I think S. Sierra's approach is fundamentally superior what I offered. I have two suggestions: one slight perspective change (which implies no change in the actual implementation), and an idea for an even simpler approach (at least from a user perspective), in that order.
While interop is the driving requirement behind AOT, I absolutely do not want to have to keep an updated enumeration of all of the classes resulting from each and every defrecord et al. usages in my pom.xml/project.clj (and I wouldn't wish the task of ferreting those usages and their resulting classnames on any build tool author).
Right now, *compile-write-classes* is documented to be a set of classname strings, but could just as easily be any other function. *compile-write-classes* should be documented to accept any predicate function (renamed to e.g. *compile-write-class?*?). There's no reason why it shouldn't be bound to, e.g. #(re-matches #"foo\.bar\.[\w_]+$" %) if I know that all my records are defined in the foo.bar namespace.
To go along with that, I think some package/classname-globbing utilities along with corresponding options to clojure.lang.Compile would be most welcome. Classname munging rules are not exactly obvious, and it'd be good to make things a little easier for users in this regard.
If there's a closed set of forms that generate classes that one might reasonably be interested in having in a build result (outside of use cases for pervasive AOT), then why not have a simple option that only those forms utilize? gen-class and gen-interface already do this, but reusing the all-or-nothing *compile-files* binding; if they keyed off of a binding that implied a diminished scope (e.g. *compile-interop-forms* – which would be true if *compile-files* were true), then they'd do exactly what we wanted. Extending this approach to deftype (and therefore defrecord) should be straightforward.
An implementation of this would probably be somewhat more complicated than S. Sierra's patch, though not as complex as my original stab at the problem (i.e. no *load-level*). On the plus side:
1. No additional configuration for users or implementation work for build tool authors, aside from the addition of the boolean diminished-scope AOT option
2. Class file generation would remain opaque from a build process standpoint
3. Future/other class-generating forms (there are a few people futzing with ASM independently, etc) can make local decisions about whether or not to participate in interop-centric classfile generation. This might be particularly helpful if a given form emits multiple classes, making the determination of a classname-based filter fn less straightforward.
I can see wanting to further restrict AOT to specific classnames in certain circumstances, in which case the above and S. Sierra's patch might be complimentary.