FETCH seems to expect deserialized form of LHS disk record as its input, rather than serialzied form.

This is rather inefficient, especially when dealing with record translation. It also means that fields such as virtual fileposition give random results/read off the end of the input buffer, since deserialize has not filled them in (not sure it knows how).

If fetch's transform expected the serialized form as input instead (as I believe diskread does) then these problems would be sidestepped I think.