Cardiac conduction disorders such as left bundle branch block (LBBB) impair ventricular synchrony and increase the risk of heart failure. Cardiac resynchronization therapy (CRT) improves cardiac function in approximately 70% of patients, yet optimizing patient selection and pacing strategy remains challenging. A key limitation of current computational frameworks is the lack of explicit uncertainty quantification, which is essential for reliable patient-specific decision support. We present an uncertainty-aware, patient-specific framework that couples a three-dimensional electrophysiology (EP) model - based on the anisotropic Eikonal equation with a parametrized Purkinje network - with a closed-loop cardiovascular model (CircAdapt). Probabilistic EP uncertainty is quantified via Approximate Bayesian Computation, yielding 30 posterior Purkinje network samples per patient, each consistent with the observed 12-lead ECG. Cardiovascular mechanics are calibrated to MRI-derived cavity volumes using Constrained Bayesian Optimization (CBO), ensuring physiologically consistent model states. The framework was applied to four CRT candidates with heart failure and LBBB, and evaluated across seven pacing strategies. Calibration was accurate (volume errors below 15% in three of four patients) and robust (synthetic recovery errors of 5.5-9.7%). Despite variability in inferred Purkinje network configurations, hemodynamic predictions - including pressure-volume loops, ejection fraction, and maximum rate of pressure change - showed negligible dispersion across posterior samples. This robustness held even for pacing strategies with inherently limited left Purkinje identifiability. Hemodynamic responses were heterogeneous across pacing strategies, with no single configuration consistently outperforming others across all metrics, underscoring the patient-specific nature of acute CRT response. These findings support the clinical applicability of cardiac digital twins for personalized CRT planning.