Have you seen the size of the reference implementations for chacha20, blake2b, or poly1305? Those are tiny, a couple hundred lines for all three! There is very little room for errors to sneak in. Between test vectors, compiler warnings, and Valgrind, we can be pretty sure all bugs have been shaked out. Add in good old code review and the safety of this stuff is pretty much guaranteed.

Of course, I would never make the mistake of rolling my own OpenSSL clone (too big, bad API, some bad primitives), or even my own AES implementation (slow or prone to timing attacks).

Of course, I don't make the mistake of going blind. I read the literature before making any choice. If I don't know why some construction is considered safe, I don't use it. Due diligence first. And I certainly don't implement the first primitive that my search engine gets its hands on. I go for the most vetted alternatives.