If oplog application of a multi document transaction is slow enough, then the PeriodicThreadToAbortExpiredTransactions which is running in the background may kill a transaction that expires.

I haven't reproduced this but it seems possible because secondary oplog application calls TransactionParticipant::Participant::beginOrContinueTransactionUnconditionally() which calls _beginMultiDocumentTransaction() (this happens when MongoDOperationContextSessionWithoutRefresh is constructed here, which calls beginOrContinueTransactionUnconditionally()) so it seems like during oplog application an ordinary multi doc transaction is getting created.