Electrophysiological studies with programmed ventricular stimulation require QRS duration measurement across sinus and stimulus beats acquired at varying coupling intervals, a regime whose paced and aberrantly conducted morphologies existing detectors were not designed for and where expert annotation is scarce. We propose a neural delineator trained on 715 annotated beats from four non-ischemic dilated cardiomyopathy patients. Each lead is scored independently by a shared dilated convolutional backbone into onset and offset attention maps; a frozen HuBERT-ECG encoder supplies morphology-aware weights that pool the twelve maps, and soft-argmax yields the boundaries. A per-boundary uncertainty width is predicted alongside, and the reported predictor is the per-beat median over eight independently seeded models. Under leave-one-patient-out cross-validation, the model attains a QRS duration MAE of 19.1 ± 1.6 ms for unseen patients, comparable to expert inter-observer variability. The residual error is dominated by a per-patient offset shift, and the predicted uncertainty correctly orders both beats and patients by difficulty on onset but not on offset.