In Whisky we had an issue related to that due to Con2Prim running after recovery from checkpoint.

E.g., you run a simulation from iteration 0 to iteration 24 (made up numbers just to explain myself).

You try the same simulation, but this time you run from 0 to 12, checkpoint, recover from checkpoint and run from 12 to 24.

The results between  these two simulations at iteration 24 were slightly different

The issue was in the Newton Raphson being called by Con2Prim and using a different initial guess when recovering from checkpoint. During a run it uses the previous values of the primitive variables as an initial guess, but these are not available when restarting from checkpoint. The solution, as far as I remember, was not to run Con2Prim during recovery from checkpoint (since the primitive variables are anyway recovered from the checkpoint files).

Cheers,

Bruno