A training-free, complex-valued differentiable focusing framework for digital holography is established utilizing band-limited angular spectrum propagation and Wirtinger gradient flow. By optimizing propagation distance as a continuous variable and introducing spatial-domain metrics (e.g., total variation), this method explicitly overcomes the zero-gradient limitation of frequency-domain quadratic criteria to ensure stable analytical gradients. A coarse-to-fine strategy then generates z-encoded multifocal complex fields for coherent extended-depth-of-field reconstruction. Quantitative validation shows that DHF-Frame-RLG localizes the optimal reconstruction plane with an axial MAE of 1.0 µm and RMSE of 1.8 µm, and achieves a 96% autofocus success rate, corresponding to a 4.5× reduction in evaluations relative to exhaustive search. Crucially, this framework mitigates boundary blurring and wavefront coherence loss caused by traditional discrete axial scans, providing a robust physical foundation for future integration with physics-informed learning models.