The growing deployment of urban surveillance systems requires adaptive anomaly detection methods capable of modeling complex, context-dependent pedestrian behavior, yet existing approaches remain constrained by static training assumptions, limited contextual reasoning, and instability under continual learning. To address these limitations, we introduce SCOUT-IPAD (Scene-conditioned Counterfactual Online Uncertainty-gated hypergraph Transformer for Incremental Pedestrian Anomaly Detection), a unified framework that combines scene-conditioned representation learning, higher-order interaction modeling via dynamic hypergraphs, counter-factual trajectory prediction, and uncertainty-guided adaptation. The method represents pedestrians through multimodal features and leverages a hypergraph transformer to capture complex multi-agent dynamics, while anomalies are detected by measuring deviations from multiple plausible future behaviors. To support reliable incremental learning, the framework incorporates uncertainty-gated updates and a dual-memory consolidation mechanism, enabling stable adaptation while reducing catastrophic forgetting. Experimental evaluation on four benchmark datasets shows that SCOUT-IPAD achieves state-of-the-art performance among the compared methods under the evaluated settings, with Frame-AUC scores of 98.3% on Ped2, 89.4% on Avenue, 84.1% on ShanghaiTech, and 76.2% on Street Scene. In localization evaluation, the model reaches 94.6% Pixel-AUC on Ped2, 79.1% Pixel-AUC on Avenue, and an EER of 12.4%. Under continual learning evaluation, SCOUT-IPAD attains 85.3% incremental AUC, with forgetting reduced to 3.2 and backward transfer improved to −1.8. These results indicate that SCOUT-IPAD effectively addresses the stability-plasticity trade-off and strengthens context-aware pedestrian anomaly detection in dynamic urban scenes.