The reason being that breeding is akin to black-box optimization -- it doesn't concern itself with genomic details at all, the phenotype being the optimization target. If we want to do gene editing directly on the other hand, we need to have some way of mapping a genotype-level edit to a phenotype-level effect.
The first CRISPR racehorses edited MSTN, which is the single best-understood speed locus because centuries of breeding already sorted it. Thus: we had a population that was varying at the MSTN gene site -> regressing by performance on the sequenced genomes of those horses flagged this site -> CRISPR was able to make a targeted edit there. Ideally we need naturally occurring, phenotyped variation to find sites worth editing.
Notice how this confines gene editing to spaces where we have confirmation that a genetic site affects the wanted phenotype. Entering a new trait-space is hard. So to your point that gene editing can introduce mutations that didn't first arise by chance: I think it's true, but novel mutations are precisely where we don't have any data (and any one experiment takes years and is very expensive).
Gene editing today is successful in specific research areas: polled cattle, PRRS-resistant pigs, disease knockouts, etc... These are places where the mechanisms that the gene-editing is supposed to target are well understood (theoretical machinery is more built out here).