fix: assorted minor defects (importance misalignment, names, docs)

[ChrisW09]

Summary

Four small, independent defects found while auditing the package. None is urgent; grouped
here to avoid four near-empty issues. Happy to split them if the maintainers prefer one
issue per item.

---

1. resolve_locations misaligns importance with the sorted locations

pretab/core/locations.py:83-87 sorts and dedupes locs and then passes the caller's
original-order importance array to trim_to_count, so index i of importance no longer
refers to locs[i]. The wrong locations are kept.

import numpy as np
from pretab.core.locations import resolve_locations

locs = np.array([5.0, 1.0, 3.0])
imp = np.array([0.1, 9.9, 0.2])   # location 1.0 is by far the most important
print(resolve_locations(locs, min_count=1, max_count=1, importance=imp))

[3.]

Expected [1.]. Currently unreachable — nothing in the package calls
resolve_locations; it is only re-exported by pretab/core/__init__.py:34. So this is a
latent trap for whoever wires it up rather than a live bug, but it is exported API.

Fix: reorder importance alongside locs, e.g. sort with np.argsort and index
importance by the same permutation before the dedupe, or drop the importance parameter
until there is a caller.

---

2. get_feature_names_out silently truncates a wrong-length input_features

pretab/core/base.py:55 uses zip(input_features, self._output_sizes(), strict=False), so
passing too few names quietly returns a short array instead of raising. scikit-learn's
convention is to validate with _check_feature_names_in and raise on a length mismatch.

Fix: validate len(input_features) == self.n_features_in_ and raise InvalidParamError, or
use strict=True.

---

3. NumPy input produces doubled name prefixes

import numpy as np
from pretab import Preprocessor

pre = Preprocessor(numerical_method="minmax").fit(np.random.default_rng(0).normal(size=(50, 2)), None)
print(list(pre.get_feature_names_out()))

['num_feature_0__feature_0', 'num_feature_1__feature_1']

Preprocessor.fit synthesises column names as feature_{i} (pretab/preprocessor.py:345)
and then names each ColumnTransformer entry num_{feature} (preprocessor.py:394), so the
ColumnTransformer's {name}__{column} joining repeats the stem. Cosmetic, but it is what
users see in the array's column labels.

Fix: nothing structural — either accept it, or synthesise plainer stems (x{i}) for ndarray
input so the result reads num_x0__x0.

---

4. LanguageEmbeddingTransformer docstring example cannot run

pretab/transformers/embeddings/language_transformer.py:76 does X.shape[1], but the
docstring example at :39 passes a plain list:

>>> embeddings = transformer.fit_transform([["red"], ["blue"], ["green"]])  # doctest: +SKIP

A list has no .shape, so this raises AttributeError. It is masked by # doctest: +SKIP,
so CI does not catch it. transform already normalises with np.asarray
(language_transformer.py:105-107); fit does not.

Fix: apply the same np.asarray normalisation in fit — which also makes the documented
example work and lets the +SKIP be narrowed to just the model-loading lines.

---

Environment

• pretab 0.1.0 (main @ 51c3043)
• Python 3.11.15, numpy 2.4.6, pandas 2.3.3, scikit-learn 1.9.0, scipy 1.17.1
• macOS (darwin 25.5.0)