nx_actor_role_llm_v2_real_test.nx

nx_actor_role_llm_v2_real_test.nx -- REAL end-to-end smoke driving the v2 LLM actor through actual transformer forward. Where nx_actor_role_llm_v2_test.nx is wiring-only, THIS smoke builds a real 12-tensor synthetic Llama GGUF (n_layers=1, hidden=2, head=2, vocab=4, ffn=4, all-zero F32 weights), a real BPE vocab, and drives the v2 actor adapter through INIT -> RUN -> EMIT. RUN actually invokes nx_llm_generate_one_v2 which executes: tokenize prompt -> nx_embedding_lookup -> RoPE + RMSNorm + multi-head attention + SwiGLU FFN + output norm + output proj + temperature + top_k + nx_sample_categorical -> token id Per the operator constraint: "real testing of real world capabilities bits up". No constructor-rejection-only paper proof; this exercises the entire other-agent runner stack through the actor system. Composes [[feedback-real-playtest-loop-not-just-logs]] + [[feedback-four-pillar-code-hygiene-untracked-imports]] (additive fix landed; preventative scanner shipped; this is the demonstration of the now-safe end-to-end path).