feature: Hy3 support

[ErikTromp]

Summary

I had the Hy3 support for Colibri already working for quite a while but now there is more focus on supporting different models from the main repo as well, figure a pull request would make sense. I made sure to rewrite anything off that I did before generalization was in place in Colibri to work in a way that consolidates with the current codebase

Validation

• make -C c check
• CUDA changes were tested with make -C c cuda-test
• Performance claims include hardware, commands, and repeatable measurements

Compatibility

• The default CPU build remains dependency-free
• No model files, generated binaries, or benchmark artifacts are included

[JustVugg]

Reviewing properly takes longer than this note, but one thing is worth telling you now rather than after a user reports it.

hy3.c emits the READY sentinel but never calls coli_serve_binary_mode().

On Windows the CRT opens stdout in text mode and rewrites the trailing \n as \r\n, so the engine sends \x01\x01READY\x01\x01\r\n, coli — which matches the bytes exactly — never recognises it, and the session hangs forever with no error at all. That is #748: a user loaded 93 Kimi layers over 42 minutes and then sat watching nothing happen.

It is one line, from compat.h, before the sentinel:

static void serve_loop(...){
    coli_serve_binary_mode();   /* #748: TEXT-mode stdout mangles the READY sentinel */
    ...
    fputs("\x01\x01READY\x01\x01\n", stdout);

colibri.c has had it since #195; inkling.c and kimi_k3.c were written without it and it went unnoticed for months, which is exactly why it now lives in compat.h instead of being pasted per engine. Yours is the fourth engine to be written without it, so this is on our onboarding rather than on you — but you get to have it for free instead of finding out from a bug report.

Credit where it is due: your coli wiring is the most complete of any new engine we have received. model_arch recognises hy_v3, the banner is arch-aware, and the serve protocol is there. #712 (Qwen3.6) has none of that and is blocked on it, so you have already done the part that usually gets missed.

The larger question — whether we take a fifth engine right now — is the maintainer’s and is being discussed separately. Nothing about that reflects on the work; four engines already multiply every maintenance problem across four files, and this week produced two Windows bugs that had to be fixed in three of them at once.

[ErikTromp]

Good catch on the ready emit, will fix it