Can't speak for DSI, but p52 from the manual says:
If you save a program that uses chord memory, the chord is saved with the program.
If you copy a non-split or layered program into a layer / split, and the chord has more than four notes saved, how would the unit consistently allocate the voices on an eight-voice unit (4 voices available per split / layer)?
The four-note chord thing seems pretty sensible: regardless of voice count (4- / 8- / 16-voice), the same notes will be played with no voice robbing, no configuration to poll ("Am I an 8- or 16-voice unit? Split or layered mode?") during voice assignment, and with portability of the chord with the program itself.
I see your point: it'd be nice to have larger mega-chords that would make use of the unit's polyphony (e.g., the PPG Wave's interval assignment in unison mode), but at the cost of increased complexity for each of four modes (4+4, 8, 8+8, 16), would it be worth it?