Ilhan Polat

Thank you, I'll rewrite this in C just to see if it will fly faster. Because I can't still see what is slowing down the code with those changes.

The main difference between numpy and scipy are how things are prepared for the underlying LAPACK routine and which routine is used. NumPy makes very little preparation as it is...

Ouch, it seems like I converted `det` to Cython in version 1.11.0 (#18225) then wrote the C code but never pushed it. Sorry for the noise. That work needs to...

In any case free-threading or not, it won't make any difference because we do not write threaded code so it won't benefit from it. Any threading would be done in...

Banded detection is never working out since there is no definition of when it is enough; is it 5% or 10% or etc. That's why that's typically left out and...

Hessenberg, tridiag, bidiag solvers are different than banded ones and require their own solvers some exist in LAPACK some need manual coding so you don't need to consider them here....

Took a while but now ready for review. Nonsymmetric tests locally pass. Symmetric solver has a minor issue that I will follow-up when I have a bit more time.

I have them, currently the symmetric part is broken so I didn't send those parts. There is really no diff on the low-level code as it is old out, new...

> I'll attempt to test this with igraph the week after next, which should help with gaining confidence in correctness. It would be much appreciated. The more tests, the merrier,...

> Are you planning to keep this API-compatible with ARPACK? It does not seem to be compatible at the moment. Yes, the API has changed, or broken, depending on how...