fix(preprocessor): keep box-cox finite below the training minimum

[ChrisW09]

Fixes #59.

Preprocessor(numerical_method="box-cox") raised a ValueError at transform for any value more than about 0.1% of the training range below the training minimum. This broke prediction on new data and plain cross-validation, where three of five folds scored nan in the issue's example.

Root cause

Box-Cox needs strictly positive input, so get_numerical_transformer_steps puts a MinMaxScaler(feature_range=(1e-3, 1)), fitted on the training data, in front of PowerTransformer(method="box-cox"). With the default clip=False, the scaler maps a value far enough below the training minimum to <= 0, and PowerTransformer.transform rejects it. Positivity was guaranteed only for in-sample values.

Fix

• New private _PositiveMinMaxScaler in pretab/compose/factory.py: a MinMaxScaler whose transform floors the output at feature_range[0] (1e-3) with np.maximum. NaN passes through.
• get_numerical_transformer_steps uses it for the scale_positive step of "box-cox".

A value below the training minimum now transforms like the minimum. Every other output, including values above the training maximum, is unchanged. MinMaxScaler(clip=True) was not used because it would also clip values above the training maximum, which extrapolate fine today.

Behaviour changes

• Box-Cox output for any value below the training minimum equals the output for the minimum. This includes values within about 0.1% of the training range below the minimum, which stayed positive before and gave a slightly lower finite output.
• The scale_positive step of a newly fitted box-cox preprocessor is a _PositiveMinMaxScaler, so its to_spec() state and fingerprint_ differ from 1.0.0. Preprocessors with other numerical methods keep their fingerprint. to_spec()/from_spec() and pickle round trips keep the floor and the fingerprint. A box-cox preprocessor saved by 1.0.0 still holds a plain MinMaxScaler and keeps the old behaviour until it is refitted.

Documentation

• Preprocessor class docstring (Notes): "box-cox" first rescales each column into [1e-3, 1] with a min-max scaler fitted on the training data, floored at 1e-3, so a value below the training minimum transforms like the minimum and values above the maximum still extrapolate.
• _PositiveMinMaxScaler docstring.
• No page under docs/ describes the box-cox internals (they only list the methods), so none needed a change.

Tests

• test_boxcox_transforms_values_below_the_training_minimum: values 0.5 and 5 below the training minimum transform to finite values equal to the transform of the minimum. Fails on main (ValueError), passes here.
• test_boxcox_survives_cross_validation: the issue's 5-fold cross_val_score returns finite scores for every fold. Fails on main, passes here.
• test_boxcox_still_extrapolates_above_the_training_maximum: a value above the training maximum still maps above the maximum's output, so the upper end is not clipped. Guard test: passes on both by design.

Reproduction

Output of the script from #59 on this branch:

[[-2.53352473]]
[[-2.53352473]]
[0.99970011 0.99958985 0.99922595 0.99956836 0.99959617]

The value 0.5 below the training minimum now transforms like the minimum, and all five folds score.

Checks

• just test: 1441 passed, 59 skipped, 7 xfailed; coverage 93.65% (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): 1428 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 PRs for add_missing_indicator / impute_with_indicator feeds the indicator through the scaler and basis; lineage crashes #62 (fix(preprocessor): keep the missing indicator out of the representation #106) and Preprocessor crashes on an all-missing column at fit (numerical, one-hot) or silently drops it (int) #95 (fix(preprocessor): keep a block for columns with no observed value at fit #130) also change pretab/compose/factory.py. add_missing_indicator / impute_with_indicator feeds the indicator through the scaler and basis; lineage crashes #62 edits the sklearn.impute import two lines below the new import numpy as np. Git merges both cleanly with this branch.
• The PRs for Preprocessor.fit crashes on any bool column under default settings #54 (fix(preprocessor): encode boolean columns as binary categoricals #99), Preprocessor.get_feature_names_out(input_features) fails after fitting on an ndarray, breaking Pipeline.get_feature_names_out() #65 (fix(preprocessor): accept input_features after fitting on a NumPy array #108), 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.

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

[mkumar73]

• Updated main,
• no issues in CI
• Changes looks appropriate for merging.