Deep-learning post-processing can correct systematic errors in numerical wave forecasts, but its reliability under high-wave conditions remains unclear. We develop an error-correction framework in which deep networks predict the full future error trajectory of operational forecasts, and compare three architectures, a pure LSTM, a pure Transformer and a hybrid LSTM–Transformer, tuned identically with Bayesian optimization, using hourly observations from six buoys in the Bohai Sea and Yellow Sea over 2023–2025. On the full test sample all models improve the 72 h forecasts by 15.37% to 23.93% in RMSE, with the hybrid model attaining the lowest pooled RMSE; on the high-wave subset with SWH above 1.5 m, however, all three degrade the forecast, increasing RMSE by 7.58% to 9.05%. An exact decomposition of the squared-error change shows that fluctuation-capturing skill remains intact while the degradation is driven by the mean term: trained on samples dominated by low and moderate waves, the correctors learn a globally averaged adjustment whose sign opposes the reversed raw bias at high waves. Neither loss reweighting nor filtered training resolves this degradation at the 72 h horizon. Shortening the horizon to 24 h restores positive skill, with RMSE reductions of 4.43% to 7.88%, and reverses the ranking in favor of the pure LSTM. These findings indicate that architecture rankings are horizon-dependent, and they are consistent with the interpretation that extreme-condition skill is limited by the information content of the predictors rather than by model capacity.