reduce(vcat and and reduce(hcat, are not consistent in the return type for a single element collection, except if that collection is an array

[oxinabox]

The reduce(vcat and hcat are a bit funny about the types they return when you reduce over a single element

2-element Array{Int64,1}:
 1
 2

julia> reduce(vcat, (1:2,))
1:2

julia> reduce(vcat, [1:2,0:-1])
2-element Array{Int64,1}:
 1
 2

julia> reduce(vcat, (1:2,0:-1))
2-element Array{Int64,1}:
 1
 2

another example

julia> reduce(hcat, ([1,2] for x in 1:1))
2-element Array{Int64,1}:
 1
 2

julia> reduce(hcat, [[1,2]])
2×1 Array{Int64,2}:
 1
 2

julia> reduce(hcat, ([1,2] for x in 1:2))
2×2 Array{Int64,2}:
 1  1
 2  2

So this issue is two fold

1. Reducing over 1 element vs multiple elements can return a different type
2. The above is not true for if the collection is a Vector. Return type is always consistent for vectors even if they only have 1 element.

---

I have this spare enhancement to the tests for these functions,
that I was writing for other reasons (I was thinking about improving their performance like #21672)

@testset "optimized reduce(vcat/hcat, A) for arrays" begin
    function check(op, args)
        Y = op(args...)

        varients = unique((
                args,
                collect(args),
                Tuple(args),
                (arg for arg in args),  # HasLength
                (arg for arg in args if true)  # SizeUnknown
            )) 

        @testset "reduce(::typeof($op), ::$(typeof(xs))" for xs in varients
            X = reduce(op, xs)
            @test X == Y
            xs isa Vector && @test typeof(X) == typeof(Y)
        end
    end

    for args in ([1:2], [[1, 2]], [1:2, 3:4], [[3, 4, 5], 1:2], [1:2, [3.5, 4.5]],
                 [[1 2], [3 4; 5 6]], [reshape([1, 2], 2, 1), 3:4])
        check(vcat,  args)
    end
    for args in ([1:2], [[1, 2]], [1:2, 3:4], [[3, 4, 5], 1:3], [1:2, [3.5, 4.5]],
                 [[1 2; 3 4], [5 6; 7 8]], [1:2, [5 6; 7 8]], [[5 6; 7 8], [1, 2]])
        check(hcat,  args)
    end
end

[simonbyrne]

This should be fixable by defining appropriate reduce_first methods.