v.1.2.0

[1.2.0] — 2026-05-11 (package) | Model: v1.1.0

Added

• NormalizingFlow.get_latent_representation() now supports v1.0.0 weights via a legacy_v1_0_0 path: pushes a zero vector through the forward transform (deterministic) rather than sampling from the conditional base distribution. NormalizingFlow.load() auto-detects the version from saved metadata and sets the flag automatically, so v1.0.0 and v1.1.0 inference are both correct without user intervention.
• docs/model_versions.rst — new reference page documenting all ISEFlow weight versions, per-version input schemas, migration guides (v1.0.0 → v1.1.0), from_absolute_forcings() usage, and weight resolution order. Linked from docs/index.rst.
• examples/recreate_manuscript_results.py — end-to-end script that downloads v1.0.0 test splits and scalers from HuggingFace Hub (pvankatwyk/iseflow-datasets), runs inference with v1.0.0 pretrained weights, and reports held-out MSE for both AIS and GrIS.

Changed

• NormalizingFlow.get_latent_representation() docstring expanded to document both the legacy deterministic path (v1.0.0) and the stochastic sampling path (v1.1.0+), including the behavioural difference and its downstream effect on DeepEnsemble inputs.