Module-scope calls (notebooks, scripts, __main__) emit no CALLS edges → find_dead_code false positives

[michael-denyer]

Bug

_extract_calls in code_review_graph/parser.py gates CALLS edge emission on enclosing_func being set:

if call_name and enclosing_func:
    caller = self._qualify(enclosing_func, file_path, enclosing_class)
    ...edges.append(EdgeInfo(kind="CALLS", source=caller, target=target, ...))

So calls made from module scope — top-level script glue, CLI entrypoints, if __name__ == "__main__" blocks, and Jupyter/Databricks notebook cells — produce zero CALLS edges. Any function invoked only from those contexts is then flagged as dead by find_dead_code (which counts incoming CALLS edges as evidence of liveness).

Notebook impact (severe)

PR #69 added notebook parsing — node extraction and IMPORTS_FROM edges work — but every cell is module-scope by definition, so notebooks emit no CALLS edges at all. This makes the dead-code detector's notebook coverage vacuous: any function called only from notebooks looks orphaned.

Reproducer

Real-world: a Databricks notebook (production inference pipeline) that calls Predict.extract_data_from_sample_ids():

# Direct parser call — bypasses any CLI/MCP layering
>>> from pathlib import Path
>>> from code_review_graph.parser import CodeParser
>>> nodes, edges = CodeParser().parse_file(Path("ML_wpredict_apply_v1.0.ipynb"))
>>> [e for e in edges if e.kind == "CALLS"]
[]
>>> [e for e in edges if e.kind == "IMPORTS_FROM"]
[<IMPORTS_FROM ... -> logging>, <IMPORTS_FROM ... -> sys>, <IMPORTS_FROM ... -> src.predict>]

The notebook contains predict_obj.extract_data_from_sample_ids(...) and similar calls in cells. Imports resolve correctly; calls are silently dropped.

refactor_tool(mode="dead_code") then flags extract_data_from_sample_ids and extract_data_from_files as dead — they're the entire reason the apply notebook exists.

Same shape reproduces with a plain .py file containing only a top-level helper() invocation, so this isn't notebook-specific — notebooks just suffer worst because they're 100% module-scope.

Scope of the fix

5 emission sites in parser.py gate on enclosing_func:

Line	Site	Languages affected
1455	Elixir call path	Elixir
2379	Generic _extract_calls	Python, JS, TS, others — the main fix
2415	JSX component invocation	TSX/JSX (a top-level <App /> render is module-scope)
2700	Solidity emit statement	Solidity
4002	R call path	R

Plus a downstream consideration: detect_entry_points in flows.py treats "is a CALLS target" as "is not a root" — so attributing a script's module-scope calls to the script's own File node would make script-only callees look "called by the script" and hide them from flow analysis. The fix needs to filter File-sourced CALLS in entry-point detection.

Why the existing convention supports the fix

_extract_value_references already attributes references to file_path when enclosing_func is None (parser.py line 2508-ish). CONTAINS edges do the same when there's no enclosing function. The fix just brings CALLS into line with the existing pattern.

Not addressed by prior PRs

Confirmed by reviewing dead-code-related history:

• fix: reduce dead code false positives #104 (decorator entry points, dunder exclusion) — different bug; doesn't add CALLS edges from module scope
• fix: filter method call noise from call graph #108 (filter external method calls to self/this) — narrows what counts as a call; doesn't restore module-scope source
• Fix Claude hook config and add JSX component CALLS edges #154 (synthesize JSX CALLS) — adds JSX edges but gates on enclosing_func (the same bug)
• feat: parser refactoring, Jedi call resolution, and performance optimizations #158 / feat: Jedi-based Python call resolution #247 (Jedi resolver) — improves call target resolution; doesn't change emission gating
• Dead code tool FPs #160 (Record<string, fn> dispatch maps) — value-reference patterns; addressed by _extract_value_references
• feat: community detection, flow analysis, and dead code improvements #249 (framework base classes, CDK) — entry-point heuristics

Fix submitted

PR pending — links 5 module-scope CALLS sites + filters File-sourced CALLS in detect_entry_points. End-to-end verification: edge count on ML_wpredict_apply_v1.0.ipynb goes from 0 to 14 CALLS edges; find_dead_code no longer flags the notebook-only methods.

318 tests pass (parser, refactor, flows, multilang, notebook).