fix(python): extract a decorated class nested in a class body

[DeusData]

What

A Python class body can declare classes behind a decorator:

class BoxStyle:
    @_register_style(_style_list)
    class Circle:
        def __call__(self, x0, y0, width, height, mutation_size): ...

push_nested_class_nodes (extract_defs.c) descended into a class body's direct children only, so a
nested class wrapped in a decorated_definition was never extracted. The class, its methods and every edge
they own were missing (matplotlib's BoxStyle / ArrowStyle styles, wagtail's Panel.BoundPanel family,
textual's nested Message subclasses, airflow's nested deadline models). For Python the walk now also
descends into decorated_definition, just as a top-level decorated class is handled. Plain nested
classes are unchanged.

Builds on #2686 (#1548 nested defs, merged). No index-format change.

Evidence

Base #2686 @090580b3 vs this change @4d71327c, 26 Python repos (manifest v1), production builds, macOS
arm64, --repeat 2, every repeat byte-identical (noise 0 on every type). The branch has since merged
main (1346ac4, which includes #2686 and #2679); its diff against main is the two files of this change,
and PR CI tests the merged tree.

type	base	this PR	+added	-removed
DEFINES	2,515,174	2,515,377	203	0
DEFINES_METHOD	397,865	397,995	130	0
DECORATES	243,846	243,951	105	0
OVERRIDE	82,497	82,576	79	0
INHERITS	78,598	78,660	78	16
MENTIONS	42,891	42,902	11	0
CALLS	1,354,959	1,355,032	234	161
USAGE	1,672,257	1,672,554	512	215
WRITES	536,179	536,222	147	104
TESTS	511,518	511,529	64	53
IMPORTS	477,476	477,476	30	30

18 of the 26 repos are identical apart from SEMANTICALLY_RELATED noise. Sampled and read (271 edges,
seed 20261017):

• New definitions and their own edges: the decorated nested classes and their methods now exist, with
  correct DECORATES, OVERRIDE and INHERITS (to _Base, _Curve, Message) and correct calls inside them
  (e.g. BoxStyle.Ellipse.__call__ -> Path.unit_circle).
• Re-attributed (same target, same line, new source): 33 CALLS, 55 USAGE, 21 WRITES. Calls and accesses
  inside a decorated nested class used to hang on the enclosing class or file. They now start at the
  nested class's own methods.
• Retargeted (same source, new target): most of the remaining removals. With the new classes in the
  registry, edges that are resolved by name alone pick a different candidate:
  • Better: a name used inside its own class now finds that class's nested member. Examples: textual's
    Input._watch_value -> Input.Changed instead of Checkbox.Changed, and Tab.Clicked inherits
    Tab.TabMessage instead of Tabs.TabMessage. wagtail's MultiFieldPanel.BoundPanel inherits
    PanelGroup.BoundPanel, and Panel.get_bound_panel -> Panel.BoundPanel. airflow's test calls through
    SerializedReferenceModels.AverageRuntimeDeadline(...) reach that class instead of the task-sdk
    namesake, and textual's TextArea message tests go to TextArea.Changed.
  • Neutral: wagtail attribute names (classname, icon, attrs, classes: 88 USAGE, 81 WRITES) resolved
    by name only move from one wrong candidate to another (Table.Row.classname or the icon template
    tag become Panel.BoundPanel.classname / .icon).
  • Worse: self-calls inside a nested class are not walked by the Python LSP, so they stay low-confidence
    suffix_match guesses. With more candidates, the guess can now pick a subclass override. Example:
    airflow ReferenceModels.BaseDeadlineReference.evaluate_with's self._evaluate_with used to reach
    the class's own method and now reaches AverageRuntimeDeadline._evaluate_with. A follow-up walks
    nested classes in the Python LSP under their outer class, which binds these self-calls.

Exposed approximation (documented, not changed here)

matplotlib's patches.py defines both the top-level Circle / Ellipse / Wedge and the decorated
style classes BoxStyle.Circle, BoxStyle.Ellipse and ArrowStyle.Wedge. With matplotlib's lib/ layout,
from matplotlib.patches import Circle misses the module-path strategy, and the symbol-name fallback
picks the lexicographically smallest QN with that name. That is now patches.BoxStyle.Circle
(< patches.Circle). Result: 30 IMPORTS are retargeted to the nested style class, 74 CALLS go to
BoxStyle.Circle (40), BoxStyle.Ellipse (23) and ArrowStyle.Wedge (11), and 30 lsp_constructor plus
53 unique_name edges are lost because Circle is no longer unique. Two things cause this: the
fallback can bind an import to a class member, and lib/ packages are not mapped. The new nodes only make
it visible. The follow-up fixes the fallback: an imported name is a module-level binding, never a class
member, and lib/ packages get a package map.

Tests

• pipeline_python_decorated_nested_class_extracted{,_parallel}: a decorated and a plain class nested
  in one class body. Both classes and their methods are extracted, on the sequential and the parallel
  pipeline. RED without the fix (ASSERT(dec_class)).
• Affected suites (complexity, extraction, registry, pipeline, cross_repo, py_lsp,
  lsp_resolution_probe, incremental) pass on the merged tree.