Gh diagnostics

[GernotMaier]

No description provided.

[Copilot]

Pull request overview

This PR adds support scaffolding for TMVA-style gamma/hadron classification features and related preprocessing/tests, alongside a small change to classification efficiency bookkeeping.

Changes:

• Adds TMVA-style feature selection/configuration paths for classification training.
• Extends classification extra columns and clipping ranges for TMVA-derived variables.
• Adds tests for binning behavior and optional SizeSecondMax handling.

Reviewed changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
src/eventdisplay_ml/features.py	Adds TMVA-style classification feature lists and clip interval updates.
src/eventdisplay_ml/data_processing.py	Adds TMVA-style training data loading path and extra classification-derived columns.
src/eventdisplay_ml/config.py	Reads tmva_style from classification model parameters during training configuration.
src/eventdisplay_ml/evaluate.py	Changes stored signal/background count columns in efficiency output.
tests/test_data_processing.py	Adds tests for zenith and energy interpolation binning.
tests/test_classification_apply_interpolation.py	Adds test for missing SizeSecondMax behavior in standard classification.
docs/changes/58.feature.md	Adds changelog fragment for TMVA-style classification support.

Comments suppressed due to low confidence (2)

src/eventdisplay_ml/data_processing.py:810

• The TMVA-style flag is only handled while loading training data; the application path still loads features.features(...) and runs the standard telescope flattening path. A model trained with this branch will therefore be applied with a different feature set, with TMVA-only inputs such as SizeSecondMax, DispAbsSumWeigth, or raw Xcore/Ycore missing or filled as NaN at inference.

    tmva_style = model_configs.get("tmva_style", False)
    if tmva_style and analysis_type == "classification":
        _logger.info("Using TMVA-style features for classification")
        branch_list = features_module.features_tmva_style(analysis_type, training=True)

src/eventdisplay_ml/data_processing.py:885

• Building the TMVA-style training frame directly from ROOT branches bypasses extra_columns() and apply_clip_intervals(). As a result, chi-square and SizeSecondMax inputs are not clipped/log-transformed and Xcore/Ycore are not converted to Core_Distance, so the model is trained on raw variables instead of the TMVA-style preprocessed features defined elsewhere.

                # For TMVA-style classification, skip telescope flattening (use event-level features only)
                if tmva_style and analysis_type == "classification":
                    _logger.info("Converting to pandas (no telescope flattening for TMVA style)")
                    # Build DataFrame directly to stay compatible across awkward versions
                    # (some versions do not provide ak.to_pandas).
                    df_flat = pd.DataFrame({name: _to_numpy_1d(df[name]) for name in df.fields})