feat: BYO-LLM pipeline — gap detection + Gemini cleaner + their LLM answers

fix: robust token tracking, retry, Jina fallback, no more 500s

fix: 503 retry + token tracking + cost calc per response

upgrade: gemini-3.1-flash-lite-preview + jina X-Engine=direct timeout=30s

perf: gemma-3-27b-it → gemini-2.5-flash-lite (3-5x faster, 1000 RPD free)

fix: jina web search with Bearer auth header

fix: full fresh acra.py with Jina web search, no duckduckgo

web search: switch to jina s.jina.ai — no rate limit issues

fix: real web search via duckduckgo-search, use_web skips local entirely

fix: use_web=True bypasses L0 early return so web search always runs

fix: use_web=True skips local entirely — pure web results

fix: use_web=True filters local docs <0.8 similarity to prevent noise

perf: batch compress — N Gemma calls → 1 call, L2 17s→5s L3 43s→12s

fix: L0 always retrieves then lets Gemma decide context vs knowledge

fix: L0 checks docs first (similarity>0.75) before model fallback

fix: acra.py — switch to google-genai + gemini-embedding-001

deploy: acra.py