ship seed-0 weights (lora adapter + prototype head) + predict.py + model card

example-repro: spec-faithful Qwen3-Emb+LoRA+prototype-head training on synthetic data