When a transaction fails during replication, it should signal the error to any
following transactions that are executing in parallel. However, the logic was
incorrect, so this only worked if the error occured during the COMMIT/XID
event at the end of the group. On other errors, following transactions would
not be signalled correctly and would not roll back, resulting in corrupt state
of replication.

This bug is also seen as one of the problems in MDEV-5863.

Many thanks to nanyi607rao for reporting this problem on maria-developer@ and
analysing the cause.