When “Good Enough” Isn’t: The Danger of Sampling in Observability