Strip base-model identity from README and training_args.json (kept only in adapter_config.json where PEFT requires it)

Initial anonymous release: Qwen3.5-9B judge LoRA (subq+human, ckpt-294)