When Arc compiles (a:b c), it compiles ((compose a b) c). When it compiles ((compose a b) c), it compiles (a (b c)). So in Arc, (do:a b) and (do (a b)) are basically equivalent, and so are (a:do b) and (a (do b)), where 'do is just an example macro. Since 'do is a macro, the equivalence does involve semantics; your transformation of (a:do b) calls a function named "do", when that function may not actually be defined.
Anyway, I'm just trying to make sure I inform you of stuff you may have overlooked. Once you know what I'm talking about regarding 'compose, it's up to you to decide whether you actually want to preserve an (a:b c)-(a (b c)) equivalence. ^_^ I am quite a fan of it, 'cause it saves some nontrivial edits (adding a closing parenthesis in a distant place and reindenting), but it's not sacred or anything.
Ah, I think I understand now. I had seen this comment from arc.arc:
; Composes in functional position are transformed away by ac.
And your examples show how 'compose gets transformed away, but I was having trouble visualizing a case where that wouldn't happen. Now it seems obvious to me: if you have (compose a b), rather than ((compose a b) c), then you can't transform compose away, because something is needed to make a and b act like one function.
