I get the same execution plans as the article author. This leads me to believe that it is serializing then filtering.
I also built octet accessor properties into my UDT, these get even cleaner plans when queried through.
Note: I in no way extensively tested this. If someone wanted to I would gladly give them my source. It could definitely be optimized as I built it quickly, simply to get a feel for how to buil a UDT.