Would it be fair to call this a bug in nextafterf? If this FPU flag is enabled it would seem like returning a denormal here is incorrect as it isn’t representable in this cpu in this configuration.
None of the relevant standards say anything about how math functions are supposed to behave in the non-standard DAZ/FTZ modes (because they're non-standard). The second you set the bit, you step off the path of being able to complain about any behavior not documented by the library.