This is an example that Csmith has to sacrifice expressiveness in order to maintain unambiguity: How to allocate bit-fields is very much implementation dependent, and we found some differences among compilers.
Yes, it's a shame. In the case of Frama-C's value analysis for instance, the remaining unionizable structs can still find the same bugs that could occur with 8-bit, 16-bit and 32-bit bitfields. But there could perfectly be bugs in the value analysis that can only be found with structs sporting, for instance, 1-bit or 31-bit bitfields inside unions.
This said, bitfields aren't used much in our field of interest for the same reason that makes them annoying for comparative testing, so perhaps I should just learn to accept that we may have bugs there but they don't matter so much.