[jonathan-taylor]

Follows #2, which was merged before this commit was pushed. As in scikit-learn, the design and the response are built separately: a ModelSpec builds X, and a response object builds y from the same data.

Responses separate from ModelSpec. ModelSpec loses its response= parameter and build_response; neither has been released, so nothing breaks. Surv, Binomial and Multinomial are fit and built on their own with fit, transform or fit_transform. A dataframe of another backend, such as polars, gives a response in that backend (for polars, a categorical comes back as an Enum).

wildcard(exclude=...) takes a response and leaves out its columns, so the response is named once and cannot leak into its own design:

response = Surv('time', 'status')
y = response.fit_transform(D)
X = MS([wildcard(exclude=response)]).fit_transform(D)

MS([wildcard(exclude='y')]) is the design of R's y ~ .. A bare wildcard is every column.

build_submodel expands a wildcard as at fit.

Missing values, following scikit-learn: transformers keep every row, fit on the observed values, and are missing where their input is. This fixes a bug where a single NaN made every value of a poly, ns or bs feature NaN (and ns could return uninitialized values). The docs show how to choose the rows R's na.omit keeps.

The conversion of dataframes of other backends moves to modelmatrix/_native.py, so response.py can use it.

Tests: 265 passed and 5 strict xfails. New tests cover missing values in poly, ns and bs; responses returned in polars and pyarrow; and exclude= given a response.

Docs: the responses page is rewritten for this design and executes locally. Sales ~ . still gives R's lm coefficients, and dropping incomplete rows of Hitters leaves R's 263.

🤖 Generated with Claude Code