fix: Fetch attribute annotations from inherited members too when pars…

…ing docstrings

I'm aware that computing inherited members will cache the MRO, but docstrings are usually never parsed *before* a package has finished loading. We do have to emphasize this in our docs though: extensions must not parse docstrings and modify parsed sections, and should rather modify the docstring value directly (even though it can be less convenient).

We should also stop caching parsed docstrings and MRO anyway: turns out the cache invalidation problem is really problematic.

Issue-mkdocstrings/python#175: mkdocstrings/python#175

src/_griffe/docstrings/google.py

@@ -286,8 +286,9 @@ def _read_attributes_section(
             annotation = parse_docstring_annotation(annotation, docstring)
         else:
             name = name_with_type
-            with suppress(AttributeError, KeyError):
-                annotation = docstring.parent.members[name].annotation  # type: ignore[union-attr]
+            with suppress(AttributeError, KeyError, TypeError):
+                # Use subscript syntax to fetch annotation from inherited members too.
+                annotation = docstring.parent[name].annotation  # type: ignore[index]
 
         attributes.append(DocstringAttribute(name=name, annotation=annotation, description=description))
 

src/_griffe/docstrings/numpy.py

@@ -558,8 +558,9 @@ def _read_attributes_section(
             name = name_type
             annotation = None
         if annotation is None:
-            with suppress(AttributeError, KeyError):
-                annotation = docstring.parent.members[name].annotation  # type: ignore[union-attr]
+            with suppress(AttributeError, KeyError, TypeError):
+                # Use subscript syntax to fetch annotation from inherited members too.
+                annotation = docstring.parent[name].annotation  # type: ignore[index]
         else:
             annotation = parse_docstring_annotation(annotation, docstring, log_level=LogLevel.debug)
         text = dedent("\n".join(item[1:]))

src/_griffe/docstrings/sphinx.py

@@ -264,8 +264,9 @@ def _read_attribute(
         annotation = parsed_attribute_type
     else:
         # try to use the annotation from the parent
-        with suppress(AttributeError, KeyError):
-            annotation = docstring.parent.attributes[name].annotation  # type: ignore[union-attr]
+        with suppress(AttributeError, KeyError, TypeError):
+            # Use subscript syntax to fetch annotation from inherited members too.
+            annotation = docstring.parent[name].annotation  # type: ignore[index]
     if name in parsed_values.attributes:
         docstring_warning(docstring, 0, f"Duplicate attribute entry for '{name}'")
     else: