The conditional in sample_ppc is not set up properly in this case - as a result the new y_hat is not conditioned on the newly generated x. The reason here is that, in sample_ppc the RVs only updated according to the inputted point (a dictionary containing one posterior sample), but not the newly generated values.
For now, the only way to achive what you want is writting your own ppc function to do posterior generation. I think this is something we should improve, could you raise an issue on Github?