No, I just wouldn't make the assumption in general. Very few BGAs I've seen have had some thought put into ease of routing it out. Many of those are assumed to go onto 6+ layer boards with blind vias the norm, not the exception, and the internals instead designed to make life easier at the silicon level.
iirc, the RockerBox was similar in layout to the Minion chip I linked to - i.e. alternating gnd/pwr for most of the chip and i/o concentrated at one edge - just with fewer balls. It's not entirely bad, but I can understand why e.g. sidehack would rather not go near it. Still, that's a lot more 'intelligent' than some of the intel stuff that is a semi-random mess with two different ball pitches in die shadow / skirt .. at least for those needing to plunk it down on a board.
This checkerboard pattern is not "making life easier at the silicon level". It is a symptom of using completely invalid approximate model of parasitic components and then using those approximated, but not really existing, parasitics as an additional power decoupling filter.
It seems like despite the claims of "full custom" Spondoolies is still using only a very approximate all-digital low-power models. They are essentially invalid because they are used outside of their range of validity (in terms of self-induced noise and dissipated power).
I'm not sure if I understood your 2nd paragraph. It seems like you think I proposed using two different pad pitches: one denser for the chip and one coarser for the board. What I intended to convey is using the same pad pitch on both chips and board. The power traces are bundled in the way similar to
https://en.wikipedia.org/wiki/Litz_wire .
The bundling and spacing is wider to so that simple faults like shorting two neighboring pads doesn't create board-level and circuit-level fault.
I'm not up to speed with the most recent digital EDA tools, but I don't believe that their low-power all-digital tools can properly model the skin-effect and long-transmission-line-effect for those types of pad and board layouts. To get accurate models one has to drop to the all-analog or nearly-all-analog modeling.