refactor: system/user prompt split + robust JSON extraction

feat: 3-step agentic pipeline — A/B comparison, red-flag triage, SOAP note

feat: AMD Live Inference Dashboard — show latency, throughput, token count after each analysis

feat: multilingual support (ZH/ES/FR/JA) + remove all mock mode

fix: add detailed connection logs to diagnose AMD Cloud failures

feat: add AMD Cloud connection status bar, replace MOCK badge

refactor: switch to AMD Dev Cloud vLLM backend with Qwen2.5-VL

Production update: full Gradio app with AMD branding, mock+real mode, bilingual support