Maybe I misunderstood your question. I thought that "as far as I know, I got the crypto right in [product where I wrote my own crypto protocols]" was a pretty clear answer to "have you ever written your own crypto and gotten it right?"
I believe I got the crypto right in version 1 of tarsnap, but I wasn't working as part of a (non-degenerate) team. I've never done crypto with a team -- so my answer to your (clarified) question must be a vacuous "no".
Not really, no. Amazon screwed up on a well-known widely-understood rule; there aren't 40 different sides of the road involved with "hash collisions are bad".
Similarly, the one cryptographic complaint I have about OpenSSL (as opposed to non-cryptographic security bugs) is their lack of attention to side channels, and the same applies here -- we KNOW that side channels are bad. The fact that the OpenSSL developers haven't made any systematic attempt to prevent them doesn't mean that cryptography is hard; it just means that the OpenSSL developers are careless.
That's an interesting criticism of OpenSSL. Do you know of a general purpose cryptographic library that has done a better job than OpenSSL at hardening against side-channel attacks?
By default, OpenSSL has: RSA blinding, constant time montgomery multiplication, careful review and correction of the conditional branches which were attacked in that amazing series of branch prediction papers.
Doesn't seem to me that OpenSSL developers are careless and indifferent about side-channel attacks.
Do you know of a general purpose cryptographic library that has done a better job than OpenSSL at hardening against side-channel attacks?
No, I don't know of any better cryptographic library. (This is why I said a few days ago that writing a better cryptographic library was something I'd like to do if I was rich.)
Doesn't seem to me that OpenSSL developers are careless and indifferent about side-channel attacks.
They're playing whack-a-mole. There are over a thousand conditional branches in the large integer arithmetic code alone. The fact that nobody has pointed out how to exploit them yet doesn't mean that they're not exploitable.
People make dumb mistakes all the time, but usually all that happens is that the pie graph doesn't render properly, or the query doesn't use the indices. That's the difference between security bugs and application bugs.
Inviting crypto to the party invites a whole host of new security bugs that you simply cannot have if you don't go down that road in the first place.