Except unlike a REPL, all your code is still there so that you eventually find yourself in an unreproducible state that on the surface looks reproducible. Referencing variables that don't exist in the notebook (only to be caught when run by someone else), cells that need to be executed out of order, etc. The problems go on and on. People sometimes go days or weeks without restarting their notebook and the amount of "shadow" state is remarkable.

I think notebooks are useful (though only for exploratory analysis), but there has to be a better way.

Part of the problem is discipline and avoiding anti-patterns. It's unfortunate that Jupyter encourages this kind of inconsistency to creep in, leaving it up to the analyst to maintain top-to-bottom "runnability".

You can go for days and weeks without restarting your REPL. It's literally the same problem.