When run after master server crash --tc-heuristic-recover=rollback produces inconsistent server state with binlog still containing transactions that were rolled back by the option.
Such way recovered server may not be used for replication. E.g when such way recovered
ex-master is demoted into slave its binlog state needs further correction to subtract
the rolled back transactions from its binlog status. Otherwise the "new" slave might claim
those transactions as locally present at the (gtid) master-slave connection protocol. At the same time the actual "new" master may never have seen those transactions (because they never arrived at it when it was formerly slave, due to the crash).
This issue should be fixed with refining the recovery procedure with truncating binlog to cut off the prepared rolled back transactions. The method is also known as pioneered by FB https://www.percona.com/community-blog/2018/08/23/question-about-semi-synchronous-replication-answer-with-all-the-details/.