As @jimhami42 has described, computer arithmetic has finite precision. The sort of small glitches you observe are to be expected in general. Comparisons of calculated values always have to include some tolerance (which is the source of the notorious vertex merging in small size models). However, it is odd that Transformation#identity? doesn’t account for this well-known fact.
This observation (which I can duplicate via an identity Array made into a Transformation by Geom::Transformation.new(array) ) is truly strange and worries me a lot. It seems to say that the identity? test relies on some undocumented flag that is only set by internal code not accessible via the API. There does not appear to be a method by which one can create a new Transformation that will pass this test!