Can LLM Agents Really Debate? A Controlled Study of Multi-Agent Debate in Logical Reasoning