Manages frank-large/small inference routing, FLM NPU constraints, and Halogen runtime kwargs for thinking block preservation.