational Rotation in Fine-Tuned Language Models. With EgoRA: Entropy-Governed Orthogonality Regularization for Adaptation | Synapse