Okay Jean-Marc Spaggiari I'll post a patch for 0.94 as well. We can address bringing 0.94's version up to feature parity with 0.96/trunk in a separate ticket.
Yes, Andrew Purtell, I considered that. I'm not a statistician, so I don't know if that approach would be sufficient to keep tabs on the extremes. I'm also interested to pursue the technique employed by Gil's LatencyUtils project. Unless someone has immediate advice on the sampling implementation, I'd prefer to implement this as a separate ticket.