I don’t think that you’re getting a less precise answer with 1.5.0. I think instead you get equally imprecise answers in 1.4.2 and 1.5.0 (like in my example, both xsvd and xqr are equally imprecise). But if you use the 1.4.2 answer as the standard, then it appears that 1.5.0 answer is in error.
Neither answer is better than the other. They’re equally imprecise. The error shows up in different ways depending on how you do the algorithm, maybe something to do with changes in promotion rules. But the change does not mean worsening. Just different.
This is expected behavior from a theory-of-numerics perspective. Maybe there’s more going on in your example than I know about. But what you’re seeing is expected. And the fix is to make your unit tests check that the computation aligns with theory (error bounded as discussion above), and not to expect higher accuracy than is possible.