Enigmata’s Multi-Stage and Mix-Training Reinforcement Learning Recipe Drives Breakthrough Performance in LLM Puzzle Reasoning