This is why I wanted to close the issue with the pure Python implementation, and punt on the question of a C accelerator for the moment.

compare_digest is effectively the same as what all the Python web servers and frameworks do now for signature checking. Yes, it's more vulnerable to timing attacks than a C implementation, but it's going to to take a sophisticated attacker to attack that through the noise of network jitter. It's sufficiently hardened that's it's unlikely to be the weakest link in the security chain.

For 3.4, I hope to see a discussion open up regarding the idea of something like a "securitytools" module that aims to provide some basic primitives for operations where Python's standard assumptions (such as flexibility and short circuiting behaviour) are a bad fit for security reasons. That would include exposing a C level full_compare option, as well as the core pbkdf2 algorithm.