Similar approach as you it looks like. The only optimization I could find was that I could recreate the bracket by comparing top to bottom over and over again which allowed me to free up some memory and vectorization. Still looking to trade mem for speed.
colab.research.google.com/drive/1_3pgu...

I now see the other side is at most 1/2. That would give 69.

"The at most half the time" makes me think it would be 68.