llama.cpp issue #27178: NVFP4-to-GGUF conversion drops most NVFP4 scales; model loads (or fails) then emits one repeated token

Root-caused to build_lora_mm in src/llama-graph.cpp applying the per-tensor scale only when the caller passes it, plus _generate_nvfp4_tensors in conversion/base.py removing quantized tensors before modify_tensors can synthesize QK-norm.