I did a quick search (here and elsewhere) and couldn't find anything on this subject. If all introns in a given primary transcript were spliced out in the same way, then this wouldn't matter. But lots of eukaryotic transcripts have alternate splicing patterns so that one version of the mRNA will contain coding regions that are absent in other alternate forms.
The problem relates to maintaining an open reading frame, avoiding a frame shift and the attendant (and unpredictable) consequences. If a splice donor site occurs within the 3-base codon, it might require that the two (or more) alternate acceptor sites would be in the same relative position in a related codon.
The simplest solution to this problem in my view is to make sure that all splice donor and acceptor sites would be at a codon/word boundary.
Is there any evidence that this is true?