For the 2026 George B. Moody PhysioNet Challenge, we developed a cross-site framework for predicting cognitive impairment 1--6 years after polysomnography (PSG) by combining frozen pretrained multimodal PSG representations with clinically derived physiological features. Model development used leave-one-site-out validation across three centres, followed by official evaluation on an unseen centre. The hybrid ranking pathway achieved an Age-conditioned AUROC of 0.708, Age-weighted AUROC of 0.755, AUROC of 0.799, and AUPRC of 0.273. However, its initial binary operating point generalized poorly, yielding a Reward of -0.293. Because threshold adjustment alone did not resolve this discrepancy, we decoupled continuous risk ranking from binary classification while preserving the ranking pathway. A clinically restricted ExtraTrees decision head using sleep, oxygenation, and age-relative features increased Reward to 0.156 while leaving all ranking metrics unchanged. These findings indicate that cross-site discrimination and threshold-dependent decision performance can generalize differently, and that a rich hybrid representation for risk ranking can be effectively complemented by a simpler clinical decision pathway.