nx_actor_role_llm_v2.nx

nx_actor_role_llm_v2.nx -- REAL transformer-forward LLM actor. Where [[nx_actor_role_llm]] drives a tiny 3-step micro-pipeline (embed + matmul + sample) over synthetic toy weights, THIS adapter wraps the full Llama-class runner [[nx_llm_run_v2]] -- multi-layer transformer forward including RoPE + RMSNorm + multi-head attention + SwiGLU FFN + output projection + temperature + top_k + sample. Cooperative phases: Phase 0 (INIT) : reserved Phase 1 (RUN) : single call to nx_llm_generate_one_v2(spec, gguf, hdr, bpe, prompt, ..., prng, rope_base, attn_scale) -- this is the monolithic generate-one-token step Phase 2 (EMIT) : fanout LLM_TOKEN with the generated token if positive; else fanout ERROR_REPORT with verdict id Phase 3 (DONE) : terminal V1 of this adapter is single-shot (one token per actor lifetime). V2 will add a reset_for_next_token loop similar to nx_actor_role_llm. Composes [[project-conductor-arc-phases-c-d-integrate-with-nishi-ai -2026-05-19]] -- the LLM substrate the other agent built becomes callable from inside the parallel-companion actor system.