You won’t believe it, but it seems like a timing issue that was revealed by us setting loglevel=3 in the latest images.
If you set the loglevel back to 7, it works again (most of the time at least, how it used to be). My hunch is that the log prints introduce tiny delays that are required for the functionality of dwc3 on LS1028A.
If you have serial connected you can test this even dynamically by doing something like this:
rmmod dwc3
dmesg -n 7
modprobe dwc3
Then USB should work. If you set dmesg -n 3, it should again cease to work.
This is so far the most ridiculous issue I have encountered. To root cause this we would need to find all the log print sites and sprinkle with delays (or it could also be cache flush related).