Technically, that is correct -- only 'eval' doesn't fork another shell. However, from the point of view of the application you're trying to run in the modified environment, the difference is nil: the child inherits the environment of its parent, so the (modified) environment is conveyed to all descending processes.
Ipso facto, the changed environment variable 'sticks' -- as long as you are running under the parent program/shell.
If it is absolutely necessary for the environment variable to remain after the parent (Perl or shell) has exited, it is necessary for the parent shell to do the heavy lifting. One method I've seen in the documentation is for the current script to spawn an executable file with the necessary 'export' language, and then trick the parent shell into executing it -- always being cognizant of the fact that you need to preface the command with 'source' if you're trying to leave a non-volatile version of the modified environment behind. A Kluge at best.
The second method is to modify the script that initiates the shell environment (.bashrc or whatever) to contain the modified parameter. This can be dangerous -- if you hose up the initialization script it may make your shell unavailable the next time it tries to launch. There are plenty of tools for modifying the current shell; by affixing the necessary tweaks to the 'launcher' you effectively push those changes forward as well.
Generally not a good idea; if you only need the environment changes for a particular application suite, you'll have to go back and return the shell launch script to its pristine state (using vi or whatever) afterwards.
In short, there are no good (and easy) methods. Presumably this was made difficult to ensure the security of the system was not irrevocably compromised.
