After applying the changes from #27636:

>>> print(repr(np.strings.replace(
            np.array(["hello planet"], dtype="U"),
            np.array("planet", dtype="T"),
            np.array("world", dtype="T"))))
array(['hello world'], dtype='<U11')

We get a unicode array back, despite the fact that StringDType explicitly installs a promoter for this case.

I thought this had something to do with the fact that the legacy string unicode promoters work by constructing a promoter DType tuple that's filled with None entries. But that's fine, since None has less priority than a specific DType.

Instead something else is going on. I got as far as tracing through the call to resolve_descriptors, which does correctly set the output DType to StringDType, so something else fishy is going on later when the actual output array is created.