> rely on the current behaviour.

I agree. The problem is that the method is used in rather complex
algos, so it's difficult for me to answer this question, esp.
when we also have to consider the changes in the high prec branch.

Gerd