What you are describing sounds like comparing traces of two supposed-to-be-identical programs to see where they first diverge. That's not what this is describing. There is no trace at all, nor any decision about what to trace or what not to trace.

There is only a decision about whether to use the old or new implementation based on the hash of the call stack at that moment. Then you binary search on hash values to identify the exact call stack hash for which the new implementation causes a problem. Then you run the program once more with the instructions "print the stack with this hash".