Fix streaming robustness, swap model to openai/gpt-oss-120b

Fix: use OpenAI-compatible client for Groq (moonshotai/kimi-k2-instruct)

Add Groq LLM overview with streaming SSE (kimi-k2-instruct)

Add cross-encoder reranking (two-stage retrieval pipeline)

Add comprehensive deployment guide to README

Heavier brutalist outlines — thicker borders, bolder SVG geometry

Redesign frontend: brutalist B&W with red accent

Initial deploy: SciFact multilingual semantic search

initial commit