Guardrail metric regressions

AirAsia tested a more aggressive seat-upsell flow. It lifted the target metric — but a guardrail is a metric you're not trying to improve and refuse to damage: page load time, refund rate, support contacts. A win on the target metric that wrecks a guardrail isn't a win.

Each row in experiment_metrics carries higher_is_better, which flips the direction of harm. From it, return every metric where treatment is worse than control by more than 2% in relative terms, with relative_change (treatment minus control, over control) rounded to 4 decimal places. Order by metric.