[0.4.4] - 2026-06-15

Added

• save(path=None) / load(path): persist a fitted model to a pickle file and restore it
  without retraining. path=None saves rfscorer.pkl to the current directory; a directory
  path saves rfscorer.pkl inside it; a file path saves directly. On major or minor version
  mismatch, load() emits a UserWarning and continues loading.
• save_zip(path=None) / load_zip(path): save/restore the model as a zip archive bundling
  rfscorer.pkl, metadata.json (version, parameters, fit statistics), probability-table CSVs,
  and plot PNGs for all computed model kinds. path=None saves scorer.zip to the current
  directory. Intended for research sharing and artifact management.
• Tutorial notebooks (tutorial_beginner_en.ipynb / tutorial_beginner_ja.ipynb): added
  Section 10 covering save() / load() usage with a Google Colab persistence guide.
  Also added a commented # !pip install rfscorer line to the import cell.

Changed

• Terminology unification: renamed all eval-prefixed names to gt (ground truth)
  to align with the unified terminology in docs/glossary.md (正解データ / ground truth data).
  Breaking changes:
  • fit(df_obs, df_eval, ...) → fit(df_obs, df_gt, ...)
  • evaluate(df_rec, df_eval, ...) → evaluate(df_rec, df_gt, ...)
  • split_by_date(..., evaluation_days=7, ...) → split_by_date(..., gt_days=7, ...);
    return value documented as (df_obs, df_gt) instead of (df_obs, df_eval).
  • Attribute record_num_eval → record_num_gt
  • Attribute evaluation_start_ → gt_start_
  • Attribute evaluation_end_ → gt_end_
  • show() output label evaluation: → ground_truth:
  • Error message "No events observed in evaluation period" → "No events observed in ground truth period"
• fit(): the datetime column in df_gt is now optional. Only user and item columns
  are required for fitting. The gt_start_ and gt_end_ attributes (which depended on
  df_gt's datetime column) have been removed.