[ChrisW09]

Fixes #93.

An invalid task such as the typo "Regression" was accepted by the knot splines, the Preprocessor configuration and the location selectors, and silently treated as classification. Users got classification-tree knots, LightGBM's binary objective on a regression target, an unrelated scikit-learn "Unknown label type: continuous" error on a float target, or the classification preset method "ple".

Root cause

task was validated only by PLETransformer, the feature maps and CrossFittedTransformer, each with its own message. BaseLocationSelector.select took task or "regression" unchecked, and the CART and LightGBM selectors dispatched with if task == "regression": ... else: <classification>, so every other value took the classification branch. PreprocessorConfig.from_params did not check task, and _preset_numerical_method returned "ple" for anything but "regression".

Fix

• New validate_task(task, estimator, *, allow_none=False) and TASKS in pretab/core/parameters.py. The match is exact, and an invalid value raises InvalidParamError through invalid_param_error, so every estimator reports it in the same format.
• validate_task runs at fit in BaseSplineTransformer (B-, M- and I-splines), NaturalCubicSplineTransformer and CubicRegressionSplineTransformer (with allow_none=True; None still means regression). PLETransformer, BaseCenterExpansion (the feature maps) and CrossFittedTransformer._fit_full use it instead of their own checks.
• BaseLocationSelector.select validates task (None is treated as "regression"). CARTLocationSelector and LightGBMLocationSelector now choose the classifier tree and the binary objective only for task == "classification".
• Preprocessor._resolved_params validates task before a preset derives numerical_method from it, so fit and get_resolved_config() both raise. PreprocessorConfig.from_params validates it as well.

Behaviour changes

• Any task other than "regression" or "classification" raises InvalidParamError: at fit of every transformer that takes task, in the location selectors, and in Preprocessor.fit / get_resolved_config(). The Preprocessor raises even when the chosen method ignores the task (for example minmax).
• Preprocessor(task=None) now raises. Before, the target-aware splines treated it as regression and untargeted methods ignored it, while ple and every preset rejected it with PLE's error. The standalone splines and the location selectors still accept None as "regression".
• The invalid-task messages of PLETransformer, the feature maps (rbf, relu, sigmoid, tanh) and CrossFittedTransformer change to the shared format, for example PLETransformer.task = 'Regression' is invalid. followed by the constraint and the valid values. The exception type stays InvalidParamError.
• New public helper pretab.core.parameters.validate_task and constant TASKS. Valid task values give the same results as before.

Documentation

• docs/core_concepts/configuration.md: the presets paragraph says that task is matched exactly and that any other value, such as "Regression", raises an InvalidParamError from fit and get_resolved_config() instead of falling back to the classification preset. The "validated at fit time" warning lists an invalid task.
• Docstrings: validate_task; BaseLocationSelector.select (task, new Raises); the pretab/compose/config.py module docstring, PreprocessorConfig and its from_params Raises; the Preprocessor task parameter, and Raises sections for _resolved_params and get_resolved_config; _preset_numerical_method; the task parameter of BaseSplineTransformer, CubicRegressionSplineTransformer and NaturalCubicSplineTransformer.

Tests

• test_invalid_task_raises[Regression, classif, None] (tests/compose/test_config.py): PreprocessorConfig rejects each value. Fails on main, passes here.
• test_selectors_reject_an_unknown_task (CART and LightGBM, each with "Regression", "regresion", "" and 1): select raises InvalidParamError. Fails on main, passes here.
• test_unknown_task_raises_at_fit[bspline, cubicspline, ispline, mspline, naturalspline, ple, rbf, relu, sigmoid, tanh]: every registered method with a task parameter raises the shared message at fit for task="Regression". Fails on main, passes here; the knot splines fitted there, and ple and the feature maps raised with their old messages.
• test_preprocessor_unknown_task[bspline, naturalspline, rbf, ple, minmax]: Preprocessor(task="Regression").fit raises Preprocessor.task = 'Regression' is invalid. Fails on main, passes here.
• test_invalid_task_raises_instead_of_resolving[None, standard, expanded, adaptive] (tests/integration/test_presets.py): get_resolved_config() raises instead of resolving a method. Fails on main, passes here.
• test_ple_invalid_task_raises, test_ple_unsupported_task, test_rbf_invalid_task, test_relu_invalid_task, test_sigmoid_invalid_task and test_tanh_invalid_task_raises: updated to the shared message. They fail on main only because of the wording.
• test_cart_task_none_means_regression and test_spline_task_none_places_knots_for_regression: task=None still gives the regression locations and knots. Guard tests: pass on both by design.

