Oh yeah, that's probably an error, I'll correct it. The way it works for <cutoff> is that it'll assume it's a cutoff of the next immediately following word, so yeah should be "<cutoff-longword> longword word"

Definitely a bug, there is logic to decrease the probability of spn for cutoffs (but keep it as possibility vs subwords), but either it's not working correctly or if it is treating it as an <unk> token explicitly, then the cutoff model isn't being applied. I'll try to take a look at it this weekend!