This seems useful, but I wonder about edge cases:

- dynamically generated code - reflection-heavy systems - implicit coupling (shared schemas, conventions, etc)

These are usually where simple maps break down and why people fall back to full-text or embeddings.