I'm actually struggling to come up with an example that specifically fails despite this, and isn't related to the properties of the elements (or missing UniformScaling methods). Perhaps I'll add some tests that are broken due to missing methods? These would hopefully be fixed in the future.

I was thinking of something like

A = Diagonal([zeros(2,3) for _ in 1:3]) # or zeros(3,3)
B = Diagonal([zeros(2,2) for _ in 1:3])
@testthrows SomeError A .+ B

I've added a test with incompatible element sizes, and also one that is directly related to this PR's changes:

julia> D = Diagonal([ones(1,1), ones(1,1)])
2×2 Diagonal{Matrix{Float64}, Vector{Matrix{Float64}}}:
 [1.0;;]     ⋅   
    ⋅     [1.0;;]

julia> size.(D)
ERROR: MethodError: no method matching size(::UniformScaling{Float64})
The function `size` exists, but no method is defined for this combination of argument types.
[...]

Thankfully, this isn't a regression, as the operation didn't work before either. In fact, the following works now:

julia> ndims.(D)
2×2 Matrix{Int64}:
 2  2
 2  2