Add escalation summary generation and product search rejection

Add production features: circuit breaker, exponential backoff, semantic caching, intent-based scope handling

Fix critical bugs: compensation for late orders, escalation detection, conversation history improvements

Fix sequential LLM timeout: Add 20s per-call timeout, 90s executor timeout, split error handling

Fix timeout bugs: Dedicated warmup script, get_running_loop, thread-safe LLM cache, proper message format

Fix 120s timeout: Add model pre-warming, LLM warmup, and async executor for graph.invoke

Fix: All 7 critical bugs + 4 security enhancements + compensation policy

Initial commit: Olist intelligence layer with database