Instrument the Call, Then Read the Receipt: TTFT Tracing in Python

Most latency tooling answers the question "how long did that take?" and stops. For an LLM call in production, the useful questions start after that: which model actually served it, did a fallback answ