It has to do with when and where the prefetches are processed vs when the relevance substitution is processed.
There is an explanation in the referenced document:
Relevance substitution is NOT performed on the prefetch action command lines. This is because these actions are scanned by other components that deliver the downloads and these other components run on different machines which do not share those client's evaluation context.
Generally I would recommend against using no hash prefetch item or wget/curl to download content unless there is no better solution due to those options being less secure and not able to make use of the Relay Caching mechanism.
There might be reasons to use them, but only with caution.
Agree, not the normal situation. However I’ve found use cases with business ERP systems, for example, where those alternative means are extremely helpful. Dynamic urls by business attributes (brand, country, location, subnet, etc) can be super useful in select circumstances.