> This is a kernel bug, not a bug in the GNU libc (ask Ulrich if you are not sure ;-)).

Kernels prior to 2.6.23 didn't know about the O_CLOEXEC flag: to catch this kind of problem, every syscall would have to check every bit when it's passed a combination of flags. This would be clumsy, error-prone and slow.
It's not a libc bug either.
The problem is really a distribution issue: using a libc defining a flag unsupported by the kernel is really calling for trouble.

> An host can have multiple kernel versions (and choose at boot time using GRUB/LILO/...)

It's possible, but it's definitely a bad idea, because of such API mismatch. For example nothing prevents a syscall from being removed/modified from one kernel version to another. If your libc doesn't follow, you're up for trouble.
Try using futexes with a kernel not supporting them :-)