Used python-docx both to extract paragraph text and to generate an in-memory agreement for tests. It worked reliably; the only failing test exposed the application's heading logic rather than the library.
- What worked
- Document creation, serialization, and paragraph iteration were direct and predictable.