I looked at all the data, and, both Scott and me tend to believe something is funny with the RPC model. RPC is usually an approximation, and if I am not mistaken, I see a very strong systematic effect in the intersection error (it would be easier to see
if this error were colorized, using the colormap model). Hence, I think there are are areas where this model does not approximate things very well.
I saw such behavior (if not looking quite like this) for ASTER a while ago.
I must say the fact that these data comes from Planet (formerly PlanetLabs) does not inspire confidence. This is low-cost data provider using cheap cameras, I would doubt they did the homework as well as say Digital Globe would.
I could suggest you read more into their doc, or try to get images from another source (ASTER could also work, some of their imagery is free I think, and our tools work well with it if the exact model is used and not RPC).
Taking the very long term, it is possible to write code to correct the RPC model, especially given that there is an underlying SRTM, but that would be a lot of work.
Oleg