The best thing might be to disable the numba JIT during coverage measurement. That relies on you trusting the correspondence between the Python code and the JIT’ed code, but you need to trust that to some extent anyway.
Not that this answers the question, but I thought I should advertise another way that someone might be interested in working on. There’s probably something really beautiful that could be done using
llvm-cov. Presumably, this would have to be implemented within numba, and the llvm code would have to be instrumented, which would require some flag somewhere. But since numba knows about the correspondence between lines of python code and llvm code, there must be something that could be implemented by somebody more clever than I am.