From Interpretation to Correction: A Decentralized Optimization Framework for Exact Convergence in Federated Learning