I have observed that the primary problem with this file is that the GCPs span the international dateline, and that the ones west of the dateline are positive (~179). The mechanisms in GDAL for doing GCP based regressions do not model the wraparound nature of the geographic coordinate systems and this gets treated as a discontinuity and the transform attempts to drag part of the map across the rest.
One potential solution is a heuristic to use values < -180 for GCPs to the west of the dateline. But this might be hard to do properly for large maps (as this one is).
So instead, I will continue to look into deriving the proper mercator coordinate system.
