fix: correct three small defects from the audit

[ChrisW09]

Fixes #21.

Three independent small fixes, grouped because each is a few lines. Happy to split if you'd
rather review them separately.

1. resolve_locations misaligned importance with locations

locs was sorted and de-duplicated while importance kept its original order, so index i
of one no longer described entry i of the other and the trim ranked the wrong entries:

resolve_locations(np.array([5.0, 1.0, 3.0]), min_count=1, max_count=1,
                  importance=np.array([0.1, 9.9, 0.2]))
# [3.]   -- expected [1.], the high-importance location

Now importance is carried through the same permutation (and the same dedupe), and a length
mismatch raises instead of silently misbehaving.

This path is currently unreachable — nothing in the package calls resolve_locations, it
is only re-exported from pretab/core/__init__.py — so this is a latent trap rather than a
live bug. It is public API, though, so it seemed worth closing off.

2. get_feature_names_out silently truncated a wrong-length input_features

BasePreTabTransformer.get_feature_names_out used zip(..., strict=False), so passing too
few names returned a short array rather than raising. Now validated against
n_features_in_, raising InvalidParamError — matching scikit-learn's convention of
rejecting mismatched input_features.

3. LanguageEmbeddingTransformer.fit rejected list input

fit called X.shape[1], which raises AttributeError on a plain list — including the list
in the class's own docstring example, which is masked by # doctest: +SKIP so CI never
caught it. transform already normalizes with np.asarray; fit now does the same.

Not fixed: the doubled name prefix

The issue's fourth item — ndarray input producing num_feature_0__feature_0 — is left as is.
It's cosmetic, and the fix would mean changing the synthesized column stems, which changes
public output names for every ndarray user. That seemed disproportionate to the complaint;
happy to do it if you disagree.

Tests

New file tests/test_minor_defects.py, 9 tests: four for the importance alignment (including
the dedupe path, the mismatch guard, and a no-importance regression check), three for the
input_features validation, two for list/array input to the embedding transformer (using a
stub model, so no sentence-transformers dependency).

Full suite: 489 passed, 9 xfailed. ruff check clean on changed files. pyright 71 → 72,
entirely the import pytest artifact that every test file produces in this checkout (pyright
is configured for a .venv that doesn't exist); no new diagnostic in pretab/.

🤖 Generated with Claude Code