fix(ipw): preserve treatment design_info instead of overwriting with covariates (#842)

Fixes #835.

`InversePropensityWeighting._build_design_matrices` assigned
`self._t_design_info` twice in a row — first to the treatment matrix's
`design_info`, then immediately overwriting it with the covariate
matrix's `design_info`. The attribute name suggests treatment metadata
but only the covariate side was being kept.

Store the covariate side under `self._x_design_info` to match the
existing convention used in `instrumental_variable.py`,
`interrupted_time_series.py`, `regression_discontinuity.py`, and
`prepostnegd.py`. The treatment design now lives under
`self._t_design_info` as the name implies.

Adds a `TestDesignInfoAttributes` regression suite that asserts the
two attributes are distinct and hold the correct sides of the
formula. The class would have failed on master because
`_t_design_info` would have held the covariate columns and
`_x_design_info` would not have existed at all.

Assisted-by: Claude Code

Co-authored-by: Benjamin T. Vincent <inferencelab@gmail.com>

Author: haoyu-haoyu

causalpy/experiments/inverse_propensity_weighting.py

@@ -106,7 +106,7 @@ def _build_design_matrices(self) -> None:
         """
         t, X = dmatrices(self.formula, self.data)
         self._t_design_info = t.design_info
-        self._t_design_info = X.design_info
+        self._x_design_info = X.design_info
         self.labels = X.design_info.column_names
         self.t, self.X = np.asarray(t), np.asarray(X)
         self.y = self.data[self.outcome_variable]

causalpy/tests/test_ipw_get_ate.py

@@ -220,3 +220,24 @@ def test_get_ate_with_unknown_method_falls_back_to_doubly_robust(self, ipw_resul
         assert np.isclose(ate_list[0], ate, equal_nan=True)
         assert np.isclose(ate_list[1], trt, equal_nan=True)
         assert np.isclose(ate_list[2], ntrt, equal_nan=True)
+
+
+class TestDesignInfoAttributes:
+    """Regression for #835: ``_t_design_info`` must hold the treatment design
+    (left-hand side of the formula) and ``_x_design_info`` must hold the
+    covariate design (right-hand side). The previous implementation assigned
+    both to ``_t_design_info``, silently overwriting the treatment side."""
+
+    def test_t_design_info_is_treatment_side(self, ipw_result):
+        """The treatment side of ``trt ~ 1 + age + race`` is the lone column ``trt``."""
+        assert ipw_result._t_design_info.column_names == ["trt"]
+
+    def test_x_design_info_is_covariate_side(self, ipw_result):
+        """The covariate side mirrors the right-hand side of the formula."""
+        assert hasattr(ipw_result, "_x_design_info")
+        assert "age" in ipw_result._x_design_info.column_names
+        assert "race" in ipw_result._x_design_info.column_names
+
+    def test_t_and_x_design_info_are_distinct(self, ipw_result):
+        """The two attributes must point to different design objects."""
+        assert ipw_result._t_design_info is not ipw_result._x_design_info