feat(eval): RAGAS evaluation framework + RAG pipeline improvements

fix: improve RAG retrieval quality and reduce generator hallucination

fix: use LiteLLMClient adapter for Gemini judge in build_judge

feat: add Gemini native judge backend to eval pipeline

fix: remove dead results list and duplicate load_dotenv in run_eval.py

feat: split run_eval.py into checkpointed phases with no_reasoning support

fix: EVAL_PRIMARY_MODEL routes graph through osmapi via LiteLLM openai-compat provider

feat: EVAL_PRIMARY_MODEL override for graph LLM, applied before civicsetu import

feat: new run_eval.py — single-pass no-phases, 3 batch_score calls total via osmapi

refactor: simplify phase 2 to sequential row-by-row scoring, comment out batch/sleep/thread code

chore: comment out inter-metric rate-limit sleep (not needed for osmapi)

feat: switch to Ollama embeddings, comment out 60s rate-limit sleeps for osmapi

feat: switch eval judge to osmapi / qwen3.5-122b-a10b, keep Gemini embeddings

fix: remove dual-key parallel mode — always use single worker with GEMINI_API_KEY_2

fix: share single genai.Client across workers to prevent global auth state conflict

fix: log batch start in score_batch_in_thread for visibility into thread pickup

fix: add 120s per-call timeout to Gemini/OpenRouter client + 5-min per-batch deadline

fix: rate-limit-aware scoring — sleep between metric calls + BATCH_SIZE=1 default

feat: add OpenRouter judge provider to sidestep Gemini 15 RPM free-tier limit

fix: raise max_tokens to 8192 for RAGAS judge LLM

fix: RAGAS 0.4.x native API + phase separation (eval-collect / eval-score)

fix: switch RAGAS judge to LiteLLM + instructor.from_litellm

fix: use new google-genai SDK in build_judge()

fix: RAGAS 0.4.x native API + phase separation (eval-collect / eval-score)

fix: pass llm/embeddings to RAGAS 0.4.x metric constructors

eval: update RAGAS metric imports to 0.4.x API

eval: fix runner script quality issues (RAGAS column_map, NaN handling, key validation)

eval: add RAGAS benchmark runner script