Canary validating the new training-time simulation_eval hook: per-sample model contributions (pred, z1_f{k}, z2_f{i}f{j}, model_z_abs*) extracted via forward_interpret on the 20260706_simulation valid split, at the final epoch (9) of a 10-epoch canary run. Baseline (all-zero) antiabsorption penalties -- validates plumbing on real data/GPU, not a scientific result.
