Physics-aware generative fusion via Complex-Flow Mamba enables robust 4D radar-vision perception

4D imaging millimeter-wave radar is quickly redrawing the map for autonomous perception. However, a significant limitation persists: conventional deep learning models often fall into a 'point cloud fallacy.' They treat radar echoes as simple discrete points, which washes out the complex-domain phase coherence. This renders systems highly vulnerable to interrupted sampling repeater jamming (ISRJ) and degrades performance on sparse targets at long ranges. We propose Complex-Flow Mamba, a framework that dives below the feature level to work directly at the signal-physics layer. We started by building a continuous Implicit 4D Complex Field. By using polar decoupling and Von Mises distribution modeling, we brought back the phase gradients that traditional discretization just throws away. To mitigate complex ISRJ interference, we developed a Complex Latent Diffusion Module (CLDM). Unlike standard filters, this generative prior performs microscopic phase correction to decouple coherent artifacts at their source. On the architecture side, we used Mamba-2's Structured State Space Duality (SSD) theory to create a bi-directional state-modulation backbone with O(N) linear complexity. Here, visual semantics act as a dynamic filter for radar, while radar dynamics serve as a physical gatekeeper for the vision stream, creating a deep entanglement between heterogeneous flows. We also added a generative detection head based on Consistency Models and a physics-consistent self-supervised loss to fix the "hollowing" effect in far-field detection. Tests on the View-of-Delft (VoD) and K-Radar datasets show that Complex-Flow Mamba hits SOTA levels in mean Average Precision (mAP) and NuScenes Detection Score (NDS). It also maintains a 0.88 jamming suppression gain under extreme electronic pressure.