The Geometry of Lies: Knowledge and Persona Directions for Detecting Deception in Large Language Models — ThinkLLM