fix(splines): expand a missing value to a NaN row in the B/M/I and thin-plate bases

[ChrisW09]

Fixes #87.

BSplineTransformer, MSplineTransformer, ISplineTransformer and a two-feature ThinPlateSplineTransformer encoded a missing input as a finite all-zero row instead of NaN. For the I-spline that row is exactly the encoding of the training minimum. This hit standalone use and the Preprocessor with missing_policy="propagate" or numerical_imputation=None, while every other spline family and the feature maps propagate NaN.

Root cause

The B/M/I design matrices pass the basis values through np.nan_to_num(..., nan=0.0). It is there to clear the NaN that SciPy's BSpline(..., extrapolate=False) returns outside the knot span (under the "extrapolate" and "warn" out-of-range policies), but it cannot tell those apart from the NaN of a missing input, so it zeroed both. In the two-feature thin-plate kernel, np.where(r > 0, r**2 * log(r), 0.0) sent a NaN distance to the 0.0 branch, because NaN > 0 is false.

Fix

• BaseSplineTransformer.transform evaluates _design_matrix only on the observed rows of each feature (after resolve_out_of_range) and leaves the rows of missing values NaN in that feature's basis block. The optional bias column is still prepended as 1. The nan_to_num calls in the B/M/I design matrices stay for the out-of-span values and no longer see missing inputs.
• ThinPlateSplineTransformer._tps_kernel selects the zero branch of the two-dimensional kernel by r == 0 instead of r > 0, so a NaN distance stays NaN.

Behaviour changes

• B/M/I splines return NaN instead of 0 in the basis columns of a missing input, standalone and through the Preprocessor with missing_policy="propagate" or numerical_imputation=None. With include_bias=True the bias column stays 1. Pipelines that disable imputation and pass this output to an estimator that rejects NaN now see the NaN.
• A two-feature ThinPlateSplineTransformer returns NaN for a row with a missing coordinate instead of an all-zero kernel row (with include_bias=True, [1, NaN, ..., NaN] instead of [1, 0, ..., 0]).
• Rows without a missing value are unchanged under every out-of-range policy, and no golden regression baseline changed.

Documentation

• BaseSplineTransformer class docstring: missing values are ignored for knot placement and, at transform time, expand to a NaN row in their feature's basis block (the bias column stays 1), as in the other spline families.
• BaseSplineTransformer._design_matrix docstring: x holds only the observed values, already resolved against the out-of-range policy; transform fills the missing rows with NaN.
• Preprocessor, numerical_imputation: the old text said that with imputation disabled all numerical methods raise on missing values. It now says the scalers, splines and the rbf / relu / sigmoid / tanh feature maps propagate a missing value as NaN in that feature's output, while "ple", "custombin", "fourier" and "polynomial" raise.
• docs/core_concepts/missing_values.md (note on disabled imputation): "tolerate it, so an affected row's output is itself undefined" becomes "propagate it: the affected row's output columns for that feature are NaN (a spline's optional bias column stays 1)".

Tests

• test_mspline_handles_nan (updated; it asserted finite output): the missing row is NaN and every other row is finite. Fails on main, passes here.
• test_bmi_spline_missing_value_expands_to_a_nan_row[{clip,extrapolate,warn,error}-{False,True}-{B,M,I}SplineTransformer]: the basis columns of missing rows are NaN, the bias column is 1, the other rows equal the transform of the same values without the missing rows, and a missing value triggers neither the "warn" nor the "error" reaction. All 24 fail on main, pass here.
• test_bmi_spline_missing_value_only_affects_its_own_feature_block[{B,M,I}SplineTransformer]: with two features, NaN appears only in the block of the feature that is missing. Fails on main, passes here.
• test_tprs_expands_a_missing_value_to_a_nan_row[2]: a two-feature thin-plate spline gives a NaN row for a missing coordinate and leaves the other rows unchanged. Fails on main, passes here.
• test_tprs_expands_a_missing_value_to_a_nan_row[1, 3]: the same for one and three features, whose kernels already propagated NaN. Guard test: passes on both by design.
• test_propagate_expands_a_missing_value_to_a_nan_basis_row[bspline, mspline, ispline]: Preprocessor(missing_policy="propagate") gives a NaN row for the missing value and finite rows otherwise. Fails on main, passes here.
• test_partial_missing_propagates_only_on_missing_rows[BSpline, MSpline, ISpline]: the existing edge-case contract test now also covers B/M/I (added to NAN_PROPAGATING, dropping the comment that they clip a missing value to the fitted boundary). Fails on main, passes here.

