Reconstructing complete, high-resolution daily meteorological fields over multidecadal periods from station observations is severely underdetermined: stations are sparse, unevenly distributed, and temporally inconsistent, especially over complex terrain. Because this ill-posedness requires prior assumptions, the choice of prior determines how well extremes are preserved. Conventional products, whether interpolation, multi-source fusion, or reanalysis, rely on prescribed background-error covariances that attenuate localized extremes as coverage declines. We introduce Generative REconstruction (GRE), a sequential assimilation framework that replaces prescribed covariances with a generative prior learned from satellite-era gridded fields. The prior, a diffusion model trained jointly on precipitation, pressure, wind, and temperature, encodes multivariate spatial structure including heavy-tailed extremes. At each step, posterior sampling conditions the prior on available stations and a forecast background, producing an ensemble whose spread reflects observational coverage. Applied over Southwest China (1951–2024), GRE yields daily 0.1° ensemble reconstructions of all four variables. Conditioned on 80 % of stations, GRE achieves<em> R</em> = 0.93 and RMSE = 5.8 mm/day for daily precipitation versus 0.69 and 7.1 for the baseline product, recovering rather than smoothing localized peaks. The joint prior propagates observational information across variables, constraining fields not directly observed. Over three decades, reconstructed precipitation extremes remain consistent with the reference climatology while ensemble spread contracts as gauges densify, evidencing calibrated long-term uncertainty. The forecast background adds most value under sparse coverage and persistent synoptic conditions, maintaining bounded physically consistent uncertainty throughout the early record.