So isn't the conclusion then that they are ignoring your robot.txt and get the names from the xml-files which are refered to in the html-files?
They are grabbing the XML's. But there are no explicit filenames to grab there. The filenames are produced from the XML's. This is how the name of the thousands jpg's are presented in the virtualtour.XML
<level tiledimagewidth="898" tiledimageheight="898">
%v and %u are counters which increase from zero to a maximum integer found or calculated from somewhere else in the XML's.
For me it looks like Google-image has reverse engineered the way krpano make tiles and has the ambition to download all the tiles, for maybe later to put them together and so get the original image. They have cooperated with NSA on other issues, so this is probably a piece of cake when you have this kind of resources.