logging every model call is the cheapest debugging upgrade you can make

I log the prompt, the response, the tokens, and the latency for every call. when something goes weird in prod I actually have something to look at. do this early.