@smarter I'm not sure if this is a good idea. I only kept it and submitted it because it fixes and issue and it avoids work. And I think it can be justified under "I said hard union so don't widen" expectations.

Hmm, so this kind of goes in the opposite direction of the just merged 5bf3227, it seems that what's important in the code sample is that the user explicitly wrote B.type and so it's the enum type itself which is "hard" here and shouldn't be widened, the fact that a hard union is also present is incidental I think. I suggested an alternative fix in #19823 (comment)