nx_llm_run_v2_test.nx

nx_llm_run_v2_test.nx -- end-to-end smoke for nx_llm_run_v2.nx. Builds the FULL Llama-shape GGUF (12 tensors) for n_layers=1 with all-zero F32 weights, builds a tiny BPE vocab, and runs nx_llm_generate_one_v2. With all-zero weights the forward yields uniform logits and the sampler returns a token in [0, vocab_size). This is the integration proof: every loader + composer brick shipped this session walks end-to-end from text prompt to emitted token ID via pure NishiLang.