Model 1 trains end-to-end on both tasks, baseline metrics logged, code in src/models/. At least one ablation run.