Part 2. Build Your Own Real-Time Translator - LLM Streaming for 500ms

The core implementation: dual prompts for speed vs quality, streaming JSON extraction, debounce logic, and progressive frontend display — with full code.