Branch nonrel_wal_test_xid_wraparound contains a test, that illustrates the problem with xid wraparound.
It fails on main too (with different symptoms), but I suggest concentrating our efforts on nonrel_wal.

In a current shape, the test is too hacky to propose as PR. I'll try to straighten it up.
If you have any ideas on how to make it better, I'm all ears.

See also a discussion in hackers about similar testing scenario:
https://www.postgresql.org/message-id/flat/CAP4vRV5gEHFLB7NwOE6_dyHAeVfkvqF8Z_g5GaCQZNgBAE0Frw%40mail.gmail.com#e10861372aec22119b66756ecbac581c