First off, the material truth DOES report the "Top Fraction" correctly in scenes where the primary source of mixing is sub-pixel sampling. This problem is specific to using the "mix adjacent materials" option in a Material Map or using the Mixture Map.
* Run the UvMapping2 demo, which employs the "mix adjacent materials" option.
* Note that the "Top Fraction" truth is always 100% for a single pixel sample.
Using a single sample per pixel will not result in mixtures, but some of the mapped materials get mixed due to the material map image having DCs that are between mapped DCs (materials).
The issue here was structural -- the counts were being stored in a multiset-wise fashion that restricted the counting to whole numbers (to track material mixing we need to store fractions of a sample). The fix was to re-work this truth collector to store partial samples.
The fix has been checked into the trunk and may be released in a (very small) incremental update to the 4.5.x series.