I would agree that its a lot more resistant than one would think, but even an innocuous Gaussian with a heavy tailed contamination confuses it.

For cases where one needs too test "is this bigger than that ?" I test for stochastic dominance rather than equality of mean.