v1.1.0

[1.1.0] — 2026-05-08 (package) | Model: v1.1.0

Added

• Restored full backward compatibility for v1.0.0 pretrained weights. ISEFlow_AIS(version="v1.0.0") and ISEFlow_GrIS(version="v1.0.0") now load and predict end-to-end; NormalizingFlow.load() auto-detects the older architecture from saved metadata.
• Extended test suite from 244 to 304 tests, adding end-to-end fit/save/load integration, pretrained predict() from synthetic inputs, and v1.0.0 compatibility coverage.

Fixed

• ISEFlow.fit() was broken: NormalizingFlow.fit() was called with mismatched positional arguments, crashing before the first epoch. Switched to keyword arguments.
• NormalizingFlow.fit() off-by-one loop guard caused single-epoch fits to skip training entirely.
• LSTM.fit() did not write checkpoints on train-only runs, causing a crash on the post-training load.
• DeepEnsemble.save() / NormalizingFlow.save() failed when trained with save_checkpoints=False; now uses getattr defaults for missing training attributes.
• ISEFlow.fit() silently overrode the caller's random seed via torch.manual_seed(random_int); removed to restore reproducibility.
• unscale_column year range was off by one (2016–2100 instead of 2015–2100); sector scaler was hard-coded to AIS. Now accepts an ice_sheet kwarg for correct GrIS (1–6) ranges.
• feature_engineer.py: split results were wiped to None immediately after creation; scale_data raised NameError in the explicit-X branch; split_training_data now honours random_state via np.random.default_rng.
• inputs.py: year=list(...) raised TypeError; _map_args was not idempotent; invalid melt_in_floating_cells value 'False' accepted.
• ISEFlow.save() now emits a UserWarning instead of silently swallowing a missing scaler path.
• Fixed HuggingFace Hub snapshot_download to use recursive glob so nested weight files are included.
• ISEFlow_GrIS_DE/NF_v1_0_0 now emit DeprecationWarning on instantiation, matching AIS counterparts.