It seems to be a problem with metis' interpretation of the vertex weights. If you set the vwgt=0x0 in a debugger when stopped at metis_nodend_ it completes successfully. Also, if you set the weights to 1 inside dmumps_part2.F like:

         DO I=1,NCMP
           FILS(I) = 1
         ENDDO

it also runs fine. I think having zero vertex weights is probably wrong anyways.

Thanks for the bug report. This accidentally happened due to my changes to patch MUMPS. This should work in petsc-dev now.