Sleep disturbances have been associated with cognitive decline, but generalizable prediction from routine polysomnography remains difficult. As team gdub, we developed a CPU-oriented system that represented each study with 830 handcrafted demographic, physiological, and annotation-derived features. Training-cohort preprocessing removed configured BMI and coverage/missingness fields, excluded the 30\% of candidate features with the largest cross-site Kolmogorov--Smirnov statistics, and applied median imputation. ExtraTrees, CatBoost, and LightGBM probabilities were averaged and calibrated with out-of-fold isotonic regression; the submitted system also included an age-binned LambdaRank score and an age-local prevalence rule for binary predictions. In strict public three-site leave-one-site-out (LOSO) evaluation of the classifier path with LambdaRank disabled, the submitted KS setting achieved a macro age-conditioned AUROC of 0.654. Official hidden validation of submission 2163 produced an age-conditioned AUROC of 0.486 despite an ordinary AUROC of 0.746. This discrepancy indicates that all-pair discrimination did not persist among age-comparable participants and identifies age-conditioned cross-source generalization as the principal limitation of the submitted approach.