You definitely have some good points - I agree on points (i) and (ii) - these are shortfalls of the language and in many cases (especially when building production level code) may be reasons to switch).

I disagree with you premise that R is not popular among machine learners and the implication that has for the "bleeding edge" comment. Perhaps it is less popular with ML people coming from a CS background - but it seems to be the most widespread language for ML (or "statistical learning") among statistics people. And that fact is not to be discounted - perhaps the preeminent book on machine learning (or at the very least one of the most popular) - "The Elements of Statistical Learning" - is written by statisticians (and in fact uses R exclusively!). The Journal of Statistical software, which features papers detailing many ML libraries, has far larger coverage of R packages than any other language: http://www.jstatsoft.org/ . I would suggest this is the evidence for the "bleeding edge" comment. Other languages' packages, say Python's ML libraries (scikit-learn, pyBrain, etc.), do note even come close to the breadth of capability in this space.

Coming from a statistics background,- I would venture to ask.. what's wrong with "pushbutton" libraries? And why are they only useful to people who do not know/understand much about ML/statistics?

Say you have a dataset and you believe ,say, a Random Forest would be well suited to predicting some response. Let's assume you have a very good understanding of Random Forests. Why would you not want a push-button library? Why would you WANT to recode the thing yourself? It's not as if your Random Forest will be any better or "more correct" than the one on CRAN or whatever language's repository you are using. I would argue it's more likely to have mistakes (coding something like this from scratch is no small task and the ones on CRAN have often gone through many iterations, improvements, and reviews from many experts in the field). And let's say you have some domain knowledge or some informed belief about how the Random Forest needs to be adapted to your particular problem (a different loss function or something) - you can easily edit the R package source code to do that - no need to rebuild the car just to give it a new paint job...