First question: is the average~ object getting zeros or just numbers very, very close to zero?

In the former case, average~ should just spit out zero. I can't duplicate your results (a test patch might help), but I can imagine exotic circumstances where squaring (the 's' in rms) values very close to zero could result in denormalized interim results. And trying to take the square root of a denormal might NaN (ie, generate a machine result that is 'not a number'... that's like dividing by zero, only worse). But I'm not convinced that's what's happening.

Test patch, Max version, and platform will probably be needed to get any further then maybes.