It's another jump, perhaps, but smaller arrays, and now we can track
ephemera efficiently without bloating child trees. (We could also
put ephemera inline with the child trees but then nth_token would be
unwieldy, and it would lower our data density.)