I logged what actually happened on this, and ended up somewhere unexpected. Strip out the noise and the size of the local pool is worth more than any feature list. It stayed consistent to a degree that mattered once I had a few weeks of data. Take it as one data point ultimately, whether others found the same thing.