Code-review fixes: guard empty SSE choices, unstick UI on truncated stream

Fix Z.ai model id, surface LLM errors as SSE frames, cache-friendly Docker layers

decompose() falls back to the default hosted model when qwen isn't enabled

Model switcher, public hardening, and CI (steps 1-3+5 of deploy plan)

Apply ponytail audit: delete dead endpoints, scripts, params, and assets

Fix "I don't know" on derivation questions (gate + over-strict refusal)

Add query decomposition to fix comparison/multi-hop retrieval misses

Add ELI5 "explain simply" button + scrape jargon from all answer prompts

Stream answers over SSE (/ask/stream, /why/stream) + stop button

Add PDF page citations (side-tab viewer) + "Why?" deeper-dive

Retrieval + generation reliability for specific computed quantities

V2: local web UI (FastAPI + React/Tailwind) + 8GB GPU fixes

Stage 7: retrieval eval loop + data-driven threshold

Full-corpus ingest (Grinstead & Snell, 518pp) + retrieval/generation fixes

Stage 6: grounded local generation (DeepSeek-R1-Distill-7B via Ollama)