Running the Postgres suites surfaced two tests asserting opposite things
about the same gesture. One was written before groups became all-or-nothing
and expected the list to survive a save that omits it, which is the behaviour
that made unset a silent no-op. It is replaced with the property that is
actually correct: resubmitting the same list alongside another flag leaves
the list and its cutover clock alone.
Nothing in the implementation changed here. Only a test that encoded the old
bug did.
Daniel Sutton
28bf05fe25