I find I am frequently removing sections of inscrutable “vectorized” NumPy code that is inefficient, not because NumPy is inefficient but because conforming the data to a structure amenable to NumPy is too inefficient in the first place, and rewriting in straightforward, simple Cython functions.

Also the next maintainer will thank you for it.

To be clear, this is not meant as a critique on C, on the contrary. Put simply, it is just that making computers do stuff is hard. This is especially true if you want it to do it in a very particular way such as in high-performance code. If you cannot already leverage the work of other experts, you need to know about modern optimizing compilers, language runtimes, modern processor architectures, OS/system, ...)

To continue your Numpy example, that framework entombs the know-how and man-years of work of specialist in a certain area. If you need to do simple dense algebra, you will a hard time beating Numpy/Julia/Eigen/MKL etc. If your problem goes slightly off-track (e.g. sparse problems) where you cannot easily leverage the hard work and expertise of others, you better start rolling up your sleeves and open that main.c