Semilocal PBE calculations can remove viable photocatalysts before screening by labeling narrow-gap semiconductors as metals. We address this failure mode in the Computational 2D Materials Database (C2DB) by combining leakage-aware repair of PBE-spurious metallicity with HSE06–PBE Δ-learning. Stage I classifies HSE06-unknown PBE metals using structural, chemical, magnetic, and stability descriptors, while excluding HSE06/GW quantities and PBE electronic shortcuts. Stage II learns E_g^HSE - E_g^PBE for the corrected insulating population. The curated XGBoost regressor reconstructs HSE06 gaps with a mean absolute error of 0.108 eV (R²=0.989), compared with 1.036 eV for raw PBE. The Stage I classifier is used only for triage because the labeled true-metal class contains 29 materials; its best holdout performance gives 87.5% accuracy, 0.286 true-metal recall, and 0.643 balanced accuracy. The corrected pH 0 electronic screen yields 10 strict and 29 initial relaxed green-hydrogen photocatalyst candidates. Four strict and 18 relaxed candidates would fail the same 1.6–2.8 eV gap window at the PBE level. Targeted VASP HSE06 calculations for six ML-predicted compounds give material-level MAEs of 0.417, 0.182, and 0.110 eV for the C2DB-native, Magpie+structural, and curated models, respectively; 1AgBr-1 shifts above the upper gap cutoff, leaving 28 retained relaxed candidates. The workflow shows that high-fidelity correction must be evaluated by candidate membership, not only by global regression error.