I'm not ready to give up but it seems all the low-hanging fruit doesn't fit into the salad.

I suspect you are correct, however, that without a new, inspired algorithmic approach to identifying optimization points in the merge, the platform constraints hamper conventional efforts to pre-select merge points (and other optimization efforts of that ilk).

In short, not enough hardware, too much data variance.

I'm not saying it can't be done. I'm saying no path to solution is clear to me at this point.

Solution to this one will require either a profound application of theoretical mathematics, or an invention -- likely one worthy of a patent.