This fix introduced an innocuous though technically breaking change in
the definition of `erc-input' (the struct). Basically, it added a new
slot, `refoldp', to allow users access to something resembling the
pre-5.6 behavior, where protocol-oriented message splitting would take
place after `erc-pre-send-functions' ran. That is, setting the slot to t
is meant to buy you some of that old functionality in the form of a
second split. (The new behavior of only splitting beforehand favors
interactive client users over bot/module authors.)

However, reflecting back on this, I think it wouldn't kill us to account
for the unlikely possibility of someone "subclassing" `erc-input' for
use outside this hook. In most cases, I believe simply recompiling their
dependent libraries would solve the issue, but why chance it if we don't
have to? Hence the attached change, which removes the slot but "spoofs"
its would-be accessor, `erc-input-refoldp', while the hook runs. I
originally went with a name that differed from that of the would-be
accessor, but this feature has been on HEAD for a while, so keeping it
seemed the less disruptive option.