Your boundary conditions are fine as you’ve previously successfully ran a simulation with them. I would probably change the top bottom BC to a slip wall so as to reduce the chances that the domain is interacting and influencing your simulation.
The issue, I believe, is your Courant number for the second simulation being significantly higher than the ideal value of 1. SimScale has a great article here that describes what the Courant number is, how to estimate it and a case study to show what can happen based on what Courant number is used.
For your first simulation, assuming I take the smallest cell length I could find (0.01m), with your timestep, your Courant number turns out to be around 23. The second simulation producing a Courant number of 459.
Now the obvious question, after you read the SimScale resource I linked, is, “In both cases, my Courant number exceeded 1 although the first simulation did converge. What gives?”. The “special sauce” in this case is your usage of the PIMPLE algorithm. As shown here in a discussion, the PIMPLE algorithm allows a Courant number of more than 1. Of course, how much more is hard to tell, but I think 459 is much too large. As for why this is possible, I haven’t read through how the PIMPLE equations work. Maybe my friend, @1318980 Darren can help you on how PIMPLE allows the Courant number to exceed 1, albeit at the cost of possibly reduced information about the flow.
So assuming I guessed correctly, lowering your timestep and keeping the Courant number as low as possible should fix your simulation divergence problem. Maybe start with the timestep you’ve already successfully used (0.05) and work from there.
Do let me know if it works. Cheers.