fix: The ordering of terms differentiated formulae was potentially in…

…correct after v1.1.0. (#236)

formulaic/formula.py

@@ -322,20 +322,17 @@ def differentiate(
         use_sympy: bool = False,
     ) -> _SelfType:
         """
-        EXPERIMENTAL: Take the gradient of this formula. When used a linear
-        regression, evaluating a trained model on model matrices generated by
-        this formula is equivalent to estimating the gradient of that fitted
-        form with respect to `wrt`.
+        Take the gradient of this formula with respect to the variables in
+        `wrt`.
+
+        When used a linear regression context, making predictions based on the
+        model matrices generated the differentiated formula is equivalent to
+        estimating the gradient of the fitted model with respect to `wrt`.
 
         Args:
             wrt: The variables with respect to which the gradient should be
                 taken.
             use_sympy: Whether to use sympy to perform symbolic differentiation.
-
-
-        Notes:
-            This method is provisional and may be removed in any future major
-            version.
         """
 
 
@@ -482,27 +479,25 @@ def differentiate(  # pylint: disable=redefined-builtin
         use_sympy: bool = False,
     ) -> SimpleFormula:
         """
-        EXPERIMENTAL: Take the gradient of this formula. When used a linear
-        regression, evaluating a trained model on model matrices generated by
-        this formula is equivalent to estimating the gradient of that fitted
-        form with respect to `wrt`.
+        Take the gradient of this formula with respect to the variables in
+        `wrt`.
+
+        When used a linear regression context, making predictions based on the
+        model matrices generated the differentiated formula is equivalent to
+        estimating the gradient of the fitted model with respect to `wrt`.
 
         Args:
             wrt: The variables with respect to which the gradient should be
                 taken.
             use_sympy: Whether to use sympy to perform symbolic differentiation.
-
-
-        Notes:
-            This method is provisional and may be removed in any future major
-            version.
         """
         return SimpleFormula(
             [
                 differentiate_term(term, wrt, use_sympy=use_sympy)
                 for term in self.__terms
             ],
-            _ordering=self.ordering,
+            # Preserve term ordering even if differentiation modifies degrees/etc.
+            _ordering=OrderingMethod.NONE,
         )
 
     def get_model_matrix(
@@ -784,20 +779,17 @@ def differentiate(  # pylint: disable=redefined-builtin
         use_sympy: bool = False,
     ) -> SimpleFormula:
         """
-        EXPERIMENTAL: Take the gradient of this formula. When used a linear
-        regression, evaluating a trained model on model matrices generated by
-        this formula is equivalent to estimating the gradient of that fitted
-        form with respect to `wrt`.
+        Take the gradient of this formula with respect to the variables in
+        `wrt`.
+
+        When used a linear regression context, making predictions based on the
+        model matrices generated the differentiated formula is equivalent to
+        estimating the gradient of the fitted model with respect to `wrt`.
 
         Args:
             wrt: The variables with respect to which the gradient should be
                 taken.
             use_sympy: Whether to use sympy to perform symbolic differentiation.
-
-
-        Notes:
-            This method is provisional and may be removed in any future major
-            version.
         """
         return cast(
             SimpleFormula,

tests/test_formula.py

@@ -197,6 +197,9 @@ def test_differentiate(self):
         assert f.differentiate("a") == ["1", "0", "0"]
         assert f.differentiate("c") == ["0", "0", "0"]
 
+        g = Formula("a:b + b:c + c:d - 1")
+        assert g.differentiate("b") == ["a", "c", "0"]  # order preserved
+
     def test_differentiate_with_sympy(self):
         pytest.importorskip("sympy")
         f = Formula("a + b + log(c) - 1")
@@ -208,6 +211,9 @@ def test_differentiate_with_sympy(self):
             "rhs": ["0", "(1/x)"],
         }
 
+        h = Formula("a + {a**2} + b - 1").differentiate("a", use_sympy=True)
+        assert h == ["1", "(2*a)", "0"]  # order preserved
+
     def test_repr(self, formula_expr, formula_exprs):
         assert repr(formula_expr) == "1 + a + b + c + a:b + a:c + b:c + a:b:c"
         assert repr(formula_exprs) == ".lhs:\n    a\n.rhs:\n    1 + b"