Typically you merge a block A and block B into a new block C that has the same length as A + B. I thought I saw a way to use a few extra pointers and some swap

Typically you merge a block A and block B into a new block C that has the same length as A + B. I thought I saw a way to use a few extra pointers and some swap operations to turn A and B into C by chipping away at their left ends, and still being a stable sort. The examples I came up with worked and confirmation bias took over. But in real data there were combinations of runs that broke the algorithm.