Scientific software should absolutely, always be verified through regression and unit tests. Anything less is non-negotiable.
In a decade of work in hpc and computational science, I have very seldom found looking at the code to be a useful tool for either verification or debugging.
Instead, use the scientific method: hypothesis testing by constructing simple examples with known analytic solutions and using that for clues as to where the real problem lies.
