I completely agree. The issue with the "clean" version of the code is that it's completely coupled together. If you touch one thing in the math code, it affects everything. At some point, it becomes easier to keep some duplicated code, point out the similarity using some code structure cues or comments, and over time, highlight any key differences using more comments.
A "clean code" approach to this kind of code doesn't handle evolving requirements well. It always ends up as a mess of helper functions with multiple code paths within depending on where the function is called from. This can sometimes be abstracted: some people might have noticed already the previous sentence is describing OOP polymorphism. But how would anyone know the exact requirements ahead of time? If we're in such an early phase, any attempt to "clean code" will only result in a bad abstraction, or in overengineering. The right approach is to do nothing, note the possible issue, and wait until more is known about the problem domain.
We could call it The Premature Abstraction Anti-pattern :)
Been a victim of this myself a few times myself so I normally do exactly what you've suggested and wait until a clear, and still performant abstraction has emerged from the code and slowly refactor-in the abstraction. But it's always better to do this on a nearly completed codebase, so at the end, where you should naturally be looking to remove lines, not add.
A "clean code" approach to this kind of code doesn't handle evolving requirements well. It always ends up as a mess of helper functions with multiple code paths within depending on where the function is called from. This can sometimes be abstracted: some people might have noticed already the previous sentence is describing OOP polymorphism. But how would anyone know the exact requirements ahead of time? If we're in such an early phase, any attempt to "clean code" will only result in a bad abstraction, or in overengineering. The right approach is to do nothing, note the possible issue, and wait until more is known about the problem domain.