[bilbo-b-baggins]

I would say since your focus is on structural or programmatic detection, and not LLM heuristics, the problem depends on language a lot.

In Rust or Go there’s super clear test markers or filenames.

In Javascript it would have to detect the framework in use then detect test files and tests embedded in program files.

And so forth.

Are you doing any call sequencing heuristics? Like if the same 5 calls (with different args) appear in the same order in multiple places (even in test files) that might be a strong signal for deduplication. Or even if the same 5 calls are in the same order with a couple different interleaved calls - the fuzziness of the heuristic might be something tunable to a language, or particular codebase, or framework, etc.