Model Presets: per-turn expert-on-demand model escalation

## Problem Right now, Hermes runs one model for the entire session. If you're on a cheap/fast model