Biases caused by modulo reduction are dangerous when the source integer is relatively close to the target distribution. This bias is unlikely to be useful for a

Biases caused by modulo reduction are dangerous when the source integer is relatively close to the target distribution. This bias is unlikely to be useful for attackers. The probability difference between the most likely output, 0, and the less likely outputs is 2^-32. For comparison, Cryptocat's bug had a much larger distance of 1/251 ~ 2^-8, which greatly reduces the amount of samples needed for distinguishing the output from random. EdDSA itself uses modulo reduction for the nonce, using a large enough source (~2^512 modulo ~2^252) value that renders the bias irrelevant.