That is exactly my experience with cython. It’s a huge pain in the ass to debug. I found pybind11 to be way less of a pain in the ass than cython.

In most cases figuring out how to do it with nutty numpy is the least time consuming and is fast enough.

If you are going to rewrite, you may want to consider Julia. It will give you close to C speed, but will be as high level (or higher level) than Python.

numba is pretty great, and i saw pretty extreme speedups (i can't remember exactly now, but they were orders of magnitude faster) when i used it in a project. however, to make full use of it, you pretty much have to isolate the numerical code that you want it to compile and be careful to ensure that it is easily translatable. (that is, avoid using libraries or python objects). essentially, for full effect, you need to turn on "nopython" mode which then severely limits what you can do in the accelerated code.

ultimately, i'd probably look at julia or maybe rust if i was building something from scratch these days though. i'll also note that adding numba to a project is a pretty heavyweight proposition (you are, afterall, adding llvm to your project).