EMUnity: Joint Foundation Modeling of Raw I/Q Across Eight Electromagnetic Tasks

Most temporal foundation models are evaluated as forecasters, while physical signal systems must also support heterogeneous classification, regression, multi-label, and sequence decisions. We present EMUnity, a raw-I/Q foundation model that combines masked self-supervised pretraining with one joint fine-tuning stage, a shared temporal backbone, task-aware sparse experts, and lightweight heterogeneous heads. Under the full-data protocol, one jointly trained model achieves comparable or superior performance to seven capacity-matched specialists across eight tasks with an 85.7% parameter reduction, while pretraining and task-aware sparsity yield targeted gains on complex physical structures. These results extend temporal foundation modeling beyond forecasting and establish concrete design principles for shared representation, conditional computation, and multi-task model consolidation in heterogeneous physical time-series systems.