[ChrisW09]

Fixes #517

Problem

LogLoss called sklearn.metrics.log_loss(y_true, y_pred) without labels. scikit-learn then infers the classes from y_true and raises when the probability matrix has more columns than observed classes. Any validation or test set that lacks a class therefore:

• stopped fit() when LogLoss was a training/validation metric;
• produced log_loss = nan in evaluate(), where LogLoss is one of the default classification metrics.

Fix

For a 2-D probability matrix with more than one column and integer labels that are all valid column indices (0 <= y < K), LogLoss passes labels=np.arange(K). The probability columns are ordered by the encoded classes 0..K-1, as in the training loop and predict_proba(). Other inputs (1-D positive-class probabilities, non-integer labels, and integers outside 0..K-1 such as raw labels {1, 2, 3}) are passed to scikit-learn unchanged, which infers the classes from y_true as before. The class docstring documents this.

Merge after #494. evaluate() and score(..., metric=(LogLoss(), True)) pass raw labels to probability metrics on main (#493). Raw integer labels that lie inside 0..K-1 without being the encoded indices (e.g. {1, 2} from classes {1, 2, 3}) can't be told apart from encoded ones here; with #494 these callers pass encoded labels, so the scores are correct in every case.

Tests

tests/metrics/test_classification_metrics.py::TestLogLossWithMissingClasses:

• test_multiclass_matrix_uses_all_columns_as_classes
• test_binary_matrix_with_a_single_observed_class
• test_evaluate_and_validation_metrics_on_data_without_one_class: an MLPClassifier with val_metrics={"log_loss": LogLoss()} and a validation set lacking class 2 fits, and evaluate() on data without class 2 returns a finite log loss.
• test_labels_outside_the_column_indices_are_inferred_by_sklearn: LogLoss()([1, 2, 3], P) equals log_loss([1, 2, 3], P).
• test_evaluate_with_labels_other_than_column_indices: for an MLPClassifier trained on labels {1, 2, 3}, evaluate() equals log_loss(y, predict_proba(X), labels=classes_).

The first three fail on main and pass with this change. The last two pass on main too; they guard against scoring raw labels as column indices.

Docs

The LogLoss docstring, shown in the API reference, describes the behaviour. No page needed a change.

Checks

• just test: 3241 passed, 46 skipped, 28 xfailed; coverage 91.32%
• just docs: build succeeded (warnings treated as errors)
• just check: all hooks passed (ruff, prettier, pyright)