This is not surprising. No refraction processing is used if IOR is 1.0, but anything other than that uses refraction code.
There's another possibility. There are two ways of calculating transparency: additive and multiplicative. Multiplicative transparency has the disadvantage of making things darker behind the transparent object, but it has the advantage that it produces an image that more closely approximates real life. Additive transparency is definitely what is needed for transparency maps and cookiecut maps, and for single surfaces with transparency on them. It could be that refractive objects use multiplicative transparency, otherwise additive transparency is used.