Accurate geological parameterization is essential for reservoir history matching, yet existing methods face a persistent trade-off between computational tractability and reconstruction fidelity. Linear methods such as principal component analysis (PCA) are efficient and statistically well-behaved but produce over-smoothed fields that suppress fine-scale heterogeneity, while deep generative models offer richer representations at the cost of training instability and poorly conditioned latent spaces. This paper presents a hybrid framework, PCA-DDPM, which combines PCA-based dimensionality reduction with a denoising diffusion probabilistic model (DDPM) trained on residual fields. The DDPM is conditioned on the PCA reconstruction and learns to predict the difference between the ground-truth property field and its linear approximation, thereby stabilizing the training process. A patch-based inference strategy with geological mask-weighted merging enables scalable application to large three-dimensional reservoir volumes. Through testing on the COSTA carbonate reservoir model, PCA-DDPM achieves SSIM > 0.93 and R² > 0.87 for permeability reconstruction, and SSIM > 0.95 and R² > 0.92 for porosity, which represents an improvement of approximately 60% in SSIM over the PCA baseline. Full-physics two-phase flow simulations across 120 wells confirm that the reconstructed fields preserve well-level production responses, with median R² exceeding 0.97 for oil and water production rates and 0.99 for injection bottom-hole pressure. The framework operates within a compact 256-dimensional latent space, making it directly applicable to practical history matching workflows.