Preventing Flash of Incomplete Markdown when streaming AI responses
Cutting 95th percentile latency from 3 minutes to 500ms