Break the Optimization Barrier of LLM-Enhanced Recommenders: A Theoretical Analysis and Practical Framework