I suspect this is due to large pages and/or the memory reservations. Large pages are not shared, but can be broken down into smaller pages by TPS. This only occurs if there is memory contention. I believe with newer hardware and 64-bit guest OS's large pages will be used by default.
With memory reservations set however, TPS will not break down large pages.
I talked to VMWare support about this. I described the situation where last night at 9pm I recomposed this pool. The recompose operation finished in about 20 minutes. From 9:20pm until this morning at about 7:30am, no one logged into one of these desktops. At 7am I looked at whether transparent page sharing actually saved memory, this is when I took the screen shot. The VMWare support person I talked to said that transparent page sharing only kicks in when the VM is active. In other words, it is only when students start logging into these VDI machines would transparent page sharing start looking for savings. Since these are linked clones set to refresh on logoff, no saving would ever be found. In the end his explanation is linked cloned set to refresh on logoff would never have memory savings through transparent page sharing.