Even that OS with the biased /dev/[u]random using RC4 because it was fast would keep a separate RC4 state in each process' libc to avoid the overhead of kernel calls.
Our crypto libraries need to provide their own belt to wear with the OS-provided suspenders. Yes, I know about potential fork() bugs, but our modern VM-crazed deployments introduce that possibility at kernel level too. We need our good crypto implementers shipping good lightweight CSPRNGs, rather than the kernel developers who try to compensate with 6000 bit entropy pools.
Meanwhile, better to have just one codebase to review, rather than 100 crappy ones, like the SecureRandom from Harmony.
I'm sure these principles of minimal redundancy will be of great comfort to those users who get their private keys exposed.
Do you see the paradox here? Find me these developers who are undoubtedly competent to implement their own CSPRNG, but prefer not to. These are the folks I want implementing the CSPRNG that will be generating my keys, rather than kernel developers who find entropy pools exciting.
We have already seen a huge number of bad keys generated. Is it that obvious at this point that kernel (and embedded system) developers are so much better at this than OpenSSL?
Is it possible that you just tend to look at more broken library and app crypto code during the course of your work? If you worked primarily with broken kernel crypto code, would you perhaps prefer (for your own use) a CSPRNG in a library written by your favorite experts?