Rows are aggregated, steps are coarsened, and dimensions are rolled up.
Rows are aggregated. Steps are coarsened. Dimensions are rolled up. Each word names one operation, and none of them is arithmetic over cells already on screen.
A dimension that rolls up has one coordinate past its last member, which holds every member together. A formula writes it as all, as in MRR[plan=all].
The rolled-up coordinate is not a sum of the members. It is the model computed at the whole dimension, so a driver that divides gets the right ratio for all plans together. One run computes it along with every member, so the roll-up is already in the grid.
A coarser view along the primary dimension, such as quarters of a monthly model, is the model computed again at the coarser calendar unit. It is a run of its own, because the steps are what offsets and running totals walk along.
Each coarse step takes the phase that covers most of its fine steps. A tie goes to the earlier phase.