And wrote:
it turned out to be even more confusing
g104 10 block s1.10G: katago 1.3 - OK, katago 1.3.2 - not working, katago 1.3.1 - OK
g170 10 block s1.97G: katago 1.3 - not working, katago 1.3.2 - OK, katago 1.3.1 - OK
g170e 10 block s1.14G: katago 1.3 - not working, katago 1.3.2 - OK, katago 1.3.1 - OK
apparently, before launch, you need to flip a coin
There's a good chance that most of those combinations are working, but you didn't let it run long enough?
When KataGo encounters a neural net that is of a different size than it has encountered before, it will need to re-tune the first time it sees that size. Different versions of the code (1.3, 1.3.1, 1.3.2) may not always be able to share this tuning, so each one separately does it. The re-tuning can take a minute or two - and during that time, KataGo will output messages that make it obvious that it is computing and tuning and working. But many GUIs are terrible at displaying those messages to the user, so it may look like it just hangs and doesn't load, if you aren't patient enough to wait out the minutes. But it only needs to do this once, so long as it is allowed to finish and succeed.
I know Jujube has reported that their issue is with running it in SmartGo specifically and that the command line works, so really maybe it has something to do with SmartGo, but how about for you? Can you try running those not working combinations directly in the command line too, so you can see if it is tuning?