why do best features selected by information gain add unsignificant improvment in text classification by SVM?

Your statistics are not really new features, but derivatives from other features (words). For the sake of simplicity, consider linear regression model instead of SVM:

h(w; k) = k0 + k1*w1 + k2*w2 + ... + kn*wm

where x1..wm are your words and k0..km are model parameters. You have added some new features - statistics of words (let's call them si and corresponding parameters - r). So your model now looks like this:

h(w, s; k, r) = k0 + k1*w1 + k2*w2 + ... + kn*wm + r1*s1 + r2*s2 + ... + rm*sm

Cool! There are new features to learn! However, these statistics themselves depend on words, e.g.:

s1 = s1(w) = f(w1, w2, ..., wm)

where f(w) is a statistics function of words in training document (e.g. average word length). f introduces some (possibly non-linear) transformation of features, which is cool by itself too. In fact, such transformations may be essential for this particular task. That's one possible reason why your statistics are rated as the most important features. However, your main model h() could already learnt most of these transformations in its own way, and that's why you've got only little improvement.

This looks like the most likely case to me, but of course there are other possibilities. Anyway, it's quite hard to make conclusions about highly dependent features. Try using more independent things like part of speech for each word (e.g. import_VERB, import_NOUN, etc.) or n-grams (e.g. not just number words "go" and "back" in text, but also number of bigrams "go_back"). This may change classification results dramatically (or not :)).