Abstract Weather foundation models pretrained on reanalysis data may encode subgrid‐scale physics beyond their forecasting objectives. Using Aurora, we probe this with a multi‐method fine‐tuning framework: agreement across structurally different adaptation strategies signals shared backbone information, while a matched randomly initialized control isolates pretraining’s contribution. We evaluate six methods (0.5%–100% of trainable parameters) for predicting ERA5‐resolved, T21‐filtered gravity‐wave momentum fluxes (313–1,820 km). Evidence for a pretraining benefit is strongest for the zonal flux u′w′ ${u}^{prime }{w}^{prime }$: several parameter‐efficient methods outperform random initialization, most clearly FISH + TADP (38.9% vs. 41.6% normalized error, p $p$ = 0.007). The meridional flux v′w′ ${v}^{prime }{w}^{prime }$ remains hard for all methods (NRMSE 60%–80%). The benefit is level‐dependent: 11–14 percentage points at 200–300 hPa but only 0.2 at 50 hPa. Approximately 1% of parameters recover part of full‐fine‐tuning skill (38.9% vs. 31.8%), indicating regularization by the frozen backbone.