fix(preprocessor): accept input_features after fitting on a NumPy array

[ChrisW09]

Fixes #65.

After a fit on a NumPy array, Preprocessor.get_feature_names_out(input_features) raised input_features is not equal to feature_names_in_, even for names of the right length. As a result, Pipeline.get_feature_names_out() failed whenever Preprocessor followed a step that outputs arrays, such as SimpleImputer with default output.

Root cause

Preprocessor.fit names array columns feature_0, feature_1, ... before fitting the inner ColumnTransformer, and get_feature_names_out forwarded input_features to that transformer unchanged. The ColumnTransformer holds the synthetic labels as its feature_names_in_, so it accepted only exactly those names. Preprocessor itself never set feature_names_in_, so it could not apply scikit-learn's rule that an estimator fitted without feature names accepts any input_features of the right length.

Fix

• Preprocessor.fit records feature_names_in_ (an object array) when the input is not a NumPy array and all its column labels are strings, and deletes it on a refit on an array.
• Preprocessor.get_feature_names_out checks that input_features has n_features_in_ entries and, when feature_names_in_ is defined, that it equals feature_names_in_. Without feature_names_in_, any names of the right length are accepted.
• New pretab.compose.inspection.feature_names_out builds the output names. For input_features after an array fit, it maps each fitted label to the given name and rebuilds every step's names from its fitted transformer with the new labels. _without_feature_names hands that call a shallow copy of the step, and of nested Pipeline and FeatureUnion steps, without feature_names_in_. scikit-learn then applies only its length check, and the fitted steps are left untouched.
• clean_feature_names takes a keyword-only rename mapping and reports a renamed column's step under the new name. An array fit with input_features=["age", "income"] therefore gives the same names as a fit on those columns (num_age, not num_feature_0). Without rename, its output is unchanged.

Behaviour changes

• A Preprocessor fitted on a DataFrame or dict whose column labels are all strings now has a feature_names_in_ attribute. A refit on an array removes it.
• After an array fit, get_feature_names_out(input_features) accepts any n_features_in_ names and returns names built from them.
• input_features of the wrong length raises scikit-learn's ValueError: input_features should have length equal to number of features (n), got k instead of input_features is not equal to feature_names_in_.
• to_spec() state and fingerprint_ of newly fitted preprocessors with string column labels change, because the fitted state now includes feature_names_in_. The golden baselines in tests/regression/ pass unchanged. A preprocessor saved by 1.0.0 still loads; it has no feature_names_in_ and therefore accepts any input_features of the right length.

Documentation

• Preprocessor class docstring, Attributes: new feature_names_in_ entry, defined only when every column label is a string.
• Preprocessor.get_feature_names_out docstring: input_features follows scikit-learn's convention; new Raises section for a wrong length or names that differ from feature_names_in_.
• Docstrings of the new feature_names_out and _without_feature_names, and of the rename argument of clean_feature_names.
• docs/core_concepts/outputs_and_inspection.md ("Feature names"): explains the convention and shows make_pipeline(SimpleImputer(), Preprocessor(numerical_method="minmax")).get_feature_names_out() returning ['num_age', 'num_income'].

Tests

• test_array_fit_uses_the_given_input_features (4 cases: minmax, polynomial, bspline with missing_policy="separate_state", ple with add_missing_indicator=True, all with missing values): an array fit with input_features=["age", "income"] gives the same names as a DataFrame fit. Fails on main, passes here.
• test_array_fit_rejects_input_features_of_the_wrong_length: scikit-learn's length message. Fails on main, passes here.
• test_named_fit_records_and_validates_feature_names_in: a DataFrame fit records feature_names_in_, matching names are accepted, other names raise, and a refit on an array removes the attribute. Fails on main, passes here.
• test_pipeline_feature_names_propagate_through_an_array_step: make_pipeline(SimpleImputer(), Preprocessor(...)).get_feature_names_out() returns ["num_age", "num_income"]. Fails on main, passes here.

