My only feature request (I'll open an issue) would be for test errors to retain the specific testset that failed instead of being flattened into the toplevel reconstructed testset. I don't know if that'll require a breaking change but since I'm not sure how to approach the implementation I don't think this should delay releasing v1.
Originally posted by @christiangnrd in #31 (comment)
More details here JuliaGPU/Metal.jl#438