Built a persistent stdio server so the model loads once per session: warm calls 0.3-1.3s (was ~3.7s reloading a 3.5GB model every call), runs on a 16GB MacBook Air M5, parallel tool calls serialize…