Reproduction

Output of the script from #93 on this branch, with each call wrapped separately. The last line is an added check that a valid task still resolves the preset as before:

PLETransformer(task='Regression') -> InvalidParamError: PLETransformer.task = 'Regression' is invalid.
Constraint: must be 'regression' or 'classification'
Valid values: ['classification', 'regression']
RBFExpansionTransformer(task='Regression') -> InvalidParamError: RBFExpansionTransformer.task = 'Regression' is invalid.
Constraint: must be 'regression' or 'classification'
Valid values: ['classification', 'regression']
BSplineTransformer(task='Regression') -> InvalidParamError: BSplineTransformer.task = 'Regression' is invalid.
Constraint: must be 'regression', 'classification', or None (treated as 'regression')
Valid values: ['classification', 'regression']
Preprocessor('bspline', task='Regression') -> InvalidParamError: Preprocessor.task = 'Regression' is invalid.
Constraint: must be 'regression' or 'classification'
Valid values: ['classification', 'regression']
Preprocessor('naturalspline', task='Regression') -> InvalidParamError: Preprocessor.task = 'Regression' is invalid.
Constraint: must be 'regression' or 'classification'
Valid values: ['classification', 'regression']
Preprocessor('bspline', task='Regression'), float y -> InvalidParamError: Preprocessor.task = 'Regression' is invalid.
Preprocessor(preset='standard', task='Regression').get_resolved_config() -> InvalidParamError: Preprocessor.task = 'Regression' is invalid.
Constraint: must be 'regression' or 'classification'
Valid values: ['classification', 'regression']
Preprocessor(preset='standard', task='regression') resolves numerical_method to 'bspline'

Checks

• just test: 1470 passed, 59 skipped, 7 xfailed; coverage 93.66% (gate 90%).
• Minimum dependencies (CI job: Python 3.10, numpy 1.24.4, pandas 2.0.3, scipy 1.10.1, scikit-learn 1.6.0): 1457 passed, 72 skipped, 7 xfailed; scripts/quickstart.py passes.
• just docs: strict build succeeds.
• just check: all pre-commit and pre-push hooks pass (ruff format, ruff lint, prettier, file hygiene, pyright with 0 errors).
• Import direction between subpackages is unchanged.

Merge notes

• Base: main.
• The PR for Target-aware location selectors return too few, duplicate, or boundary locations on tied/discrete features #57 (fix(placement): fill tied-data locations with distinct interior points #102) conflicts in pretab/core/selectors.py in two places. In the imports, Target-aware location selectors return too few, duplicate, or boundary locations on tied/discrete features #57 adds uniform_knots to the .knots import and this PR adds from .parameters import validate_task. In the docstring of BaseLocationSelector.select, Target-aware location selectors return too few, duplicate, or boundary locations on tied/discrete features #57 rewrites the Returns text and this PR adds a Raises section. Keep both sides in each.
• The PR for LightGBM placement with task="classification" always uses the binary objective on raw labels #61 (fix(placement): label-encode classification targets for LightGBM #105) rewrites the LightGBM objective in LightGBMLocationSelector._ordered_candidates (label encoding, binary or multiclass). Keep LightGBM placement with task="classification" always uses the binary objective on raw labels #61's objective block and **objective, and turn its condition around so that it dispatches on "classification" (if task == "classification": <label encoding> else: <regression>), as this PR does for the inline conditionals it replaces.
• The PR for A 2-D y (multi-output target) crashes the B/M/I splines even when unsupervised, and target-aware methods with unrelated errors #88 (fix(splines): ignore y when unsupervised and reject a multi-output target #123) conflicts in pretab/core/selectors.py in the imports (keep both validate_task and A 2-D y (multi-output target) crashes the B/M/I splines even when unsupervised, and target-aware methods with unrelated errors #88's single_target) and in the select docstring (take A 2-D y (multi-output target) crashes the B/M/I splines even when unsupervised, and target-aware methods with unrelated errors #88's y entry and this PR's task entry).
• The PR for CrossFittedTransformer converts DataFrames to ndarray; a wrapped Preprocessor treats every column as categorical #58 (fix(cross-fitting): pass DataFrames to the wrapped transformer unchanged #103) conflicts in CrossFittedTransformer._fit_full: keep validate_task(self.task, type(self).__name__) together with CrossFittedTransformer converts DataFrames to ndarray; a wrapped Preprocessor treats every column as categorical #58's X_arr = _as_2d(X).
• Each of these four pairs, resolved this way, passes the full test suite (with Target-aware location selectors return too few, duplicate, or boundary locations on tied/discrete features #57: 1494 passed; LightGBM placement with task="classification" always uses the binary objective on raw labels #61: 1474; A 2-D y (multi-output target) crashes the B/M/I splines even when unsupervised, and target-aware methods with unrelated errors #88: 1523; CrossFittedTransformer converts DataFrames to ndarray; a wrapped Preprocessor treats every column as categorical #58: 1474; each with 59 skipped, 7 xfailed). The PR for CART location selector can drop the dominant split because spacing is enforced in ascending location order #70 (fix(placement): rank CART splits by impurity decrease before spacing #110) changes the CART candidate ranking next to the tree construction rewritten here; git merges the two without a conflict.

This change was previously part of #97, which is closed in favour of one pull request per issue.