Thanks everyone c: I got some useful pointers here.

Came up with a pretty simple solution: I just added an import sub to the package that implements the cache loading; all it does is register other packages using it. The load function is then responsible for registering the objects that need to be read from disk or regenerated. Which gives us the following structure:

$cache={
  'package'=>{
     'object'=>[\$result,\&generator,@data],

  },
};
[download]

The order in which the packages and it's objects is registered is kept also; from this we make a stack of indices. And then, way I handled writing to disk:

218   for my $block(@blocks) {
219 
220     $block=freeze($block);
221     $body.=$block;
222 
223     push @header,$header[-1]+length $block;
224 
225   }
[download]

Where 'blocks' is the objects themselves and 'header' is a list of offsets into the file, both corresponding to the index stack. One can then just write it all into a single file:

229   unshift @header,int(@header);
230 
231   my $header=$signature.(
232     pack 'L'x@header,@header
233 
234   );
235 
236   print {$FH} $header.$body;
[download]

And then open, seek and read; so I can save data in big blocks but only keep in memory the ones I need.

I will have to implement a mechanism for minimizing reads on consecutive entries, but since the entries are sorted by access, successive cache lookups can be done in one go. Squeaky clean ;>

So yeah, I wasn't sure how I was going to solve the problem, and less than a day later it is essentially fixed. Nice.