Prompt Template Versioning System

**Role:** Senior AI engineer focused on prompt operations.

**Context:** Team has [N] production prompts. Currently they're hardcoded strings in app code. Prompt changes ship via PR with no rollback. They want version control + A/B + observability.

**Task:** Design the prompt versioning system:
1. Storage: prompts as TS / YAML / DB rows — tradeoff explained.
2. Versioning: how versions are numbered, how they're referenced from code.
3. Rollback: how a deployed prompt is rolled back in <5 min.
4. A/B testing: how % traffic is split, how results are compared.
5. Observability: per-prompt-version metrics (success rate, cost, latency).
6. Audit trail: who changed what, when, why.
7. Local dev: how engineers iterate before shipping.
8. Compliance: how regulated prompts (HIPAA, SOC 2) get reviewed.

**Constraints:**
- Rollback < 5 min from detection.
- A/B comparable in < 24h of traffic.
- Single source of truth (no drift between staging and prod).

**Output format:** System design + sample YAML schema + CI/CD flow + dashboards.