Turbulent flows occur in a wide range of aerodynamic and industrial applications where reliable flow information is often required. Such information is desirable at high spatial resolution and in real time, yet remains challenging to achieve. In practice, experimental techniques provide valuable real-time information but are often limited by spatial resolution and constraints on sensor placement. Conversely, high-fidelity simulations such as direct numerical simulation provide fully resolved flow fields but at a computational cost incompatible with real-time use. These limitations have motivated data-driven approaches that enable high-resolution, real-time prediction by combining the strengths of experimental and computational methods. In this regard, for many real-world applications, sensor data is limited to wall measurements, motivating the development of wall-to-flow reconstruction models that infer volumetric flow fields from surface data. Despite recent advances, existing wall-to-flow reconstruction models tend to exhibit a progressive loss of accuracy with increasing distance from the wall. Inspired by temporal information propagation in image-to-video methods, the proposed approach reformulates volumetric flow reconstruction by treating the wall-normal direction as a pseudo-temporal axis to enhance accuracy across the entire wall-normal span. This work proposes a physics-guided deep generative model for reconstructing wall-bounded turbulence from wall-pressure and wall-shear data. A compact U-Net predicts a sequence of wall-parallel velocity slices, with wall measurements injected through a control branch. Within the backbone, a lightweight motion transformer operates along the wall-normal direction and promotes consistency of coherent structures across adjacent slices. The method is assessed for incompressible turbulent channel flow at a friction Reynolds number Re_tau = 180. The model recovers instantaneous three-dimensional velocity fields with high near-wall fidelity while maintaining accuracy away from the wall. It accurately reproduces first- and second-order statistics as well as pre-multiplied energy spectra, showing excellent agreement with reference data. Compared with a representative state-of-the-art 3D generative adversarial network, the proposed model further achieves lower normalized mean absolute error profiles for all velocity components and physical consistency of the reconstructed fields.