fix: emit CALLS edges for module-scope code (closes #284)

[michael-denyer]

Summary

Closes #284.

_extract_calls and 4 sibling helpers gated CALLS edge emission on enclosing_func being set, so module-scope calls (top-level script glue, CLI entrypoints, if __name__ == '__main__' blocks, Jupyter/Databricks notebook cells) produced zero CALLS edges. Any function invoked only from those contexts was flagged as dead by find_dead_code.

Notebooks were hit hardest: PR #69 added node + IMPORTS_FROM extraction, but every cell is module-scope by definition, so notebooks emitted no CALLS edges at all — making the dead-code detector's notebook coverage vacuous.

What changed

Parser (5 emission sites): when enclosing_func is None, attribute the CALLS edge to the File node instead of dropping it. Matches the existing convention used by _extract_value_references and CONTAINS edges.

Site	Language(s)
_extract_calls (the main path)	Python, JS, TS, generic
Elixir call path	Elixir
JSX component invocation	TSX/JSX
Solidity emit	Solidity
R call path	R

Downstream fix in detect_entry_points: without filtering, a script's module-scope calls would attribute to the script's own File node, making script-only callees look "called by the script" and hiding them from flow analysis. Added get_all_call_targets(include_file_sources=False) so detect_entry_points excludes File-sourced CALLS. Implementation joins against nodes.kind = 'File' rather than pattern-matching source_qualified so future changes to File-node naming can't silently miscategorize edges.

End-to-end verification

Real-world repro: a Databricks notebook (production inference pipeline) calling Predict.extract_data_from_sample_ids().

Before:

>>> CodeParser().parse_file(Path('ML_wpredict_apply_v1.0.ipynb'))
nodes: 1, edges: 3 (all IMPORTS_FROM, zero CALLS)
>>> find_dead_code(...) → flags extract_data_from_sample_ids, extract_data_from_files

After:

nodes: 1, edges: 17 (3 IMPORTS_FROM, 14 CALLS)
>>> find_dead_code(...) → no longer flags either method

Tests

5 new tests, all passing:

• test_parser.py::test_module_scope_calls_attributed_to_file — bare .py script
• test_parser.py::test_module_scope_calls_in_notebook — .ipynb file
• test_flows.py::test_detect_entry_points_module_scope_caller_is_still_root — flow analysis treats File-sourced CALLS correctly
• test_refactor.py::test_module_scope_caller_prevents_dead_code_flag — end-to-end parse → store → find_dead_code
• test_refactor.py::test_if_main_block_caller_prevents_dead_code_flag — same for __main__ block

Full impacted suite: 318 passed, 0 failures (parser, refactor, flows, multilang, notebook).

$ uv run pytest tests/test_parser.py tests/test_refactor.py tests/test_notebook.py tests/test_multilang.py tests/test_flows.py
================== 316 passed, 2 xpassed, 1 warning ===================

Test plan

• Parser emits CALLS edges from module-scope code (Python .py)
• Parser emits CALLS edges from notebook cells (.ipynb)
• detect_entry_points excludes File-sourced CALLS so script-only callees remain roots
• find_dead_code does not flag module-scope-called functions
• find_dead_code does not flag __main__-block-called functions
• No regressions in existing parser/refactor/flows/multilang/notebook tests

Not addressed (scope kept tight)

Reviewed all dead-code-related PRs/issues (#104, #108, #154, #158, #160, #247, #249) — none address module-scope CALLS emission. The other 4 helper sites (Elixir, JSX, Solidity, R) had the same gating shape so are fixed in the same PR for consistency, even though the original repro only needed the Python path.