Inverting point-source emissions from sparse atmospheric observations is difficult because emission rate, release height, wind speed, turbulence, and plume morphology can compensate for one another. We present NeuPlume, a neural-physical probabilistic inversion framework that returns an ensemble of data-compatible plume fields rather than a single point estimate. A Lagrangian stochastic model builds a scenario-specific forward library, a conditional neural field compresses the concentration fields, and a latent diffusion model learns the feasible field prior. During inversion, diffusion posterior sampling combines this prior with sparse observations to infer emission rate, effective release height, wind speed, turbulence intensity, and the full concentration field. The present implementation targets passive, low-height, near-neutral releases over flat terrain. Across six synthetic cases within this scope, NeuPlume achieves mean errors of 10.0% for emission rate and 1.4% for effective release height, outperforming Gaussian plume and mass-balance baselines under the same volumetric observation protocol. On 30 holdout cases, the nominal 68% credible interval attains 73.3% empirical coverage. As an illustrative field-transfer check, NeuPlume is applied to four UAV methane transects above a coal-mine ventilation shaft; known-U posterior intervals from three flights overlap the same-shaft hourly inventory, with diagnostics identifying wind-speed and height-boundary sensitivities under model mismatch. NeuPlume provides uncertainty-aware source-parameter constraints within the physical scope of its forward library and can be re-instantiated for other regimes by rebuilding the scenario-specific simulation ensemble and retraining the neural prior.