Reproduction

Output of the script from #87 on this branch:

BSplineTransformer   NaN -> [nan nan nan nan nan]  x_min -> [1. 0. 0. 0. 0.]
MSplineTransformer   NaN -> [nan nan nan nan nan]  x_min -> [0.8 0.  0.  0.  0. ]
ISplineTransformer   NaN -> [nan nan nan nan nan]  x_min -> [0. 0. 0. 0. 0.]
PSplineTransformer   NaN -> [nan nan nan nan nan]  x_min -> [1. 0. 0. 0. 0.]
ThinPlateSplineTransformer (2 features), [NaN, 5] -> [nan nan nan nan]
Preprocessor(ispline, propagate): NaN row = [nan nan nan nan nan nan nan] | identical to the x_min row: False

Checks

• just test: 1474 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): 1461 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.
• No textual conflicts: this branch merges cleanly with each of the other per-issue branches. The PRs for A partial spline policy mapping drops the "clip" default, and "extrapolate"/"warn" give all-zero basis rows #84 (fix(splines): keep family policy defaults and extrapolate out-of-range rows #121), ISplineTransformer: 200-point trapezoid integration produces wrong and all-zero basis columns #53 (fix(splines): evaluate the I-spline basis in closed form #98), MSplineTransformer: absolute clipping thresholds make the basis depend on feature scale #63 (fix(splines): drop absolute clipping thresholds from the M-spline basis #107), 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), B/M/I spline knot_locations are sized by output_dim and clipped to the data range instead of used as given #81 (fix(splines): honour explicit knot_locations regardless of output_dim #118), Spline knots on tied data land on the range boundary or repeat, producing dead columns and all-zero rows #56 (fix(splines): keep knots unique and strictly interior on tied data #101), Target-aware splines trim selector knots by position, dropping the dominant split, and cap adaptive width at 15 #69 (fix(splines): search the resolved knot window in target-aware placement #109), Standalone transformers drop DataFrame feature names, so reordered columns are silently misaligned #85 (fix(transformers): record and check DataFrame feature names #135) and An invalid task such as "Regression" is silently treated as classification by the splines and selectors #93 (fix(validation): reject an unknown task instead of using classification #128) also edit the spline modules, and 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) edit missing_values.md and pretab/preprocessor.py, in other places.
• The PR for A partial spline policy mapping drops the "clip" default, and "extrapolate"/"warn" give all-zero basis rows #84 (fix(splines): keep family policy defaults and extrapolate out-of-range rows #121) switches the B/M-spline bases to extrapolate=True, so out-of-range rows under "extrapolate" / "warn" are extrapolated instead of zeroed. The two changes are independent: this branch only removes the missing rows before _design_matrix, and the out-of-range rows are left to A partial spline policy mapping drops the "clip" default, and "extrapolate"/"warn" give all-zero basis rows #84. The PRs for ISplineTransformer: 200-point trapezoid integration produces wrong and all-zero basis columns #53 (fix(splines): evaluate the I-spline basis in closed form #98) (closed-form I-spline) and MSplineTransformer: absolute clipping thresholds make the basis depend on feature scale #63 (M-spline clipping) change _design_matrix internals, which now receive observed values only. With this branch merged together with the branches for A partial spline policy mapping drops the "clip" default, and "extrapolate"/"warn" give all-zero basis rows #84, ISplineTransformer: 200-point trapezoid integration produces wrong and all-zero basis columns #53, MSplineTransformer: absolute clipping thresholds make the basis depend on feature scale #63, A 2-D y (multi-output target) crashes the B/M/I splines even when unsupervised, and target-aware methods with unrelated errors #88 and add_missing_indicator / impute_with_indicator feeds the indicator through the scaler and basis; lineage crashes #62, the full test suite passes (1573 passed, 59 skipped, 7 xfailed), and so does this branch merged with B/M/I spline knot_locations are sized by output_dim and clipped to the data range instead of used as given #81 (1490 passed) or with Preprocessor crashes on an all-missing column at fit (numerical, one-hot) or silently drops it (int) #95 (1545 passed).

This change was previously part of #97, which is closed in favour of one pull request per issue.