When Marginal Risk Hides Routing Failure: Path-Aware Selective Prediction in Multimodal Neurodegenerative Assessment

Adaptive multimodal predictors increasingly decide not only how to combine evidence, but also which evidence to acquire and when to stop. These decisions induce policy-dependent terminal populations, so acceptable pooled selective risk can conceal either a supported high-error path or a statistically fragile low-support branch. Unlike fixed subgroup auditing, each candidate policy induces its own terminal groups and statistical support, making path reliability part of policy selection itself. We introduce \method, a predictor-agnostic policy-selection framework in which terminal-path support and one-sided risk screening directly determine whether an adaptive acquisition policy is feasible during development. In a nested participant-disjoint evaluation of 902 participants from the Parkinson's Progression Markers Initiative (PPMI), marginal selection attained 0.675 automatic coverage and 0.087 selective risk, yet its supported clinical-plus-DTI terminal path made 14 errors among 68 automatic predictions (risk 0.206; one-sided 95\% upper bound 0.303). Under the same candidate family, nested path-aware selection yielded 0.600 automatic coverage and 0.072 selective risk, with no imaging acquisition and 0.400 deferral in this cohort. Across 500 controlled two-modality repetitions, paired ablations show that explicit support substantially reduces under-supported routing when added to path-empirical feasibility, while path-specific Clopper--Pearson screening further discourages small paths and provides a distinct uncertainty-screening effect. This retrospective, development-only study supports terminal-path evaluation as a complement to marginal reporting; it does not establish post-selection guarantees, prospective clinical safety, or that imaging lacks diagnostic value.