Having submitted quite a few bugs to ESRI over the years, I am frankly a bit... jaded... and am now reluctant to submit every little issue I find (of which there are many). However, if I can't find a good work around to an issue, I will still take the time to submit a formal bug fix and repro case... In this case though, I have a good work around (keep scratch raster in GRID format).
I did try to post my dataset to this thread, but it is too large (1.9 MB zipped)... A clue that helps: I can repoduce the error with my original dataset that is 8287 cols x 12513 (~30,000 unique values)rows, but when I try to repro with a smaller dataset (4913 cols x 7165 rows and 18,000 unique values - which I created in order to make some repro data small enough to post to the forum), the output result is correct (i.e. I can't reproduce the issue with the smaller dataset). Like Pavan from ESRI said, it seem to only affect "fairly large" datasets.