I wrote:
> It looks to me like the ON CONFLICT code has exactly zero chance of
> working with non-btree (or more generally, non-amcanunique) indexes.
Nah, I take that back --- I was expecting this to be handled somewhere
else than it is. It looks like the actual cause is just sloppy
handling of noDupErr in ExecInsertIndexTuples; the attached fixes it
for me.
regards, tom lane