Given a database SOA serial X, and a SOA-EDIT that changes the SOA to Y, where Y>X, if an IXFR comes in with a serial N (X < N < Y), we send an empty reply (SOA Y+optional RRSIG), trying to indicate to the slave that it is up to date. However, because Y > N, the client interprets this as 'incremental updates coming' and thus is confused about the connection getting closed.
Thanks @sfrost for debugging this.
I suspect this was introduced with #1199
honor SOA-EDIT while considering "empty IXFR" fallback, fixes #1835
Reopening to remind us we really need a test for this.