Apart from making information easily retrievable, which ironically academia at least used to be notoriously bad at, inventing an algorithm and using it effectively in an application in many cases are two very different matters.

These arguments often get handwavy rather quickly. Real-life applications have to accommodate edge cases and non-functional requirements, something a scientific paper doesn't need to account for.

That's perfectly fine. It isn't a scientist's job to worry about implementation details but those shouldn't be brushed off lightly either.