The prototype is being written in Python (and some C for fast CUDA kernels that aren't actually crucial). It's just easier, transparent to read and interpreter takes care about boring stuff. It's for purpose. The target client is meant to be written in a language with low level access to memory (and some other spices). So, C for instance (it's not yet decided, I'd prefer it). Thus it's easier to prepare cleanly readable code first, test it, and implement into less readable and easier to abuse language. Besides, my style keeps no standards, so it's better if it's at least readable.
The choice of language with low level access to memory is due to things that always made me feel uneasy while analyzing other coins' clients. Would prefer having means to at least obfuscate crypto operations in memory. Especially since Spectre came into light (just wrote a small, sketchy art about it: https://medium.com/@Hibryda/on-melting-spectre-or-how-soft-hardware-can-lead-to-coitus-interruptus-with-your-coins-6f975e8b22b9 ). Call me paranoid, but it never failed me.