Previously, it would just pass back a partially-uninitialized tupdesc,
which doesn't seem like a safe or useful behavior.

Backpatch to v10 where this code came in.

Discussion: https://postgr.es/m/30830.1544384975@sss.pgh.pa.us