Production AI & LLM Pipelines: Guardrails, Streaming Resilience, and Cost-Aware Fallbacks