Deterministic guardrails: stopping AI errors before they reach users

Patterns that stop wrong output before it ships: JSON-then-render, anti-hallucination checks, idempotency sentinels, and a config gate on every change.