Democratizing Large-Scale Re-Optimization with LLM-Guided Model Patches