The keys are integers, not strings.  The code accidentally worked on
little-endian machines, at least up to 256 distinct record types within
a session, but failed utterly on big-endian.  This was unexpectedly
exposed by a test case added by commit 4452000f, which apparently is the
only parallelizable query in the regression suite that uses more than one
anonymous record type.  Fortunately, buildfarm member mandrill is
big-endian and is running with force_parallel_mode on, so it failed.