MELITE v0.2.1. Hardening strict tabular workflows for v0.2.1

[NanoBiostructuresRG]

Summary

This PR prepares MELITE v0.2.1 as a focused hardening release for the strict tabular dataset architecture introduced in v0.2.0.

It does not add a major new feature. Instead, it strengthens model selection, export consistency, installed-package validation, and documentation/metadata.

Changes

Model training

• Fixed [models].active / ACTIVE_MODELS so it now controls which model families are trained.
• MultiModelTrainer now validates active model keys before training.
• Empty active model lists now fail clearly.
• Unknown model keys now raise a clear ValueError.
• Added tests confirming that single-model configurations such as ["svc"] and ["rf"] train only the requested model.

Export hardening

• Hardened export_best_model to use strict dataset loading for registry-based datasets.
• Removed the unsafe legacy fallback to arr.files[0].
• Individual legacy .npz files now require an explicit X array.
• Legacy reduction/level export compatibility remains supported.
• Aggregated legacy archives remain supported, with validation for:
  • 2D feature matrices
  • numeric feature matrices
  • matching X/y lengths

Installed-wheel smoke workflow

• Added an installed-wheel smoke workflow using a minimal strict [datasets.toy] dataset.
• The smoke workflow:
  • builds the wheel,
  • installs it into a temporary environment outside the checkout,
  • verifies public API boundaries,
  • verifies the installed package is imported from the temporary environment,
  • runs melite run --config toy_config.toml --smoke,
  • runs melite export --config toy_config.toml --row 0 --force,
  • verifies expected output artifacts.
• Updated CI to run the new smoke workflow.

Documentation and metadata

• Updated project metadata to v0.2.1.
• Updated CHANGELOG.md with the v0.2.1 hardening entry.
• Updated README and documentation to reflect:
  • strict .npz dataset expectations,
  • [models].active behavior,
  • export hardening,
  • installed-wheel smoke validation.
• Public API remains unchanged:

["Config", "load_datasets", "plot_cv_distributions", "predict", "__version__"]

Validation

The following validations passed locally:

python -m pytest
python -m build --no-isolation
python scripts\smoke_install_wheel.py
python -c "import melite; print(melite.__version__, melite.__all__)"
git diff --check

Build output confirmed:

Successfully built melite-0.2.1.tar.gz and melite-0.2.1-py3-none-any.whl

The installed-wheel smoke workflow completed successfully.

Notes

• This PR intentionally keeps legacy reduction/level export compatibility.
• Legacy individual .npz files must now contain an explicit X key.
• No public API expansion is introduced in this release.
• No PyPI publishing, tagging, or release creation is included in this PR.