There was never a bug here. The doc is quite explicit
http://www.lmdb.tech/doc/group__mdb.html#ga73a5938ae4c3239ee11efa07eb22b882
After txn_abort the only valid call on a still-open cursor is
cursor_renew, not cursor_close.
However you want to call it, are you open to improve this workflow?
API-wise, it would be so much simpler just to free a cursor after it became
obsolete. It's counter-intuitive to open a new txn, renew the cursor, close
the cursor, and abort the txn - just to close a cursor.
And without MDB_VL32, plain cursor closing after txn abort seems to work
just fine (I did not consider "closing" as "using" when I read the
docs).
One could say that the inconsistent behavior is a result of the complexity
MDB_VL32 introduces.
If you don't want to change the current behavior, it would still be nice for
future developers to null the obsolete txn ptr in the cursor on txn abort to
make the unwanted close call crash always, not just occasionally.