One has to say it aloud at last: let's kill it. Say, in 5.4.
It's useless on modern hardware anyway, and if you look at re-designing it, it turns out there is need for multiple solutions, since no single one performs well in all caching workloads.
What would it give our users? One less gotcha-like feature in the server that they need to learn, learn to forget. It will open up a niche for third-party projects to contribute. And the core server development team will have one less thing to worry about.