From what I can see this is not (necessarily) a failure of CarpetInterp,
but of Carpet itself. This testsuite uses the three-level
initialization, coupled to a few refinement levels. This scheme seems to
have problems in that case. We have other testsuites using it, but not
with mesh refinement. Using only unigrid with the failing testsuite
seems to work (doesn't produce nans and non-sensical time values like
when using mesh refinement).
Without a deep analysis it looks like that either the three-level
initialization in Carpet is broken when using mesh refinement, or there
is some error in the parameter file that both I and Carpet didn't catch.
Before spending time by looking into a possible fix in Carpet: is anyone
actually using or interested in this feature? If not, we might as well
use the time elsewhere and disable/remove it.
I would like this to work.