This is also very true. And I think a part of the reason why this is worse with research code is that the meaty part of the code tends to be (at least in ML/NL

This is also very true. And I think a part of the reason why this is worse with research code is that the meaty part of the code tends to be (at least in ML/NLP) a few equations from the paper, in a hacky and convoluted way, and unless you're a world-class expert on keeping track of indexes and one-greek-letter variable names, there's very little to get from the code to a well-written paper. I make an exception for tuning parameters, constants, tweaks, etc, but these shouldn't matter much anyway.