nx_actor_role_llm.nx

nx_actor_role_llm.nx -- LLM-inference actor role adapter. REAL integration with the other-agent Nishi AI stack: this adapter wraps a cooperative-scheduled LLM inference step sequence as an actor. Each call to nx_lr_actor_step() executes ONE phase of the micro-LLM forward pass by calling the other-agent primitives: Phase 0 (EMBED): nx_embedding_lookup (other agent) Phase 1 (PROJ): nx_blas_matmul (other agent) Phase 2 (SAMPLE): nx_logit_apply_temperature + nx_logit_top_k_mask + nx_sample_categorical (other agent) Phase 3 (EMIT): nx_ms_send_fanout (this agent) + mark complete This proves the parallel-companion substrate doesn't just COMPOSE structurally with the Nishi AI port -- it actually drives real inference primitives. V1 ships a tiny 1-token-from-prompt path using i64 Q10 storage; multi-layer + attention defer to future composes per [[feedback-honest-perf-verdict-no-aspirational-claims]] (the v2 LLM runner stack is what handles full Llama, not this). Composes [[feedback-parallel-companion-multimodal-dnd-real-time]] + [[project-conductor-arc-phases-c-d-integrate-with-nishi-ai-2026-05-19]] (the integration project).