Revision or Re-Solving? Decomposing Second-Pass Gains in Multi-LLM Pipelines