Failure of richards benchmark

There was a spurious failure of the Richards benchmark in the CI. No such failure had been previously observed in many runs. The job log is attached job_log

Seems to be quite a consistent failure after !2736 (merged) when using the new runner on sal.