3D Gaussian Splatting (3DGS) is celebrated for its high-quality and real-time rendering capabilities, yet its performance degrades severely when faced with real-world mixed blurs caused by both lens defocus and camera motion. Existing methods need to handle different blur types separately and rely on deterministic optimization that is sensitive to initialization, struggling to provide a robust and unified solution. We propose SPMCGS, a deblurring framework for 3DGS. We formulate the reconstruction as a posterior probability sampling problem to enhance robustness against poor initializations. Our core contribution is a unified, state- and pose-aware blur decomposition modulator: this single network uniformly models both defocus and camera motion blur. Furthermore, it handles “dormant Gaussians” from our MCMC process by using opacity as an input. Simultaneously, it is guided by a sharpness prior generated by a pre-trained network to focus the optimization on difficult regions. Experiments on several benchmarks demonstrate that our method achieves highly competitive performance and exhibits superior visual quality. Crucially, our framework introduces no inference overhead, fully preserving the real-time rendering capabilities of 3DGS.