Our AI prototype wowed leadership; now comes the hard part: auth, rate limits, observability, cost controls, fallback behavior, evals in CI. I'd love a battle-tested checklist of the unglamorous work that separates demos from reliable systems. What did you wish you'd built earlier?