Two history-only weather models lost most Swiss plant timing skill in tests within 2005-2024

We tested how much published ecological prediction skill depends on observed target-year weather. We used signed annual model statistics for Swiss plant dates and British butterfly timing and growth. We fit a recent bias correction and two history-only driver models on ten years, then scored later years. In 2015-2024, mean spring plant skill fell from 0.691 with observed weather to 0.014 with a one-year history model and 0.034 with a history trend. Later plant scores were 0.369, 0.025 and -0.742. An older-model test in 2005-2024 gave the same plant pattern. Butterfly timing and growth had small or negative history-only scores. An exact loss decomposition showed that weather mismatch explained the spring plant loss. In butterflies, its interaction with ecological model error offset part or all of that cost. With independent unbiased common temperature errors, median critical RMSE was 0.967 degrees C for 19 spring phases and 0.795 degrees C for eight later phases. These are conditional error budgets, not measured forecast errors. The study reuses published statistics and does not test a weather service. Short test periods, changing site coverage, weak driver benchmarks and fixed calibration limit inference. The package includes code, source hashes, performance statistics, a figure and independent arithmetic checks.