Residual-Preserving Reconciliation and Seasonal Uncertainty in Foundation-Model Forecasts of Surface Heat Fluxes at a Tibetan Plateau Alpine Site

Forecasting alpine turbulent heat fluxes requires distinguishing predictive skill from assumptions about energy closure. We evaluated hourly sampled sensible and latent heat fluxes at Gyirong, Tibet, using 168 h histories to forecast 1, 6, and 24 h endpoints. A frozen Chronos-2 model predicted both fluxes, available energy and an accounting residual. A locally trained Cholesky network conditioned positive-definite reconciliation weights on historical state while permitting nonzero residuals. Local fitting used 2023, model selection and marginal interval calibration used disjoint 2024 subsets, and fixed settings were assessed in 2025 and partial 2026 against foundation, statistical and neural comparators. In 2025, reconciliation reduced sensible-heat mean absolute error by 2.68–4.16% relative to unreconciled Chronos-2, although XGBoost or GRU performed better in several comparisons. State conditioning improved latent-heat errors relative to constant reconciliation, but winter latent-heat accuracy deteriorated relative to unchanged Chronos-2. Forcing zero residual substantially worsened sensible-heat forecasts. Calibrated 80% intervals attained 78.5–83.9% aggregate coverage in 2025, while 24 h winter latent-heat coverage was only 62.3%. Beyond point accuracy, the contribution is a reproducible protocol that separates foundation-model skill from local adaptation under strict temporal extrapolation and exposes season- and target-dependent uncertainty concealed by aggregate scores.