[1.0.0] - 2026-06-30

First stable public release. The public API under the top-level xai4tsc
namespace is now considered stable and follows semantic versioning.

Added

Dual-use design: standalone experiment runner (python -m experiment_runner.main)
and an importable, config-agnostic xai4tsc package.
Data (xai4tsc.data): UCR/UEA loaders, pre-split local datasets, and a
synthetic-dataset framework (SyntheticDataset ABC) with a shipped
freq_shapes dataset, label/multi-hot encoders, and ground-truth
localization metadata.
Models (xai4tsc.models): FCN, LeNet, ResNet (Wang et al. 2017),
LSTM, XLSTM (Beck et al. 2024), and PatchTST (Nie et al. 2023), all
clean-room PyTorch implementations behind a common ModelBase wrapper.
Explainers (xai4tsc.xai): Captum-backed time-domain attributions,
frequency / time-frequency explainers (incl. a paper-faithful FreqRISE), and
TSHAP, all behind the Explainer ABC with declared data_applicability.
Evaluation (xai4tsc.evaluation): a unified EvaluatorBase design — the
Quantus library is adapted through a single QuantusEvaluator, alongside native
frequency / time-frequency perturbation metrics (FrequencyEvaluator,
TimeFrequencyEvaluator) and the ground-truth TimeFrequencyAUCEvaluator
localization metric, all with declared domain applicability.
Extensibility: runtime registries (MODELS, EXPLAINERS, METRICS) with
register_* hooks.
Documentation, contribution guide, and a research disclaimer.