Yes, FME is pretty much the home team, so we're all kinda partial to using it where it fits. And generally we've had good experiences with the software and the support (not to mention that the licensing is cheaper!)
We've done some more testing, and found out that:
* indeed, using a pure ESRI pipeline to load the FGDB into SDE (via ArcCatalog) SOLVES the polygon problem. The polygon structure is reshaped into an ESRI-valid structure (an inverted polygon).
* It turns out the the latest FME Beta (2012?) still exhibits the load failure on the invalid data.
* That let us verify that re-shaping the polygon using Union did actually fix the problem, since teh updated dataset loaded into SDE via FME 2012 Beta with no problems.
Thanks for the warning about Union. We realize that it's a pretty blunt tool, so we want to be very careful about what we update using it. Since Check Geometry and Repair Geometry are non-starters about fixing these invalid polygons, I"m hoping that we can write some custom Python code to identify just the polygons in error, and then fix each one individually, to cause the least amount of change in the dataset.