Bignums "contaminate" surrounding operations. Adding a long and a bignum yields a bignum. Seeding an equation with a single bignum (42N is a bignum representation of 42) will prevent overflow.
No one has come forward with a single real-world scenario where they're using bignums in Clojure. Choosing a default that is only theoretically useful rather than a 10x performance improvement seems a bit silly.
In any situation where the compiler cannot be sure you're using primitives everywhere, it will emit non-overflowing bytecode.
By my measure, there's nothing premature about this optimization. In Rich Hickey's words, Clojure is a replacement for Java, not Ruby. Giving up Java-like performance makes the language quantitatively less useful.
That Clojure's target is to replace slow old Ruby (actually that never happens, rather a successful language gains dominance in some niche and spreads from there) is only more evidence that it isn't worth breaking the abstraction called "numbers" to get merely an order of magnitude better performance. Seriously, I almost can't believe this is still an open issue in 2010. And Rich Hickey, who has designed such a great language otherwise, is on the wrong side of it? I'm boggled.
EDIT: If peregrine's comment is right, then we're arguing about something that isn't happening. I hope so.
Premature? No people have been complaining (or just getting very confused) about Clojure's numeric performance without oodles of invasive type-hinting for two years now.
More importantly with these changes it's now possible to reimplement Clojure's core datastructures in Clojure without sacrificing performance.
The majority of people who care about BigInts are solving Project Euler problems, not writing Clojure libraries or deploying apps into production.
The situations that deal with numbers that large are usually a bit nuanced. (i.e., "it's 1 in 2^63 of the time that I need bignums for day-to-day programming"). I agree that losing accuracy needlessly sucks (and I've been burned by floating-point), but 64 bit ints are usually good enough.
Correct, promotion will no longer be automatic because the semantics of primitive and auto-promoting arithmetic can't be unified (on the JVM with no fixnums, the return types are disjoint). However, missing from the blog post and discussion here is the fact that, should there be overflow, an exception will be thrown. We are not adopting the silent error strategy of C/Java etc.
Oh no! Clojure is choosing the route of premature opt. What a terrible shame.