The "Average document size" stat in cbhealthchecker comes from the "item_alloc_sizes" histogram in the cbstats timings report.
I've been told that item_alloc_sizes is just a rough measure on incoming sets to get an average item size, but does not reflect non-resident items or info since a reset, and is not really accurate enough to be called "Average document size".
Basing on another stat, like ep_kv_size, is preferred I think.
Having a histogram of item sizes is great, and if it were available and mostly accurate for the entire data set, that would be wonderful. But that sounds like a different feature request for ep-engine.