Strict taxonomic cold-start benchmarking reveals when graph learning adds value beyond microbial lineage priors

Prioritising disease associations for newly observed microbial taxa is a central challenge for microbiome-driven discovery: curated links are sparse, experimental validation is costly, and deployment often begins with lineage and incomplete functional descriptors rather than known disease edges. Yet standard random or transductive benchmarks can reward taxonomic or association-derived leakage, leaving unclear whether graph learning contributes information that would remain available for an unseen taxon. We address this gap by formalising strict microbe-side taxonomic label cold start as a deployment-aligned benchmark for microbe-disease association prediction. The leakage-audited benchmark comprises 1375 MDIPID microbe-neurological disease associations across 626 taxa and 32 diseases, with Genus and Family-fixed splits, sampled-negative discrimination, and full-candidate retrieval. TaxFuse-HGCL serves as an inductive, prior-calibrated reference framework: a hypergraph neural ranker is trained inside each fold, guided by taxonomy-kNN listwise targets, and calibrated with the taxonomy-neighbour score available for the held-out taxon. Across experiments, taxonomy kNN emerged as a high-value biological prior rather than a weak baseline. In the Genus split, TaxFuse-HGCL improved the neural backbone from sampled AUC/AUPR 0.8307/0.6213 to 0.8481/0.6490, reaching the taxonomy kNN range, while taxonomy kNN retained stronger full-candidate recall. In the Family-fixed split, TaxFuse-HGCL reached sampled AUC/AUPR 0.8516/0.6591 and micro-AUPR 0.1368, the strongest discrimination and micro-averaged retrieval profile among protocol-matched baselines. Residual, oracle, and 16S probes showed that the dense neurological benchmark is dominated by lineage-carried signal, whereas on a larger, sparser Disbiome dataset fold-internal fusion produced paired gains over tuned taxonomy kNN on sampled AUC/AUPR and full-candidate AUPR endpoints. Together, these findings make strict taxonomic cold start a biological model-selection criterion: they show when graph learning adds deployable signal beyond lineage, when a simpler lineage prior is sufficient, and why prior-calibrated neural systems should be evaluated before experimental follow-up.