I have a unit test that shows there's a race condition (first observed by David Lloyd) in com.arjuna.ats.internal.jta.transaction.arjunacore.jca.TransactionImporterImple#importTransaction(javax.transaction.xa.Xid, int).
If two threads call this method at the same time, two separate transaction objects may be created. Here's the sequence of events:
T1: call importTransaction for XID1
T2: call importTransaction for XID1
T1: getImportedTransaction returns null
T2: getImportedTransaction returns null
T1: create new transaction, add to map
T2: create new transaction, add to map (overwriting T1's)
There is nothing in the documentation to indicate that this is not a valid situation or that access to the TransactionImporter has to be single-threaded in any way.