Adds the .safetensors checkpoint and restyles the README top to match the efficient-cube3d card.
- multi_part_dit_int4_hqq_g128.safetensors (1,478,462,752 bytes): same weights as the .pt, stored as 1163 plain tensors, no pickle. Verified in a fresh process: vertices and faces bit for bit on all 3 parts of the idx 130 bar stool.
- README: Cube3D-style header (badges, intro, emoji table, highlighted note); HQQ (Half-Quadratic Quantization) defined at first mention; .safetensors load snippet and Files row. All numbers unchanged.
- quant_config.json: safetensors file name and load recipe.
Latency badge says INT4 is slower (66.5 s vs 34.7 s, eager mode), since that is what we measured.
