Handle UnionAll Broadcasted in broadcast override inference

[vchuravy]

Problem

Inference crashes with a FieldError when Enzyme's broadcast mapreduce/foldl override is consulted for a Broadcasted type that widenconst left only partially specified (a UnionAll with free Axes/F/Args parameters):

FieldError: type UnionAll has no field `parameters`, available fields: `var`, `body`
  [2] bc_or_array_or_number_ty(Ty::Type, midnothing::Bool)
      @ Enzyme.Compiler.Interpreter src/compiler/interpreter.jl:1077

bc_or_array_or_number_ty and has_array match Ty <: Broadcast.Broadcasted{...} and then directly read Ty.parameters[4] (the broadcast args tuple type). That access assumes Ty is a concrete DataType, but a UnionAll has no parameters field. This was hit in the wild by the SVD reverse rule in TensorKit, which produces such a partially-specified Broadcasted type.

Fix

Guard both helpers so they bail out (return false) when Ty is not a fully-specified DataType. Returning false is the conservative, correct choice: when the args tuple type is unknown, Enzyme simply skips the broadcast override and falls back to the normal AD path instead of crashing.

Testing

BC = Base.Broadcast.Broadcasted
DAS = Base.Broadcast.DefaultArrayStyle
T1 = BC{DAS{2}, A, F, Args} where {A, F, Args}   # UnionAll, previously crashed
bc_or_array_or_number_ty(T1, false)  # => false (was: FieldError)
has_array(T1, false)                 # => false (was: FieldError)
# concrete Broadcasted still returns true as before

🤖 Generated with Claude Code