“In structured domains like coding, synthetic data can be generated without performance degradation because there are methods to programmatically inject diversity, preventing distributional collapse.”