We present a multimodal model for predicting future cognitive impairment (CI) from a single overnight polysomnogram (PSG), designed to generalize across unseen sites with differing recording systems and patient populations. The model integrates frozen electroencephalography (EEG) embeddings from the CBraMod foundation model with automated sleep-stage (YASA) and chin- and leg-electromyography (EMG) features. To limit site-specific information without increasing capacity, we remove site-dominated embedding dimensions using mutual information and train bagged gradient-boosted (LightGBM) and linear heads with age-matched pairwise ranking and channel dropout. This work is part of the George B. Moody PhysioNet Challenge 2026, "Screening for Cognitive Impairment During Sleep Studies." On the hidden validation set, our model (team OCA-CENTINEL) achieved an age-conditioned AUROC of 0.658 and AUPRC of 0.257, ranked 155th of 514 submissions; leave-one-site-out cross-validation yielded a mean AUROC of 0.69 and worst-site AUROC of 0.67. Increasing model or representation capacity improved within-cohort validation but not held-out-site performance, whereas capacity-neutral refinements transferred consistently. These findings suggest that cross-site generalization, rather than representation capacity, is the main bottleneck for robust PSG-based prediction of future CI.