"You cannot debug a framework you do not understand" is the right motivation for writing out the math, and the dependency-free implementations make this far more useful than the five-line tutorials.
One thing I would add to a piece this comprehensive: decoding parameters interact with retrieval quality in a way that surprises people. Turning temperature down does not fix a grounding failure, it makes the wrong answer more consistent, and confident low-temperature errors are harder to catch in review than variable ones. It would also help to pair each stage with the metric that detects its failure, because a reader who follows all of this still needs to know which number moves when the reranker regresses versus when the chunker does.