AST: @property/@staticmethod methods get class-unqualified node ids but class-qualified rationale_for edge targets → docstring nodes orphaned at build #1050

Summary

The Python AST extractor generates inconsistent node ids for methods decorated with @property / @staticmethod: the method node gets a class-unqualified id, but the rationale_for edge that links the method's docstring/comment targets a class-qualified id. The two ids don't match, so build_from_json treats the edge as dangling and drops it (the src not in node_set or tgt not in node_set: continue branch). The result is orphaned (degree-0) rationale nodes — every docstring/comment on a decorated method becomes a disconnected node in the graph.

The same class-qualified-vs-not mismatch can also drop calls / references edges that point into decorated methods, so those methods are under-connected in general.

Minimal reproduction

repro.py:

class Bar:
    @property
    def baz(self) -> int:
        """Return the baz value."""
        return 1

    def normal(self) -> int:
        """A normal instance method."""
        return 3

from pathlib import Path
from graphify.extract import extract
from graphify.build import build_from_json

ast = extract([Path("repro.py")])

# method NODE ids:
#   baz (@property)      -> id "..._baz"          label "baz()"      (NO class segment, NO leading dot)
#   normal (instance)    -> id "..._bar_normal"   label ".normal()"  (class segment present)

# rationale_for EDGE targets:
#   baz docstring    -> target "..._bar_baz"     -> target_in_nodes = False  (class-qualified, does not match node)
#   normal docstring -> target "..._bar_normal"  -> target_in_nodes = True

G = build_from_json(ast)
# baz's docstring rationale node  -> degree 0  (ISOLATED)
# normal's docstring rationale node -> degree 1 (connected)

Observed output:

NODE ids:
  id=..._baz           label='baz()'
  id=..._bar_normal    label='.normal()'
rationale_for edges:
  target=..._bar_baz       target_in_nodes=False   <-- dropped at build
  target=..._bar_normal    target_in_nodes=True
after build:
  rationale ..._rationale_4   degree=0   'Return the baz value.'   <-- orphaned
  rationale ..._rationale_13  degree=1   'A normal instance method.'

Root cause

For a decorated method, the node-id derivation omits the enclosing class (id file_method, label method() with no leading dot), while the rationale_for edge-target derivation keeps the class (id file_class_method). The two code paths disagree, so the edge endpoint never resolves and the edge is silently discarded as "external/stdlib".

Larger-corpus evidence

On a real ~8k-LOC Python project (294 docstring/comment rationale nodes): 38 rationale nodes were isolated. For all 38, the source endpoint (the rationale node) was present and only the target (the parent method) was missing; all 38 targets resolved by removing exactly one interior id segment (the class), and every parent was verified to be a @property or @staticmethod.

Suggested fix

Make the decorated-method node id consistent with the edge-target id — i.e. include the enclosing class segment for @property / @staticmethod methods just as for regular instance methods (and ideally keep the leading-dot label convention too). Either derivation is fine as long as both sides use the same one.

Workaround (post-extraction)

Remap an edge endpoint that is missing from the node set by dropping one interior id segment when exactly one candidate exists:

def declassify_unique(eid, node_set):
    p = eid.split("_")
    hits = [h for i in range(1, len(p) - 1) if (h := "_".join(p[:i] + p[i+1:])) in node_set]
    return hits[0] if len(hits) == 1 else None

This reconnected 37/38 orphaned nodes in the corpus above (the 1 remainder had an ambiguous declassification and was left alone). Empty __init__.py files and pure HTML header boilerplate remain legitimately isolated — they have nothing to connect to.

Environment

graphify v8 (installed via uv pip install; graphify.__version__ is not exposed and importlib.metadata.version("graphify") raises PackageNotFoundError, so I can't give an exact build string). Python 3.12. The behavior is deterministic and reproduces from the snippet above.