Johannes Schindelin [off-list ref] writes:
In my hands, I need to tell rerere to forget, *and then recreate the merge
conflict* before I can resolve it again and let rerere learn the new
resolution.
I can believe that---that is how I originally desiged "forget" to
behave.
Even though My quick tests to prepare a reply to Martin seemed to
show that unmerging was not strictly needed in today's Git in order
to forget and learn new, I wouldn't be surprised if an explicit
unmerging in addition to "rerere forget" was necessary, especially
if the original preimage were very different from the state before
the botched resolution was applied.