nx_dual_companion_compose_test.nx

nx_dual_companion_compose_test.nx -- two-actor concurrent inference proof. The parallel-companion north star concretized: NARRATOR + NPC actors running interleaved cooperative steps, each generating 3 tokens autoregressively, output ordered through one multiplexer. What this proves end-to-end: 1. Two independent LLM-actor contexts can share the same scheduler + message bus + multiplexer without state corruption. 2. Cooperative pick_next interleaves them by priority + fairness. 3. nx_lr_actor_reset_for_next_token lets each actor roll into the next token using its own previous output (autoregressive loop). 4. Multiplexer correctly orders concurrent tokens by stream class priority (NARRATOR LLM_TEXT > NPC LLM_TEXT same class, FIFO tie-break + arrival-time tracking). 5. The other-agent's embedding + matmul + sample primitives are invoked 12 times (6 tokens x 2 actors x 3 primitive calls each) without leak / overflow / state cross-contamination. This is the smallest substrate-honest demonstration of the [[feedback-parallel-companion-multimodal-dnd-real-time]] cardinal.