LLM_MODE

Also written: fake mode, live mode, replay mode

An environment variable that chooses where model calls go: `fake` (a free stand-in model), `live` (the real APIs, with replies saved in `recordings/`) or `replay` (saved real replies played back).