In my experience, OpenGL walks the middle line. There is certainly a core set of functions that (almost always, discounting buggy drivers) work. But the core set doesn't span all the critical functions you need for what modern game players would consider a performant game engine (such as hardware synchronization to eliminate "tearing"). So game engines will need to factor in the GL extensions, which put us in the "up to third parties to follow up" world. It's a frustrating environment to work in; you can't really ever trust that your code will either succeed or fail on any given hardware configuration, and you're stuck playing whack-a-mole on bug reports that you lack the hardware to reproduce.
> As for support of the latest features on my old Geforce 7600, I guess I should accept the fact that they cannot be implemented efficiently
I wish it were that simple. That, I could deal with.
I've worked with a card that had a bug in the GLSL compiler. A particular sum simply wasn't compiled, and we had to work around the problem by multiplying by 1.000000000000000001 to force the compiler to generate the bytecode for the whole calculation (the fact this trick works is something one of our veteran engineers "just knew would work," so we got lucky). There is functionally no chance of that software bug ever getting patched; card vendors don't care about older versions of their technology, and even if a driver version were out there that patched the bug, you can't trust machine owners to keep their drivers up-to-date.
More frustratingly, as I mentioned elsewhere, I've worked with cards that implement things that should be high-performance (like stages of the shader pipeline) in software, just to claim they have the capability. Since OpenGL gives you no way in the API to query whether a feature is implemented in a reasonably-performant way, you either do some clever tricks to suss this stuff out (F.E.A.R. has a test mode where it runs a camera through a scene in the game and quietly tunes graphics pipeline features based upon actual framerate) or gather bug reports, blacklist certain card configurations, and keep going.
Old cards not being as powerful as good cards I can deal with; if we could simply say "Your card must be X or better to play," we'd be fine. New cards with bugs and under-performant cards that lie about their performance to clear some market hurdles are the maddening corners of the ecosystem.