The Geometry of Truth: Layer-wise Semantic Dynamics for Hallucination Detection in Large Language Models