Add ComfyUI-native NAR LoRA (fused qkv/gate_up layout, vae2llm/llm2vae as diffs)

ckpt_io: also load AR-branch LoRA safetensors

Add safetensors releases (fp32 + bf16) with named LoRA tensors; scripts load either format

README wording

README: recommend the 0.666 VAE merge

v4 references

Delete files tokenizer_head_joint_v3.pt with huggingface_hub

rename weights to v4

sanitized tokenizer head + NAR LoRA + training/inference scripts

initial commit