perf(preprocessor): slice transform blocks from output_indices_

[ChrisW09]

Fixes #20.

Problem

transform(..., return_array=False) — the default — transformed the data once through the
ColumnTransformer, then re-ran each per-feature pipeline a second time purely to read
.shape[1]:

width = transformer.transform(X[columns]).shape[1]

So every pipeline ran twice per call, for identical numbers.

Fix

column_transformer_.output_indices_ already records the exact output slice for each
transformer — and output_dims_, on this same class, already reads it. Use it here too.

It is also strictly more precise than what it replaces: the width = 1 fallback for a
non-estimator entry was a guess, whereas the indices are authoritative.

Result

return_array=True :  21.9 ms -> 21.9 ms
return_array=False:  43.5 ms -> 21.7 ms      (20k rows x 20 features, PLE)

The dict path now costs the same as the array path, which is what it should always have
cost — the split is pure bookkeeping.

Returned values are unchanged: same keys, same widths, same numbers.

Tests

Four added to tests/test_preprocessor.py:

• test_dict_transform_calls_each_pipeline_once — wraps each fitted pipeline's transform
  in a counter and asserts every count is exactly 1. This is the direct regression guard and
  fails on the old code with counts of 2.
• test_dict_blocks_reassemble_the_stacked_array — hstacking the blocks reproduces
  return_array=True
• test_dict_block_widths_match_output_dims — widths agree with output_dims_ and sum to
  len(get_feature_names_out())
• test_dict_keys_cover_every_input_feature

Full suite: 484 passed, 9 xfailed. pyright back to its pre-existing 71 errors — my first pass
added 12, all the familiar "transform returns dict | ndarray, pyright can't narrow it"
noise, resolved with the cast(...) pattern the repo already uses in
test_adaptive_output_dim.py.

One incidental cleanup: tests/test_preprocessor.py had a pre-existing ruff I001
(unsorted imports) on the exact block this change edits, so I sorted it rather than leave a
lint error attached to my diff. That file is now ruff-clean; the unrelated RUF100 in
pretab/__init__.py is untouched and still present on main.

🤖 Generated with Claude Code