Cardiovascular flow simulations require high-fidelity computational fluid dynamics (CFD) to resolve hemodynamic metrics such as wall shear stress, a key indicator of vascular health associated with atherosclerosis, thrombosis, and aneurysm rupture risk. Despite their predictive value, such simulations impose enormous computational burdens that hinder clinical translation. Super-resolution neural networks, such as TEECNet, can recover high-resolution flow fields from low-resolution CFD solutions, and domain decomposition methods that enforce locality enable scaling to larger domains. However, existing graph-based formulations face limitations in approximation capacity when constrained by explicit mesh connectivity, reducing flexibility in resource allocation and adaptability to varying mesh topologies. Moreover, the data volume of complete vascular simulations makes it inefficient to capture all spatial regions with equal fidelity, particularly near-wall boundary layers critical for wall shear stress prediction. Finally, very few studies have explored generalization from synthetic vascular trees to patient-specific geometries, limiting the clinical applicability of learning-based methods. In this work, we extend TEECNet to a mesh-independent formulation that operates on local geometric features rather than explicit mesh connectivity, enhancing approximation capability and enabling flexible allocation of computational budgets. We integrate a gradient-based adaptive sampling strategy that prioritizes regions with significant discretization error, particularly near-wall boundary layers. The domain decomposition framework is extended to three-dimensional vascular geometries, improving generalizability toward transitional flow dynamics. We validate the approach on 24 synthetically generated left coronary arteries spanning healthy and diseased cases with stenosis severity ranging from 10–60%. Results demonstrate computational speedup exceeding 30 times compared to high-resolution CFD. Despite training solely on synthetic geometries, the method generalizes to patient-specific coronary arterial trees in a zero-shot manner through geometric registration, enabling accurate recovery of time-averaged wall shear stress distributions without retraining.