Code doing strange things.

I was investigating the part of my analysis that is responsible for recentering on the middle of the collapsed object.  I have not figured out whether it is working the way I expect it to, or not, but in the process I discovered that the code is behaving strangely in some of the runs.  For example, here is a normal sequence, in the warm sim with initial density profile $x^{-0.8}$, dumping out the particles every 20 ATUs (time units).  This run had 10004 particles in it.

Now here is a run that crashed (in the sense that all the positions were NAN, not in the hanging up on timestep 0, which is the other failure mode).  The only difference between this run and the previous one is that this one had 10007 particles, rather than 10004.  And of course the thermal part of the velocity has a random component.  What happens is very hard to explain:

Note the x axis.  The very next time dump was all NAN.  10010, 10013, 10016, and 10019 particles all looked identical to the blue run above, and ran in a well behaved way until the end.    If you zoom into one of those regions in Dumpout 3 of the crashing run, it looks like this:

So some of the coherence in the structure is still being maintained...   Thoughts anyone?