python: implement TreeVisitor.adapt() so generic visitors can visit Py/J nodes

[jkschneider]

Summary

TreeVisitor.adapt() was stubbed in rewrite-python (# FIXME implement the visitor adapting) — it just returned self. That's load-bearing for correctness because each language's LST root routes its accept(v, p) through the adapter:

class J:
    def accept(self, v, p):
        return self.accept_java(v.adapt(J, JavaVisitor), p)

class Py(J):
    def accept(self, v, p):
        return self.accept_python(v.adapt(Py, PythonVisitor), p)

When v is a bare TreeVisitor (or a direct subclass that overrides only pre_visit/post_visit — a common pattern for cross-language collectors), the stubbed adapt flowed v straight into accept_java/accept_python, which immediately calls language-specific dispatch:

class CompilationUnit:
    def accept_python(self, v, p):
        return v.visit_compilation_unit(self, p)

The bare TreeVisitor doesn't implement visit_compilation_unit → AttributeError, wrapped in RecipeRunException by the visitor framework's outer try/except and propagated up. Reproducible failure:

class _Collector(TreeVisitor):
    def pre_visit(self, tree, p):
        return tree

_Collector().visit(some_py_compilation_unit, None)
# RecipeRunException: AttributeError: '_Collector' object has no attribute 'visit_compilation_unit'

• This blocks any correct implementation of _collect_search_result_ids (and similar cross-language tree walks) from extending the bare TreeVisitor. python: replace _Collector visitor with iterative tree walker in _collect_search_result_ids #7589 worked around the symptom by replacing the visitor-based collector with a hand-rolled iterative walker; closing that PR was the right call — this is the right layer to fix it.

Implementation

• A class-level registry on TreeVisitor maps each language visitor base class to a small adapter class. Each language module (rewrite.java.visitor, rewrite.python.visitor) registers its adapter at import time via TreeVisitor.register_adapter.

• adapt(tree_type, visitor_type) returns:

  • self if the visitor is already an instance of visitor_type
  • otherwise an adapter instance that is-a visitor_type (so the language-specific visit_* defaults in JavaVisitor / PythonVisitor are available for child traversal) and forwards pre_visit / post_visit / default_value / is_acceptable plus _cursor / _visit_count / _after_visit to the wrapped visitor — so user-defined logic on the original generic visitor still runs against the right cursor and sees every traversed node.

• When adapt is called on an existing adapter (e.g. cross-language traversal: a Py node visited via a JavaVisitor adapter chain), the wrapped visitor is unwrapped and re-wrapped in the correct adapter rather than nesting adapters.

is_adaptable_to is updated to match the new semantics: a visitor is adaptable to a target type if it's already an instance, or if it's an existing adapter, or if a registered adapter exists for the target type.

Test plan

• New tests/test_visitor_adapt.py covers:
  • The regression — a bare TreeVisitor subclass visiting a Py.CompilationUnit no longer raises
  • pre_visit actually fires on the visited node
  • End-to-end: _collect_search_result_ids (which uses a TreeVisitor subclass) finds a SearchResult marker on a Py root
  • adapt is a no-op when the visitor already is-a target type
  • adapt unwraps an existing adapter rather than stacking adapters
• Manual smoke test against the production install path (~/.moderne/cli/python/openrewrite/8.81.6/) — same minimal repro that crashed before this fix now succeeds and pre_visit is invoked.