Flow-based intrusion detection in unmanned aerial vehicle (UAV) networks can underuse protocol relationships and model complementarity. The Protocol-Semantic and Confusion-Pair-Aware Heterogeneous Ensemble (PS-CAHE) combines semantic features, training-only confusion-pair selection and two fusion weights. On 113,952 deduplicated UAVIDS-2025 flows, PS-CAHE achieved 95.58% Macro-F1 and 93.88% Matthews correlation coefficient on the fixed holdout. The full pipeline improved Macro-F1 by 0.83 percentage points over Original XGBoost; the pair-aware increment over global out-of-fold (OOF) weighting was only 0.15 points. Across 25 endpoint-pair-disjoint folds, this increment averaged 0.14 points (corrected 95% confidence interval: 0.06 to 0.22), positive in every fold. Three-base and matched two-base probability stacking scored higher in both nested protocols, with row-stratified advantages of 0.08 and 0.05 points. Strict no-port node-disjoint evaluation yielded 93.54% versus 93.78% for Original XGBoost; the corrected difference interval spanned zero, establishing no unseen-node advantage. Training-only temperature scaling reduced negative log-likelihood from 0.1097 to 0.1074 and expected calibration error from 0.88% to 0.49% without changing decisions. The small fusion increment has no demonstrated operational benefit. PS-CAHE offers directly inspectable, low-dimensional fusion on one simulator-generated benchmark; real-flight data, independent external datasets, temporal/mission shifts, open-set attacks and adversarial robustness remain unevaluated.