Stop Trusting Vibes: Build a Minimal LLM Eval Harness Before You Claim Your Model Got Better