I have a hunch. timestamper uses an exntesion point called console annotation which stores a small compressed gzip stream of data into the log file. It uses a lot of those, because each timestamp it's adding, I believe, is a new annotation.
If there's a race condition in the write operations, it could corrupt the gzip encoded stream, and I wonder if it can happen just in the right away that breaks the decompressor like this.
If my hypothesis is correct, trying to look at the problematic build log will consistently reproduce the problem (as opposed to "sometimes I can see build #15 but some times it crashes when looking at the same build #15.)
If this fits the issue you are seeing, please attach your console output (or send it personally to me if it's sensitive), and that'd greatly simplify our efforts to fix the issue.