v1.4.0

Highlights

• Registry-based feature generator system enabling pluggable feature engineering for models (#245).
• Predicted vs actual scatter plots for backtests with improved visualization and 1-based horizon distance (#248, #270).
• Flaky CI tests fixed via GitHub fetch caching for tests that pull from raw.githubusercontent.com (CLIM-552).

Features

• Add registry-based feature generator system (#245)
• Add predicted vs actual scatter plot for backtests (#248)
• Improve predicted vs actual plot and use 1-based horizon distance (#270)
• Support URL input for dataset_csv in chap eval (#250)
• Add --plot flag to chap eval command (#251)
• Unique predict filenames, Runner._execute refactor, and --dry-run for eval (#252)
• Add MAPE metric (#265)
• Cache GitHub fetches in tests to fix flaky CI (CLIM-552)
• Also cache pooch downloads from GitHub in tests

Bug Fixes

• Filter locations with no disease data in first train split (#255)
• Pin AR model versions with python <3.13 constraint (#249)
• Exclude doc tests from default pytest run and fix placeholder bash blocks
• Exclude tests/.github_cache from pyright

Dependencies & Build

• Migrate pydantic-geojson to geojson-pydantic
• Switch from hatchling to uv build backend
• Upgrade all dependencies to latest versions
• Bump ipython to >=9.12.0, pygments to 2.20.0
• Bump cryptography and requests

Documentation

• Update chap evaluate to chap eval with URL datasets (#271)
• Flesh out predict and train section in describe_model.md
• Consolidate R-model renv docs and remove r_models page
• Train/predict documentation improvements (#253)
• Final CHAP->Chap naming cleanup (#269)