I'm aware that the implementations are not strictly equal, but how is libsodium doing it 'wrong'? Is it because it does not partition reads into smaller chunks (avoiding the unspecified >= SSIZE_MAX read)? I consider any reasonable answer a valid bug report.
I should also point out that randombytes is technically not part of NaCl. The authors maintain that random byte generation is out of scope for NaCl, and what actually happens is that randombytes.o is bundled separately from libnacl.a [1]. In TweetNacl the randombytes implementation is completely omitted.
[1] https://bugs.debian.org/cgi-bin/bugreport.cgi?msg=20;bug=694...