Well, I gave up my variable caching ideas, even if just for a while. I wanted to master the standard caching techniques (there had always been an excuse, some extension to write...). And I am very glad now because it starts to feel like having real control over my eZ implementations ;)
Among things that I've discovered were persistent variables, recommended by Gaetano in response to my var cache struggling, and completely unknown to me before. Even if not a complete substitute, still a very nice solution, one of those that you ask yourself how you could have lived without...
So what's so great about persistent variables without going into much detail?
They don't require any additional control. Persistent variables are compiled as part of viewcache, which means that they only expire when the viewcache expires, and they naturally follow the smart viewcache clear rules as well. What more control needed than that?
They seem to be available at all times, if not cached before, calculated upon request for the full view that stores them.
This also means that expiry times of cache-blocks that depend on that data do not have to be synchronized in any way (not that it is even possible...). Gaetano mentioned the relationship between viewcache and cache-block clear being an important issue, but I haven't been able to spot any unwanted or problematic link, yet.
Persistent variables can indeed store a variety of data useful for further generating parts of the pagelayout. Even if the singular cost of viewcache generation goes slightly up, this is usually benefitial: once generated, the variables are simply there as long as needed. And in most cases all I need is some $node-related operations.
Even if some more expensive data is to be fetched, once you're through the effort of composing you persistent hash, it's there!
Since the variables are physically serialized, it is impossible to pass complex objects, such as $node directly. You have to precisely choose which simple data you want (numbers, strings, arrays, hashes...). Luckily, arrays do just fine in most cases, so that's probably as much as one may need before going straight to a fetch outside of a full view.
Then, some includes seem to be able to damage a persistent var that had been generated few lines before, so you have to carefully test if the persistent var set is available in all the full views required. Haven't found an exact rule, yet.
If you're interested if a very nice use example, look here in the comments.