@dkazakov are you sure replacing a fully-default-ctor/dtorable, zero-copy struct with N*M virtual calls is the way to go? The rest looks solid, but please take the opportunity and fix that `reverese` typo :D