So, it looks like this tries to build a model (or rather, multiple layers of models combined with something called context mixing[0]), potentially analyzing the whole file first? Is it fair to call this a very fancy entropy coder, replacing the last phase of what a traditional audio codec does (the lossless part, after quantization)?

I like the code style, this is pretty nice C. It's refreshing to see human code.

[0]: https://en.wikipedia.org/wiki/Context_mixing