DLM serve demo

Ask a question against an in-context table. Three arms: confluence (position-code model), plain (same-data LoRA), base (untuned Llama-3.2-1B), llama8b (Llama-3.1-8B-Instruct).

Arm Table (first line = header; cells separated by commas or tabs) Question

API: POST /predict with JSON {"arm","header","rows","question"} — see /docs. First request after idle pays a cold start (~30-60 s).