Dev/v0.2.3. Add experimental stacking classifier workflow

[NanoBiostructuresRG]

Summary

This PR adds the MELITE v0.2.3 experimental stacking workflow.

Main changes:

• Adds an opt-in stack model key based on sklearn StackingClassifier.
• Keeps F1 macro as the main scoring and model-selection metric.
• Keeps standalone SVC behavior unchanged as StandardScaler -> SVC.
• Uses stack_method="predict_proba" for stacking.
• Uses a stacked SVC base estimator with StandardScaler -> SVC(probability=True).
• Keeps Random Forest and XGBoost unscaled because they are tree-based models.
• Uses LogisticRegression as the initial final estimator.
• Preserves .pkl export through joblib.
• Adds export and prediction support for stacking models.
• Documents the stacking workflow and its internal CV behavior.
• Adds .tmp_stack_smoke/ to .gitignore for local smoke-test outputs.

Notes

The stacking-internal CV uses the configured split count and random state without repeated splits because sklearn stacking builds out-of-fold meta-features with cross_val_predict. This ensures each training sample contributes exactly one out-of-fold prediction for training the final estimator.

The outer MELITE grid search and reporting workflow still uses the existing repeated CV/F1 evaluation.

Optuna, MLflow, model registry behavior, and .pkl replacement are intentionally out of scope.

Validation

Unit and integration validation:

• python -m pytest -p no:cacheprovider tests\test_model_training.py tests\test_export.py tests\test_predict.py tests\test_version.py
  • 47 passed

Full test suite:

• python -m pytest -p no:cacheprovider
  • 125 passed, 1 warning

Version checks:

• melite --version
  • MELITE 0.2.3
• python -m melite.cli --version
  • MELITE 0.2.3
• python -c "import melite; print(melite.__version__)"
  • 0.2.3

Real stack smoke test:

• Temporary TOML copied from examples/example_config.toml.

• Changed [models].active to ["stack"].

• Changed output path to .tmp_stack_smoke/output/.

• Ran python -m melite.cli run --smoke --config .tmp_stack_smoke\stack_smoke.toml.

• Generated results.csv with model_name = StackingClassifier.

• Exported with:

  python -m melite.cli export --config .tmp_stack_smoke\stack_smoke.toml --csv .tmp_stack_smoke\output\results.csv --outdir .tmp_stack_smoke\output --row 0 --force

• Exported model:

  .tmp_stack_smoke/output/Model_StackingClassifier_sample_pca70.pkl

• Loaded exported .pkl successfully as:

  sklearn.ensemble._stacking.StackingClassifier

• Confirmed:

  • predict: True
  • predict_proba: True
  • prediction output example: [1 1 1 1 1]
  • probability output shape: (5, 2)

Other checks:

• git diff --check
  • passed, with only Windows LF-to-CRLF warnings
• Final branch status before push:
  • clean