Think of it as a DSL, statisticians like to use. While it has its dark sides, its vector-orientation is well suited for a language whose main purpose is data manipulation.
I can't let this stand as a blanket defense of crappy languages. We have Lua. We have Python. We have Scheme. We have multiple other languages that can be used as a good way to access the goodies provided by your wonderful hand-crafted libraries regardless of what they do. There is no reason for you to invent your own language just so people will be able to use your library code from a REPL.
Sadly, back in the dark ages, this was not true. However, in the future, anyone who tries to defend the horrible design of a new language with "Think of it as a DSL" gets to debug a 1000 KLOC application written in ANS COBOL 1968, which is a DSL for fixed-field database munging.
Let's put things in perspective: R is S and S had it's first appearance in the 1970s. It's current reincarnation is from the late 1980s. I remember that other statistical packages back then had a similar looking syntax. This is also the reason why lua/python stand no chance against R: they simply lack the vast abundance of statistical packages and tutorials. The best thing they can do is use R as an inferior process (IIRC there is a package for incanter that does that).
The point is though that S/R isn't that badly suited for what it was created for. There are a few pitfalls but those are explained in the langage definition. But unfortunately nobody RTFM nowadays (like people used to do back then in what you call the dark ages) because it's easier to start screaming for Mommy and write stupid blog posts that prove nothing but that those people haven't read the language definition.
