feat(api): re-resolve corpus_version on demand when empty at startup

refactor(pipeline): pass corpus_version explicitly, stop mutating settings

fix(api): log exception type, not message, for infra errors

fix(api): return 503 with Retry-After on pool exhaustion

refactor(api): run /query as sync handler in the threadpool

fix(backend): el rate limit lee rate_limit_per_min/day de settings

refactor(rag): strategy pattern for LLM providers — eliminate provider if-chains

feat(llm): add multi-provider abstraction supporting Groq and OpenAI-compat

fix(hardening): address verify warnings W1 W2 W3

test(hardening): unit + integration tests for all subsystems

feat(hardening): pipeline integration -- cache + tracing + post-gen validation

feat(hardening): rate-limiting -- slowapi middleware + query endpoint limits

fix: allow startup without corpus data, return 503 on empty corpus

feat: wire fastapi app with lifespan, query router, and di dependencies

chore: validate python 3.14 wheel compatibility for psycopg2 and sentence-transformers