Reproduction

Output of the script from #65 on this branch:

['age' 'income']
['num_age' 'num_income']
['num_age' 'num_income']

Checks

• just test: 1445 passed, 59 skipped, 7 xfailed; coverage 93.58% (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): 1432 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 Preprocessor.fit crashes when a column name starts with '_' or contains '__' #71 (fix(preprocessor): keep ColumnTransformer step names valid for any label #111) rewrites clean_feature_names in pretab/compose/inspection.py around a new _block helper and a step_to_columns mapping, so the docstring and the prefix logic conflict. Take Preprocessor.fit crashes when a column name starts with '_' or contains '__' #71's version, add the rename=None keyword with rename = rename or {}, map each step's columns through it when building step_to_columns ([rename.get(str(column), column) for column in columns]), and keep the docstring sentence on rename. Resolved this way, the full suite passes (1454 passed, 59 skipped, 7 xfailed).
• The PRs for Standalone transformers drop DataFrame feature names, so reordered columns are silently misaligned #85 (fix(transformers): record and check DataFrame feature names #135) and CrossFittedTransformer(Preprocessor(add_missing_indicator=True)) crashes when a fold's training rows contain no missing value #89 (fix(cross-fitting): keep missing indicators of a wrapped Preprocessor #124) change the imports at the top of pretab/compose/inspection.py (Standalone transformers drop DataFrame feature names, so reordered columns are silently misaligned #85 adds import pandas as pd; CrossFittedTransformer(Preprocessor(add_missing_indicator=True)) crashes when a fold's training rows contain no missing value #89 adds cast and from sklearn.base import clone). Keep both sides. Standalone transformers drop DataFrame feature names, so reordered columns are silently misaligned #85 records feature_names_in_ on the standalone transformers, which _without_feature_names already strips on the copies it uses.
• The PR for output_format="auto" chooses sparse vs dense per transform batch, breaking dense-only estimators at predict #73 (fix(preprocessor): resolve output_format='auto' once per fit #113) also changes Preprocessor.fit near the new lines; git merges it cleanly. output_format="auto" chooses sparse vs dense per transform batch, breaking dense-only estimators at predict #73 also changes the fingerprint of newly fitted preprocessors.
• The PR for Preprocessor routes integer column labels positionally, silently swapping columns #55 (fix(preprocessor): route columns to the ColumnTransformer by label #100) conflicts in get_feature_names_out (the input_features docstring and the method body), where it compares names by their string form. Keep this PR's side of both hunks: it already accepts the labels seen at fit. With that resolution the full suite passes (1455 passed).
• The PRs for Preprocessor.fit crashes on any bool column under default settings #54 (fix(preprocessor): encode boolean columns as binary categoricals #99), Preprocessor(numerical_method="box-cox") raises at transform for values below the training minimum #59 (fix(preprocessor): keep box-cox finite below the training minimum #104), Preprocessor.transform does not check n_features_in_ for ndarray input; extra columns are silently misaligned #72 (fix(preprocessor): reject NumPy input of the wrong width at transform #112), Preprocessor fitted on a DataFrame rejects a same-width NumPy array at transform (and vice versa) #78 (fix(preprocessor): match array input by position after a named fit #116), 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) and Preprocessor rejects polars DataFrames, so pipelines with set_output(transform="polars") crash #94 (fix(preprocessor): accept polars DataFrames as input #129) also append tests at the end of tests/integration/test_preprocessor.py, so expect an append/append conflict there: keep both blocks.
• Follow-up with Preprocessor fitted on a DataFrame rejects a same-width NumPy array at transform (and vice versa) #78: once both land, Preprocessor._align_input from Preprocessor fitted on a DataFrame rejects a same-width NumPy array at transform (and vice versa) #78 should use feature_names_in_ to tell a named fit from an array fit. See the PR for Preprocessor fitted on a DataFrame rejects a same-width NumPy array at transform (and vice versa) #78 (fix(preprocessor): match array input by position after a named fit #116).
• Follow-up with Preprocessor rejects polars DataFrames, so pipelines with set_output(transform="polars") crash #94: this PR detects an array fit with isinstance(X, np.ndarray). Preprocessor rejects polars DataFrames, so pipelines with set_output(transform="polars") crash #94 reads a list of rows like an array and names its columns feature_0, feature_1, .... Once it lands, the check should be not has_column_labels(X), or a fit on a list of rows records feature_names_in_ = ["feature_0", ...].

This change was previously part of #75, which is closed in favour of one pull request per issue.