In the discussion of the problems that occur if you load the same DLL by both its short and long names, Xepol asserted that any program which loads a DLL by its short name "would have ONLY itself to blame for making stupid, unpredictable, asinine assumptions" and that Windows should "change the loader to NOT load any dll with a short name where the short name and long name do not match."
The problem with this rule, well, okay, one problem is that you're changing the rules after the game has started. Programs have been allowed to load DLLs by their short name in the past; making it no longer possible creates an application compatibility hurdle.
Second, it may not be possible to obtain the long name for a DLL. If the user has access to the DLL but not to one of the directories in the path to the DLL, then the attempt to get its long name will fail. The "directory you can access hidden inside a directory you cannot access" directory structure is commonly used as a less expensive alternative to Access Based Enumeration. Maybe your answer to this is to say that this is not allowed; people who set up their systems this way deserve to lose.
Third, the application often does not control the path being passed to LoadLibrary, so it doesn't even know that it's making a stupid, unpredictable, asinine assumption. For example, the program may call GetModuleFileName to obtain the directory the application was installed to, then attempt to load satellite DLLs from that same directory. If the current directory was set by passing a short name to SetCurrentDirectory (try it: cd C:\Progra~1) then the program will unwittingly be making a stupid, unpredictable, asinine decision. (But since the decision is being made consistently, there was no problem up until now.)
When you call CoCreateInstance, there is nearly always a LoadLibrary happening behind the scenes, because the object you are trying to create is coming from a DLL that somebody else registered. If they registered it with a short name, then any application that wanted to use that object runs afoul of the new rule, even though the application did nothing wrong.
Now, you might argue that this just means that the component provider is the one who made a stupid, unpredicable, asinine assumption. But that may not have been a stupid assumption at the time: 16-bit applications see only short names, so it might have been that that's the only thing they can do. Or the component may have made a conscious decision to register under short names; this was common in the Windows 95 era because file names often passed through components that didn't understand long file names. (Examples: 16-bit applications, backup applications.)
Read more: The old new thing