I think that this is the topic, dfsort-icetool-icegener/topic9351.html
I think there is a little confusion with an example showing x'0000nnnn' for the RDW, which should be x'nnnn0000'.
The question itself is a little obtuse. What happens when you SORT first on the RDW? Then the records are primarily in RDW order. For non-spanned records, that will be definitely record-length order. For spanned records, I'm not sufficiently interested to check (set up a test/read the manual) whether there is anything in the IOCS by the time SORT happens which would modify the behaviour.
If the previously-spanned-multiple-records are presnted to the SORT statement as one record (expected) I'd expect x'0000' in the low-order bytes of the RDW. If possible to read the constituent records individually, I'd expect the low-order two bytes to be non-zero.
I've only ever SORTed on the record-length to get data into record-length order for analysis of a file. If I wanted data also sorted withen equal record-lengths and the key happened to start at 5, I'd not consider clumping it all together, as it would hide my specific abnormal wishes. Normally I'd have no problem with clumping consecutive keys (although these days no real problem, I believe, in leaving them separate, it's an old habbit). It is just that the RDW is not "data" so I'm not going to confuse things by pretending it is "data".