So I think I got lucky. It appears that 3.11rc1 has a patch fixing how the transmit power was being set.
This patch also went into 3.10.2.
Apparently this bug has always been in rtl8192cu there, and was going to be fixed in 3.1rc1, but then it accidentally got reverted. I tested 3.11rc5, and everything is good so far (using the device to write this message).
As for why the ARM device was not affected, somebody slipped the realtek alternate driver into the raspberry pi tree
This also shows me why it would have been difficult to bisect.
I don't see any movement on problems with rtl8192se, but if you believe this is a regression, you might be able to bisect it.