Implementing TurboQuant in llama.cpp: CUDA Scars and What Actually Ships

Part 1 of 2. What it took to land TurboQuant quantisation in llama.cpp's CUDA path, what broke, and which of it survived contact with a real workload.