The scale parameter returned by GEE estimate_scale is the inverse of the scale parameter returned by GLM estimate_scale. Everything is internally consistent so this doesn't affect any of the derived results.
I should address this at some point but it is a bit tricky, so I'm just leaving this note here for now.
This issue is now fixed in my GEE branch.
Fixed long ago and merged.