a few comments on the fly:

• can you revert your changes to the submodules kokkos and reference so that the tests will pass, and we don't downgrade kokkos accidentally
• I agree that slope limited reconstruction is probably not required for the gravitational potential.
• I will have a look at your test and why it might fail. Have you tried to run this on different architectures without MPI? Do the results match?

For the test: I tried it on A100 (no MPI) and :

• the custom testidefix.py test yields the exact same error and passes
• the non-regression test fails with similar error (but slightly different ~2e-8)

Some of the variables in the test were not initialised, so this can certainly lead to different results.

In addition, the test provided is not great (as indicated by the large error of testidefix.py), as it seems like it assumes perfect advection of the potential by the shear, but at the same time, density changes because of sound wave propagation, and there is this weird VX1 that is I believe a residue from the hydro shearing wave test (is it?). This should really be cleaned up.

Can't we have a shearing-box+self-gravity following a self-gravitating shearing wave? Would make more sense IMHO.

I added a test for the shearing wave. I mostly copied the setup section 4 of from Paardekooper 2012. But there is a key difference explaining the difference with this test and their Fig 3: This setup is purely 2D, not with a razor thin disc. This is why the ODE in the python script is different from their eq 89 (there is a factor 2/k different) and the produced result is not the same.

On my side it passes with MPI and on a single A100.

I can remove the test I first implemented, as the shearing wave is a better one.

This test also passes with no self gravity, but I did not include that version, to avoid redundancy with the already exist shearing box test. Let me know if you think it should be included as well.

Rebased to have #346