How Much Structure Does Band-Gap Prediction Need? A Fold-Controlled Representation–Learner Audit from Composition to Graphs on MatBench mp_gap

We audit three pipelines on MatBench mp_gap (106,113 PBE band-gap targets): composition-only Magpie/XGBoost, engineered structural descriptors with XGBoost, and ALIGNN. Targets, official test folds, metrics, and non-negativity post-processing are controlled; the Level-2-to-Level-3 step changes learner, direct model-fitting fraction, and optimization budget. MAE falls from 0.3380 to 0.2855 and 0.1797 eV. Gross coordinatewise SHAP shares rank coordination fingerprints (26.2%) above global symmetry (11.8%). Under the fixed single-seed 8,000-round cap, deleting coordination does not worsen MAE, whereas deleting symmetry raises it by 0.0287 eV. Attribution share and deletion response are distinct. On 20,299 entries in 6,681 same-fold repeated-composition groups, ALIGNN MAE is 0.0088 eV above a fold-conditioned oracle; the aggregate 95% formula-cluster interval spans zero. Only the post hoc >1 eV global within-formula-spread stratum has an unadjusted pointwise interval entirely below zero. At the zero-gap boundary, the comparable Level-1-to-Level-3 raw-prediction interquartile range contracts about 36-fold, and 93.3% of entries move closer to zero. Clipping creates a growing zero-error atom. On positive-gap entries, median absolute error falls 2.47-fold, but the advantage reverses between P95 and P99. Over the full dataset, the worst 1% of ALIGNN errors contribute 56.0% of total squared error. Suppressing explicit angle values in a configuration-matched fold-0 ablation raises clipped MAE by 8.2%. The audit separates representation–learner performance, attribution density, deletion response, boundary behavior, and tail risk.