Also, I was curious why you bother to localize $/ ...

Untested code and two minds about how I would implement it. It also requires that the OP change his existing code to assign the returned filehandle, rather than passing it as a parameter per open. Which I considered a good thing.

Like you earlier, I'm not really sure what circumstances this would be useful, so I raised the possibility without putting too much effort into trying to make it bullet proof. I wanted the OP to either be sufficiently aware to fix it up himself, or ask.

I like your re-write++. One additional change I would make is to use a hard coded '<:raw' mode on the real open and the user supplied mode on the ramfile open. As you have it, if he passes a non-read mode things will go wrong. Though, that might be a good thing also...ponder...undecided.

It might also be worth doing some rudimentary, Is this a huge file? check. I don't like arbitrary limits, but issuing a warning if the file is bigger than say 100MB might be the clue stick to avoiding mysterious failures.

I also thought some about using one of those modules I never use to canonal...canonica....to ensure the paths are absolute and unique--save loading the same file twice--but with all the convolutions possible on *nix, it would take some serious thought.

Oh. And I'd definitely use unless( exists $cache{ $path } ) { :)

All of that said, from the OPs latest description of the application (nested CGI calls), none of this is likely to help, as the cache will get re-built every time the scripts are run.