When we look at latency issues during loading, we see that one of the culprits is that we won't apply any sort of limiting during inference of Core.Compiler.return_type, and that makes the answer more likely to be wrong by the time it gets to runtime. This is also for a code path (empty arrays) that really shouldn't matter at runtime, so that is very unfortunate. I think preferably here we would limit how much inference is done, as normal, by respecting the max_methods parameter, and return no information from this tfunc (so that it falls back to runtime lookup, in the rare case this is needed at runtime). For example, SnoopCompile reported:

inserting (::Type{T})(x::Static.StaticInteger) where T<:Real @ Static ~/.julia/packages/Static/tQ45Q/src/Static.jl:415 invalidated:
   mt_backedges: 1: signature Tuple{Type{Int64}, Any} triggered MethodInstance for Base.Broadcast.combine_eltypes(::Type{Int64}, ::Tuple{Base.Broadcast.Broadcasted{Base.Broadcast.DefaultArrayStyle{1}, Nothing, typeof(+), Tuple{Vector{Any}, Vector{Int64}}}}