Regression Suite Design

**Role:** AI engineer who has watched 5+ companies regress quality silently when model versions rotated. You build the suite that catches it.

**Context:** Team ships a product with [N] customer-facing LLM-powered features. They're nervous about a model upgrade [LIST: e.g., Claude 4.5 → 5].

**Task:** Design the regression suite:
1. Cover all [N] features with at least 20 test cases each.
2. For each test: input, expected behavior, observable signals.
3. Grader per test (string match / LLM-judge / human-required).
4. Suite execution: per PR (subset, 5 min), nightly (full, 1h), pre-release (full + red-team).
5. Drift detection: what signal triggers a "model has regressed" alert.
6. Sign-off rubric: what % pass-rate green-lights deploy.
7. Manual review backlog: which failures get human review vs auto-fail.
8. Historical comparison: how today's results are compared to last week's.

**Constraints:**
- LLM-judge graders must be calibrated to human ratings (κ ≥ 0.7).
- Every threshold has a justification.
- Include 3 known-failure cases that the suite MUST catch.

**Output format:** Test-suite spec + sample YAML test definitions + CI pipeline.