I feel like if you’re having enough incidents where you can accurately calculate a P95 for recovery then that isn’t the metric you’re going to be caring about. Instead you’re going to be focused on change failure rate.
Also for failures I would imagine they aren’t heavily skewed to small amounts.