Yes, that's what I would expect. There is only one "backend_xmin" in "pg_stat_replication", which corresponds to the snapshot held by the oldest query in any database on the standby server.
Thanks for the pointer to pg_stat_replication.
In my situation there aren't any queries against the standby server, both queries are running against the primary, but backend_xmin does contain the open transaction's txid_current() value. Does the standby somehow report the txid of the open transaction on the primary back to the primary as if it were running on the standby?
Is it an unavoidable limitation of the standby feedback mechanism that xmin is not tracked by database? It was certainly a surprise to me to find that hot_standby_feedback can trigger cross-database dependencies like this.