Replace an LLM call with a typed decision model, CPU-only. 77.5–84.1% top-1 at 4.0–60.8 ms p50, single-thread CPU.