Martijn van Oosterhout <kleptog@svana.org> writes:
> Ok, 2661 is definitly pg_cast_source_target_index on my system so the
> -P should have caused postgres to ignore it.
That was my first thought too. However, if the pg_internal.init file
were missing/broken then the thing would try to rebuild it, and I think
that would involve opening most system indexes. Try inserting a correct
pg_internal.init file --- ideally from a recent backup, but if you
haven't got one then it most likely will work to copy it from another DB
in the same installation, or in extremis do an initdb in a temporary
directory and take the file from there.
Of course, if the data loss extended to more than just index(es) you're
still hosed ...
regards, tom lane