fix(python): resolve module-qualified calls from inside class methods

[rijefff]

Summary

In _resolve_python_member_calls() (graphify/extract.py), qualified calls of the form
module.func() resolve only when the caller is a module-level function. When the same
call appears inside a class method, no calls edge is produced.

The module-alias arm bails because caller_file is None.

Reproduction

# pkg/util.py
def helper(x):
    return x + 1

# app.py
from pkg import util
from pkg.util import helper


class Machine:
    def method_qualified(self):
        return util.helper(1)      # <-- NO calls edge (bug)

    def method_bare(self):
        return helper(1)           # <-- ok


def function_qualified():
    return util.helper(1)          # <-- ok

caller	shape	calls edge on main
function_qualified()	module-level + module.attr	yes
Machine.method_bare()	class method + bare name	yes
Machine.method_qualified()	class method + module.attr	no

Root cause

file_of_node is populated only from contains edges:

caller_file = file_of_node.get(caller)
mods = [t for t in imported_by_filenode.get(caller_file, ()) ...]
if len(mods) != 1:
    continue

Methods are linked to their class through the method relation, not contains, so
methods get no entry in file_of_node. caller_file is therefore None,
imported_by_filenode.get(None, ()) is empty, and the arm continues.

Confirmed by instrumenting the resolver: caller_file=None for Machine.method_qualified(),
while function_qualified() resolves to its file.

Fix

1. Populate file_of_node from method edges (class -> method) as well.
2. Walk transitively method -> class -> file when resolving caller_file.

16 added lines, no existing line changed.

Test

tests/test_python_member_calls.py (new, follows the existing
test_<lang>_member_calls.py naming). It asserts the previously missing edge and the
two shapes that already worked, so the fix cannot silently regress them.

• Without this change: fails — assert _has_edge(result, method_qualified, helper, "calls") -> False
• With this change: passes

Full suite

pytest tests/ on this branch (base v8): 3651 passed, 14 failed, 170 skipped.

The 14 failures are pre-existing and environmental, not caused by this change — verified by a
control run: reverting only the patch and re-running the full suite produces the same 14
failures, and comm of the two sorted failure lists is empty. They are test_terraform.py
(7, tree_sitter_hcl not installed), test_ollama_retry_cap.py (4), and one each in
test_extract.py, test_install_references.py, test_manifest_ingest.py.

Impact

Measured on a private ~272-file Python codebase where most functions are methods on classes:

metric	before	after
total calls edges	5,436	8,191 (+51%)
cross-module edges between two core modules	0	17

On that repository's full graph the change adds +2,755 edges (+22.9%).

The failure is silent: the graph simply reports fewer callers, so an agent can conclude
"nothing else calls this" when callers do exist.

Related

Similar resolver gaps are open for other languages: #1313 (Go), #2135 (Rust), #1682 (PHP).
I did not find an existing issue for the Python class-method case.

Not covered here

from pkg import util as u2 inside a method (lazy import + alias) still produces no edge —
the alias is bound in function scope rather than file scope. Left out deliberately to keep this
change small; happy to file it separately.