Reconstructing unsteady, vortex-dominated flows using Space-Time Super-Resolution (STSR) remains a significant challenge. The primary difficulty lies not only in increasing resolution, but also in preserving the sharpness of vortex cores and maintaining temporal consistency to prevent phase errors. This study introduces a framework that addresses these issues by combining state-space sequence models with physics -informed regularization, thereby ensuring that the reconstructed flow fields are both visually sharp and physically consistent. To evaluate the proposed approach, a moving-cylinder benchmark was developed, departing from the use of static datasets. This benchmark introduces the challenge of time-varying immersed boundaries, requiring models to account for dynamic geometry while reconstructing velocity and pressure fields and synchronizing drag and lift signals. This demanding setup is intended to assess whether these models can achieve the long-horizon consistency necessary for real-world engineering applications. Four different approaches were compared: UNet (spatial), ConvLSTM (recurrent), Transformers(attention), and Mamba (state-space). By embedding physics-based constraints directly into the loss function, the framework effectively suppresses non-physical artifacts and high-frequency noise that often affect purely data-driven methods. The results indicate that, although all models benefit from additional temporal context, the Mamba architecture is particularly effective. It matches the accuracy of traditional recurrent models while providing substantially greater stability across long sequences. For practical applications such as load prediction or the development of digital twins, this approach yields cleaner and more reliable force histories and flow structures, which can be used directly in downstream analysis without the need for additional numerical filtering.