It seems as if there's an internal heuristic that says "if the number
of tuples generated by a series of crossjoins will exceed 1 billion
then don't return any positions on the axis."

Is anyone aware of such a limit inside PTS/AS?

In a highly sparse cube it's entirely feasible to have:
... non empty crossjoin(crossjoin(...)..)
return a small number of rows. However the crossproduct can easily
exceed 1B tuples scanned prior to the non-empty processing.