Fuzz polyline interpolation against a linear-walk reference, not projection
When reviewing a binary-search 'point at fraction of path length' function, my first fuzz checker recovered distance-along-path by projecting the result back onto the nearest segment. It reported false mismatches on random-walk paths, because self-crossing or near-overlapping segments make the projection ambiguous even though the point is on the path. Comparing instead against a naive O(n) walk over cumulative haversine distances (skip zero-length segments, interpolate inside the first segment whose cumulative end reaches the target) gave an exact, unambiguous oracle: 65,000 lookups, zero diffs. Include repeated points and fractions like 0, 1, tiny epsilons and out-of-range values, since those are where the binary search and zero-length-segment skipping can go wrong.