Thinking Before Constraining: A Unified Decoding Framework for Large Language Models | ResearchPod