fix(preprocessor): resolve output_format='auto' once per fit

[ChrisW09]

Fixes #73.

With output_format="auto", Preprocessor.transform chose between a dense array and a CSR matrix separately on every call, from that batch's density. The same fitted preprocessor therefore returned sparse output for some single rows, and dense-only estimators such as HistGradientBoostingRegressor failed at predict although fit had succeeded.

Root cause

"auto" was resolved inside compute_output_report, which transform calls with each batch's output. fit stored no decision, so the container followed the density of whatever was being transformed. Rows from PLE or one-hot blocks easily fall below the 0.3 threshold on their own.

Fix

• New resolve_output_format and private _density in pretab/compose/output.py. resolve_output_format resolves "auto" to "dense" or "sparse" from an array's density. compute_output_report uses both; explicit formats behave as before.
• Preprocessor.fit calls fit_transform on the inner ColumnTransformer instead of fit, at no extra cost because ColumnTransformer.fit runs fit_transform internally. It resolves the format once from the training output and stores it in the new fitted attribute output_format_. A dense training output goes through np.asarray first, as in transform, because a global transform_output="pandas" can make the inner transformer return a DataFrame.
• Preprocessor.transform passes output_format_ to compute_output_report, so every call returns the container chosen at fit. output_report_ still reports each call's own shape and density. A preprocessor without output_format_ (fitted by 1.0.0 and loaded from a spec or pickle) keeps the per-call choice.

Behaviour changes

• With output_format="auto", every transform returns the container chosen at fit, including for a single row.
• New fitted attribute output_format_ ("dense" or "sparse"). Explicit formats are stored as given.
• output_format_ is part of the fitted state, so to_spec() and fingerprint_ of a newly fitted preprocessor differ from 1.0.0 for every output_format, not only "auto".

Documentation

• Preprocessor docstring: the output_format parameter says "auto" is decided once at fit from the training output's density; new output_format_ attribute.
• Docstrings of compute_output_report and the new resolve_output_format.
• docs/core_concepts/outputs_and_inspection.md: the output_format entry says the choice is made once at fit, from the training output, and stored in output_format_, so every transform returns the same container.

Tests

• test_auto_format_is_fixed_at_fit_for_dense_training_output: on the diabetes data, output_format_ == "dense", all 442 single rows come back as ndarray, and output_report_ describes the one-row call. Fails on main, passes here.
• test_auto_format_is_fixed_at_fit_for_sparse_training_output: a 20-level one-hot column gives output_format_ == "sparse", and one row and the full frame both come back sparse. Fails on main, passes here.
• test_explicit_formats_are_stored_as_resolved[dense], [sparse]: explicit formats are stored as given. Fails on main (no output_format_), passes here.

Reproduction

Output of the script from #73 on this branch:

ndarray ndarray
0 of 442 single rows come back sparse
[118.12429962]

Checks

• just test: 1442 passed, 59 skipped, 7 xfailed; coverage 93.66% (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): 1429 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 Preprocessor.fit crashes on any bool column under default settings #54 (fix(preprocessor): encode boolean columns as binary categoricals #99), Preprocessor routes integer column labels positionally, silently swapping columns #55 (fix(preprocessor): route columns to the ColumnTransformer by label #100), Preprocessor with one-hot encoding or preset="expanded" fails under set_config(transform_output="pandas") #76 (fix(preprocessor): keep the internal ColumnTransformer on default output #114) 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) change the same self.column_transformer_.fit(X, y) line in Preprocessor.fit, so each conflicts there. Keep this PR's fit_transform block and apply the other change to it: fit_transform(self._column_transformer_input(X), y) for Preprocessor.fit crashes on any bool column under default settings #54 and Preprocessor routes integer column labels positionally, silently swapping columns #55, Preprocessor with one-hot encoding or preset="expanded" fails under set_config(transform_output="pandas") #76's set_output(transform="default") line above it, and Preprocessor crashes on an all-missing column at fit (numerical, one-hot) or silently drops it (int) #95's try/except around the fit_transform call. With Preprocessor with one-hot encoding or preset="expanded" fails under set_config(transform_output="pandas") #76 resolved this way, the full suite passes (1445 passed, 59 skipped, 7 xfailed). Once Preprocessor with one-hot encoding or preset="expanded" fails under set_config(transform_output="pandas") #76 lands, the inner transformer always returns arrays, so the np.asarray conversion here is redundant but harmless.
• tests/integration/test_output_format.py also gets an append/append conflict with the PR for Preprocessor with one-hot encoding or preset="expanded" fails under set_config(transform_output="pandas") #76 (fix(preprocessor): keep the internal ColumnTransformer on default output #114): keep both blocks.
• The PRs for 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) and 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) change Preprocessor.fit and transform nearby. Git merges both cleanly. Preprocessor.get_feature_names_out(input_features) fails after fitting on an ndarray, breaking Pipeline.get_feature_names_out() #65 also changes the fingerprint of newly fitted preprocessors, because it adds feature_names_in_.

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