When the base field is pseudo-Mersenne, taking mods is actually fine, since the statistical distance to uniform of the resulting distribution is O(2^{-λ}) at the λ-bit security level. In other words, an attack that exploits the bias is at least as costly as breaking the discrete log problem directly, and therefore not a security concern. So I wouldn't worry about implementations of secp256k1 that do it like that (although it would be simpler and less "leaky" to just use the non-reduced 256-bit MAC value as the nonce; of course, anything shorter than 256 bits would be a huge problem).

But careful approaches like RFC 6979 are very important for curves over random fields, like the Brainpool parameters.

[By the way, a more regular contributor of this site suggested that it would be appropriate for me to mention that I'm one of the authors of the OP.]