I know of 2 examples of Lisp being used to develop an OS (there may be more): one by Symbolics (called Genera) and the other by Lisp Machine Inc (aka LMI). Apparently both of these descended from code developed at MIT's AI Lab. Googling, it sounds like some of the source code for one of the versions was released as open source.
Of course, Lisp Machine Lisp (or Zeta Lisp, et al.) are not the same as Arc, but I expect that what you can do in one, you can do in the other. Small matter of programming :-)
The interesting question is not the implementation language, but what you intend to build.
I guess the quote idea came from me thinking that the strings were being evaluated. My untrained mind thought "strings evaluate to themselves", but apparently not. So I used the single quote to prevent evaluation.
Your use of hash tables to represent nodes in the AVL tree bothers me a lot. Yes, you get convenient notation, but it's disgusting, both practically and algorithmically, in terms of both space and time.
To see why, re-implement your stuff in something simple, like C or assembly, then measure space and time consumption. Compare using structs for nodes versus hash tables.
Just because it's wrapped up in pretty syntax doesn't make it efficient for use in all circumstances.
Might play with a C version of alists, versus hash tables, versus binary tree while you're at it. They're all useful, but in different circumstances. Implementing every "dictionary" with one or the other data structure is a mistake. You need to choose the correct data structure to match your application. It's an interesting engineering problem.
What I was doing was getting the AVL tree to work (which, by the way, took a certain amount of thinking, debugging, and rewriting). Implementing them with hash tables, with convenient notation, was entirely appropriate for the task at hand. And now that I've got it in working form, it's easy to go and replace the hash table parts with structs or something. In fact, I think that was pretty clearly on my mind, because I described in footnote 2, in some detail, how one might implement the nodes using things other than hash tables. It was not my intention that the nodes should forever remain hash tables. I think you misunderstand me greatly.
And by the way, I did re-implement it in C, using structs, before I wrote the original post. Here's an excerpt.
Feels good, man. (The above code, by the way, fails to free() the obsolete nodes. That could be fixed if necessary.) Fortunately, I didn't really have to debug that code; I cribbed directly from my Arc code.
The "constant numbers of fields" is the reason there's a better alternative.
A hash table has an _expected_ constant lookup time. But it might be as bad as O(number of fields). He blows all his O(log n) complexities by putting in an expected O(1) lookup for all his field manipulations.
It's probably premature optimization to use an AVL tree in place of an ordinary binary tree, or even an alist.
I didn't see him promoting AVL trees for all uses, but I did see him suggesting they could/should replace alists.
Of course I don't mind a guy learning about AVL trees, or anything else, by implementing them in Arc or anything else. Indeed, I recommend such exercises to everyone; we learn a lot. But my feedback regarding his implementation of the nodes stands.
Yeah, it is distracting and I'm sorry about that. Similarly, I'm sorry to publicly snipe, 'cause I really do admire folks who make an effort to learn by doing, and better yet, write up their experiences.
But "disgust" was the word I wanted. I can't point at his choice and say "this is all wrong". Instead, I point at it and claim that it "smells" wrong to me, I'd never do it that way, bad taste, etc. Then I tried to argue about why my taste is offended.
Regarding O(1) versus O(k), I actually wrote "expected O(1)". That is, _expected_ constant time, versus guaranteed constant time. Though your point that the number of fields is a constant, 4, is certainly true.
I didn't find it at all easy to learn what I wanted from the source code. I mean, imagine I'm sitting here trying to write my first program in Arc. I know I need a loop, say, but which one? I don't even know the names yet. So I search backwards and forwards through arc.arc (or was this one in ac.scm, I can't remember!), finding first one variety then another. And interpreting the meaning of the source when I'm still learning is slow going. Perhaps you've forgotten :-)
Searching the forum wasn't much better. Imagine searching for "cache". Short words are handy to type, but don't yield useful search terms.
So I decided to write things down where I could print them out in a single command, yielding a relatively compact description, and look them up conveniently.
I can definitely imagine better documentation easing the learning curve for new Arc programmers. Source code can be self-explanatory, but only after you've attained a certain base level of fluency with the language. Just as the axioms have to be defined in Racket  before Arc can work, they have to be defined in your head before it can make any sense. So I applaud your effort.
One resource that really helped me learn faster was rntz's help system, available in Anarki . It's very convenient because it gives you access to documentation right at the repl:
[fn] (acons x)
Determines if `x' is a `cons' cell or list.
Unlike 'alist, this function will return nil if given an empty list
See also [[atom]] [[alist]] [[dotted]] [[isa]] [[cons]] [[list]]
[fn] (alist x)
Return true if argument is a possibly empty list
Unlike 'acons, this function returns t when given an empty list
See also [[atom]] [[acons]] [[dotted]] [[isa]] [[cons]] [[list]]
[mac] (afn parms . body)
Creates a function which calls itself with the name `self'.
See also [[fn]] [[rfn]] [[aif]] [[awhen]] [[aand]]
Source code can be self-explanatory, but only after you've attained a certain base level of fluency with the language. Just as the axioms have to be defined in Racket before Arc can work, they have to be defined in your head before it can make any sense. So I applaud your effort.
Well put. My thoughts exactly. The suggestion wasn't necessarily to learn Arc by spending a couple days reading the source. I mean, that's what I did, but I was already familiar with Common Lisp. Just that once you reach a certain point (as Preston seems to have), it's not so daunting to say "hey, I wonder what filechars does?" and go read
(def filechars (name)
(w/infile s name (allchars s)))
The ambiguities of certain names in Arc really used to confuse me. acons is the cons predicate and alist the list predicate, but afn isn't the function predicate. Moreover, alist has nothing to do with alists (association lists).
Now that I have better command over the language, these things don't bother me so much. They have their own sense, and I increasingly appreciate how "externally exposed names [are] approximately hamming encoded." 
Offhand, I can think of 3 places where dots are used in Arc.
1) for improper lists (where the cdr of the final cons is not nil),
2) in parameter lists, to indicate that all remaining arguments are to be collected into a list and assigned to a single parameter, and
3) as a syntactic abbreviation.
I'm guessing you're referring to the 3rd use, since it's a departure from earlier dialects of Lisp.
When the compiler sees a sequence of non-whitespace characters, it has to decide what it represents. Might be something important like a parenthesis, a character constant, a string, or a number. If it's none of the above, then it's probably a symbol. But before it decides on symbol, it first checks to see if it's a syntactic abbreviation, looking for for one of the special characters: ".", ":", "&", "!", and "~". If any of those are present, then it's some sort of abbreviation and is handled specially; otherwise it's a symbol.
You can search in ac.scm for "ssyntax" to see how things are implemented. You can experiment using "ssexpand".
Usefulness... Saves typing, right? If you want to see the 5th element of a list, you could type fred.5 or (fred 5)
Seems natural for strings, tables, etc.
Same for ":". We all know the classical abbreviations like "cadr" and "cddr", but we can also gen up new ones like car:cdr:cdr:cdr:cdr which yields a function like caddddr that we can pass around in a 1st-class way.