Here the canonical representation refers to one single definite and unambiguous encoding for given data. This requirement is very common in cryptographic applications and also commonly demanded when the deterministic processing is desired. Technically the "canonical" and "deterministic" encoding can differ (e.g. ASN.1 CER vs. DER), but there is not much value to have two distinct encodings.

On graphs: as you've said recursive code has to be careful, but the recursion itself is not very frequent in normal applications and they are more susceptible to attacks if the recursion is built into the serialization format. XML billion laughs attack is a famous demonstration for this issue. It is still worthwhile to have parallel standards to specify how recursive structures should be encoded in the basic format, and possibly to tweak the basic format to better accommodate such standards, but I believe such needs can be met without making the basic format bigger.