In my experience valgrind does produce false positives, especially in
optimized release builds.
I'm not saying that this is the case here but it is a possibility.
To better understand the issue you can use valgrind's capability to
launch gdb when this happens and you from current instruction,
My stress test is showing a problem with PDFReference15_v5.pdf (doesn't appear
to occur with PDFReference15_v6.pdf).
I'm having a lot of trouble understanding the valgrind backtraces:
Doing PDFReference15_v5.pdf :