SAVGO: Learning State-Action Value Geometry with Cosine Similarity for Continuous Control — ThinkLLM