fix: overhaul follow-up detection for conversation memory + add diagnostic logs

fix: switch HF endpoint to universal router URL (auto-selects provider)

fix: add stream error handling + startup diagnostics, bump CACHEBUST

feat: overhaul model routing — GLM-5.1 primary, strict 4-model fallback chain, purge all stale refs

fix: update deprecated Hugging Face api-inference URL to router.huggingface.co

fix: update Qwen model identifier to Qwen3.5-9B due to deprecated support

feat: Add full multi-turn conversation memory and context rewriting

fix(backend): bust docker cache + fix async streaming generator to not block event loop

fix(api): improve stream token chunks robustness and sync year filter

feat: implement ChatGPT style UI, streaming generation, math/code blocks, citations & multi-model fallback

fix: f-string syntax error in reranker block for python 3.10 compatibility

feat: ResearchPilot v2 — Next.js frontend, 358k embeddings pipeline, LaTeX math rendering, dual-GPU Kaggle scaling

feat: FastAPI backend complete

feat: RAG pipeline complete - ResearchPilot end-to-end working

feat: retrieval optimization pipeline complete

feat: vector database indexing complete

feat: document chunking pipeline complete

feat: data ingestion and processing pipeline complete