nx_zimage_gguf_probe.nx

nx_zimage_gguf_probe.nx -- read the REAL Z-Image text-encoder GGUF structure, sovereignly. sd-server -> Nishi migration, the real-weight bridge: proves our sovereign GGUF reader (`nx_gguf_parse` + `nx_gguf_find_tensor` + `nx_gguf_tensor_at`, the same path `nx_real_gguf_test` gates on a 491MB Qwen) reads the ACTUAL Z-Image model weights -- `Z-Image_Qwen_3_4b-Q6_K.gguf` -- so real tensors can flow into our gated f32 organs (-> dequant `nx_q6_k_to_f32` -> the DiT/attention path). Bounded PREFIX read: the GGUF header + tensor-info + KV-metadata live at the FRONT of the file; tensor DATA (the multi-GB bulk) is lazy and not needed to enumerate/locate tensors, so we read only a 192MB prefix instead of the 2.5GB file. Pass: file opens, >=1KB read, GGUF v3, tensor_count sane, and token_embd.weight is found with dim_0>0. license_tier: ORIGINAL