Isn’t the answer just: all models are wrong, some models are useful? And which model is most useful depends on what you’re using it for?

It seems like any fixed rule (like “use things out there in the real world, not vibes”) will be optimal in some situations but suboptimal in others. If I need to look up a product in the self-checkout machine, I hope they are categorized by vibes and not by lineage or morphology, or else I’m never going to find the tomatoes.

Really the trick is just to be clear what definition we’re using, whenever we use it. There’s no need for just a single definition of “fish” or “vegetable.” Imagine if our language let us label our words to indicate which definition we’re using: like 1=vibes, 2=morphology, 3=lineage. The self-checkout screen would not have a “vegetables” category, it would have “vegetables1”. Two scientists talking to each other would say “fish3”.

That would avoid a lot of the hand-wringing over what “the” definition of fish should be.