Fix ambiguous cross-file call inference for duplicate helper names

[alceops]

Fixes #543.

This keeps cross-file calls inference from guessing when an unqualified callee label resolves to multiple functions across the corpus. Today the global label map stores one node per normalized label, so common helper names like log(), execute(), or main() can resolve to whichever duplicate was seen last, creating false INFERRED calls and inflating graph metrics such as god_nodes.

Changes:

• track all global label candidates instead of overwriting duplicates
• resolve an unqualified cross-file raw call only when the label has exactly one candidate
• add a regression with two log() helpers proving run() -> log() is not guessed when ambiguous

Verification:

• python3 -m py_compile graphify/extract.py tests/test_extract.py
• fresh venv: pip install -e . pytest
• pytest tests/test_extract.py -q → 21 passed
• graphify update .