Before They Can Solve: Predicting Post-Training Coding-Agent Performance from Base Models alphaXiv