Intra-class method calls resolve to the class node: jedi inference at ast.Call col_offset hits the receiver, not the callee

[rahlk]

PROBLEM

For an intra-class method call (self._method(...)), the emitted call graph points the edge at the CLASS node instead of the callee method — and the callee method may not appear as a graph node at all. The symbol table is correct (the method exists as a PyCallable with the right signature); only the call-site/edge resolution is wrong. This silently truncates call chains at their deepest hop (the codellm-devkit/python-sdk#65 symptom: web_auth_reset_password -> ResUsers.reset_password -> _action_reset_password loses the last hop).

MINIMAL REPRO (verified against installed 0.3.0)

# pkg/mod.py
class ResUsers:
 def reset_password(self, login):
 return self._action_reset_password([login])
 def _action_reset_password(self, ids):
 return ids
def entry():
 return ResUsers().reset_password("x")

Raw Codeanalyzer.analyze() output (no SDK involved):

source='pkg.mod.entry' target='pkg.mod.ResUsers.__init__' type='CALL_DEP' weight=2 provenance=['jedi']
source='pkg.mod.ResUsers.reset_password' target='pkg.mod.ResUsers' type='CALL_DEP' weight=1 provenance=['jedi']

Expected second edge target: pkg.mod.ResUsers._action_reset_password.

Raw call-site record for the intra-class call — callee_signature is the class:

{"method_name":"_action_reset_password","receiver_expr":"self","receiver_type":"ResUsers",
 "callee_signature":"pkg.mod.ResUsers","is_constructor_call":false,"start_line":3}

ROOT CAUSE (localized in 0.3.0)

codeanalyzer/syntactic_analysis/symbol_table_builder.py, _call_sites (~lines 590-594): _infer_callee(script, node.lineno, node.col_offset) is called with the ast.Call node's position. For attribute calls (self.method(...)), node.col_offset is the column of the RECEIVER (self), not the method name — so Jedi infers the receiver's type, whose full_name collapses to the class's qualified name. _infer_callee's constructor-rewrite guard (d.type == "class", ~line 92) does not fire because an instance's d.type is "instance", so the class name passes through unrewritten. Bare-name calls (ResUsers()) work only because the call's col_offset coincides with the callee expression there.

Likely fix direction: for ast.Attribute callees, infer at the attribute NAME's position (node.func.value.end_col_offset + 1 / node.func.attr position), or resolve via the attribute node explicitly.

DOWNSTREAM / PROPAGATION

• The SDK (python-sdk) is a verbatim pass-through (verified: _build_call_graph and get_callsites_for forward analyzer output unmodified) — no SDK-side fix applies.
• python-sdk 1.x is pinned to 0.3.0: fixing this for 1.x users needs a 0.3.1 backport, or ships naturally with 0.4.0 (schema v2) if the v2 emission path inherits the same _call_sites inference — PLEASE VERIFY the v2 branch against this repro before cutting 0.4.0; the L1/L2 jedi path likely shares the code, and the superset gate would then bake the wrong edge into v2 goldens.
• Suggested regression gates: (1) edge reset_password -> ResUsers._action_reset_password exists; (2) NO callable -> class edge for a non-constructor call; (3) callsite callee_signature for self.x(...) ends with .x.

Reported from a python-sdk audit session; full repro scripts available on request.

[rahlk]

Verified fixed in v0.3.1 against the original repro: the intra-class edge now targets pkg.mod.ResUsers._action_reset_password (method, not class), callee_signature matches, and the previously-missing entry -> ResUsers.reset_password edge now exists as well. All three regression gates from the issue pass. SDK uptake: codellm-devkit/python-sdk pin bump to 0.3.1 in flight.