Data-Ready Blind-Start CPI Evaluation through Fold-Safe Metabolic Transfer

We ask whether reaction-derived metabolic co-associations can add useful information to double-cold CPI prediction, and whether that transfer can be made data-ready enough to evaluate credibly. V3.7 freezes the ColdstartCPI reproduction, learns a side representation from fold-filtered KEGG co-associations, and trains rescue/suppress/abstain routing from dual-entity-blind out-of-fold errors. Held-out structures and sequences are excluded before metabolic pretraining and router supervision stays inside outer training. Across five predefined folds, mean AUC/AUPR improve from 0.836145/0.782751 to 0.842217/0.792995, with gains on every fold. Metabolic transfer is the changed biological information path; machine-checkable boundaries make the comparison credible. We do not claim metabolic semantics alone causally explain the gain.