Filing for future reference (would be too much of an interruption to fix now):

julia> a = rand(UInt8, 5)
5-element Array{UInt8,1}:
 0x08
 0xcd
 0x4a
 0x5d
 0x1f

julia> b = rand(UInt8, 5)
5-element Array{UInt8,1}:
 0x7c
 0xf4
 0x2c
 0x86
 0xef

julia> dot(a, b)
0x9b

The fix is less trivial than it seems: the r_promote infrastructure in reduce.jl seems promising, but it doesn't handle the product of elements, only the sum of the accumulator. We might need a widen_op analogous to promote_op to prevent these kinds of overflows.