Bump. I was going to make the same PR. @JeffBezanson Any objections?

Though I'm not sure the test needs to be so complex. A much shorter reproducer is:

julia> mapreduce(cos, +, Union{Int,Missing}[1], dims=1)
ERROR: InexactError: Int64(Int64, 0.5403023058681398)

EDIT: actually, more changes are needed to fix the above test case. I guess the only really correct solution will be to use the type of z and widen progressively if needed, as we do for map. Cf. #10533 (comment).

EDIT2: maybe as a temporary fix we could do typeof(z) <: T ? T : typeof(z)? That's really what matters here.

maybe as a temporary fix we could do typeof(z) <: T ? T : typeof(z)? That's really what matters here.

That does sound a bit better to me --- using T can't possibly work unless z isa T.

I've implemented that in the first commit of #28027.

Fixed by #28089, but would be good to add the test case from here.

Force-pushed this to a commit with just the test-case if there's still interest in including it.