Lots of little, simple functions, composed... You can still use the threading at a higher function, but instead of doing the actual work at each step in the pipeline, you call a well-named function which does what that step was doing. Yes, it may feel redundant, but it absolutely improves human comprehension because it reads like clear human instructions.
Tests become much simpler too.
The only two challenges I face with this approach is 1 - naming things, and 2 - organizing. The function names tend to become quite long, since they ideally describe what the function does.
I haven't done this lately, now that the AI tools have become pretty great; I can imagine that an AI tool would be an excellent knowledge base for these many functions, and it would probably be very successful in providing additions or modifications given the clean, clear function names and brief contents.