Stop using JSON for LLM structured output

For simple extraction tasks, a delimiter-separated string uses 11 tokens vs 35 for JSON. Output tokens are the latency bottleneck.