Certainly if you do traditional TDD or something like REPL-driven development, you see the intermediate results and validate the correctness of your code as you go.
Those are great! But they are coding disciplines one must choose to follow, and continue following every step of the way. You aren't inherently following them because that's how the tool works. Sometimes the tool-enforced discipline has advantages.
How would one write tests for a complex, mission-critical Excel spreadsheet? Or use version control?
Spreadsheets often mix the data and code/formulas and the formulas are hidden behind the sheet view and sprinkled across many cells. At least Python scripts separate the code from the data so you can write tests using known good or fuzzing data. And you can use version control to track and review code changes.
I agree that, in the hands of a hypothetical ideal person, Python should be better than Excel for almost everything. But as my story above shows, even people you'd expect to be experts, seemingly following best practice, don't end up being very close to this ideal person.
Honestly, your story just indicates that someone senior needs to have a discussion with the applied mathematician about their coding practices (and they may even need a bit of training in that regard).
His code style is fine. His problem is that he is short on the discipline and focus to validate his results in meaningful ways. An exhortation to follow some vague "coding practices" won't fix that - he'll respond e.g. by writing lots of tests, none of which catch the problem his code actually has.
The two ways I know to get correct work out of him are (1) review it and kick it back to him when problems are found, (2) have him implement what he's trying to do in Excel.
So I was mainly a Haskell programmer in grad school before getting into scientific computing, and the main good habit I picked up was working at the REPL and breaking my code up into small, testable functions/components, and thinking very hard about state. The main bad habit I had was expecting the Julia compiler to be as helpful as GHC when refactoring code.
While I was writing my thesis and job hunting, I attended a few workshops aimed at “grad students breaking into industry”. The main thing I noticed from applied mathematics students in particular was they would write out long functions (really hard to debug) or they worked exclusively in Jupyter notebooks (these have super complicated state, so it takes a lot of discipline to be able to translate these into usable code).
