Thought my WDN feature was clean ... until one assert fired due to a nan < 1e-6 comparison. Turns out a missing value in a long window propagated all the way to the causality check. Now I precheck for empty diffs and drop NaNs before running any leak detection. If there’s nothing to compare, raise early. Don’t fake a pass. Lesson: if the test passes due to NaN, the test is broken. [[ML]] [[Serendipity]]