In discussion on IRC, darkblue and I have discussed a few different use cases that we should try to make sure happen in the short term, without neccesarily doing any work with localization.

Some possible test cases to determine where GDAL breaks, without localization being neccesary:

• Convert UTF-8 encoded KML file to a shapefile using ogr2ogr. Ensure that the same bytes from the KML file are maintained in the ogrinfo output of the shapefile.
• Store UTF-8 encoded bytes in a shapefile attribute. Ensure that when fetching the attribute value, the bytes are the same. (crschmidt has tested this and confirmed it worked for him.)