Fascinating stuff. Here are some largely unrelated hasty generalizations based on my limited experience in the world of software development.

I'l classify the software development work I've done as falling in to three broad categories: frontend, backend, and data analysis:

• Frontend software development is the most forgiving of mistakes & imperfect code. There's little opportunity to do permanent damage to production data. It's typically easy to test your code by clicking around a bunch, so thinking through the problem carefully to ensure correctness by design is often overkill. Also, requirements are more subject to change and code gets thrown away quicker, so code written to be correct by design has less time to accrue benefits over the long term.

• Backend software development tends to be less tolerant of mistakes.

• Data analysis is arguably deserving of the highest quality code, because it can fail silently. Illustration: At the last company I worked at, I was the maintainer of the company's email system, including marketing emails. I once found a subtle bug in code written by one of the company's data scientists that caused it to double or triple count purchases generated as a result of marketing emails. Sometimes you can write test cases for data analysis code, but it was very difficult in the environment we were using, and it can be hard to capture all of the corner cases. However, software developers are perennially overconfident in their ability to produce bug-free code (the more experienced you get as a backend developer, the more paranoid you become), and I wouldn't be surprised if the industry as a whole has yet to catch on to this problem of analytics code silently giving bad numbers. (Also, much data analysis programming is surprisingly unsophisticated mathematically... it's mostly sums and averages.)

I'm highly intrigued by Steve Yegge's liberal/conservative dichotomy of software development cultures. Under this paradigm, liberal cultures (more common at smaller/startup companies or those working on less critical applications) are more tolerant of errors, while conservative cultures are less tolerant.

If anyone is interested, I recently scribbled down some tips on writing reliable software for a friend working in finance who is learning to code to automate aspects of his job. (Code review is something I didn't mention that works well if you're working in a group.)

By the way, does anyone have any idea why, as Jonah says, there is so little metacognition among mathematicians? I found the same thing to a surprising degree in software development (by thinking about my thought process while coding, I was frequently able to generate lines of inquiry I hadn't read or heard about from anyone).