Knowing without Acting: The Disentangled Geometry of Safety Mechanisms in Large Language Models — arXiv2