I'm noticing that this test is quite fragile, it creates noisy linear data where the standard deviation of the residual is 50, but it expects a fairly tight match.

It's seeded, so it'll continue to pass provided that .NET doesn't change the behaviour of System.Random, which they do fairly often, and provided that we don't change our implementation. But I think it's probably a bad test. I noticed this when I changed the implementation to one which (accidentally) mimicked the old behaviour, but with extra calls to rand.NextDouble.