What broke when we ran an AI coding agent on local Ollama models (including a 4GB GPU)