Repairing Learned Controllers with Convex Optimization: A Case Study