Self-Hosting GLM 5.2 with Hybrid Quantization

@jcartu shared a MXFP8/NVFP4/NF3 hybrid quantization of GLM 5.2 that stays within 1% KLD of BF16 — better than z.ai's FP8 cloud serving. He clarified GLM is mainly useful for compliance-sensitive tasks, as Fable and GPT-…