I see your and MaxBarraclough concerns. In my case, there exist 5-6 codes which do -at their core- the same thing as ours does and they all have been cross-checked against each other within either theoretical or numerical precision (where possible). That's the spirit that sjburt was referring to, I guess, and which triggered me because it is only true to a certain extend.

The cross-checking is anyways good scientific practise, not only because of bugs in the code (that's actually a sub-leading problem imho), but because of the degree of difficulty of the problems and the complexity of their solutions (and their reproducibility). In that sense, cross-checking should discover both, scientific "bugs" and programming-bugs. The "debugging" is partly also done at the community level - at least in our field of research.

However, it is also a matter of efficiency. I -and many others too- need to re-implement not because of bug-hunting/cross-checking but simply because we do not understand the "ugly" code of our colleagues and instead of taking the risk to break existing code we simply write new one which is extremely inefficient (others may take the risk and then waste months on debugging and reverse-engineering which is also inefficient). So my point on writing "good code" is not so much about avoiding bugs but about being kind to you colleagues, saving them nerves and time (which they can then spend on actual science) and thus also saving taxpayers money...