> I've also seen people mangle well-factored but untestable code in the process of writing tests, which can be a tragedy when dealing with a legacy codebase that was written with insufficient testing but is otherwise well-designed.

Have you read Michael Feathers' Working Effectively with Legacy Code? [0]

In his definition of legacy code it is any such code that has no test coverage. It's a black box. There are errors in it somewhere. It works for some inputs. However you cannot quantify either of those things just by "inhabiting the mind of the original developers." The only way to work effectively with that code base in order to extend it, maintain it, or modify it is to bring it under test.

This is far more difficult than it sounds with legacy code than with greenfield TDD for the aforementioned reasons: there are unquantified errors and underspecified behaviours. You can't possibly do it in one sweeping effort and so the strategy is to accept that tests are useful and to add them with each change, first, before making that change and using the test to prove the change is correct.

Slowly, over time, your legacy code base surfaces little islands of well tested code.

You have to be deliberate and careful. You have to think about what you're doing.

This is a much different experience than writing greenfield code. TDD is effortless and drives you towards the answer in this case.

[0] https://www.amazon.com/Working-Effectively-Legacy-Michael-Fe...