People who do bignum application perhaps shouldn't use a scripting language like newLISP. Those bignum applications would probably involve heavy number-crunching which is more suited to do in a compiled language.
The best CL for bignums is actually CLISP, surprisingly. It is a byte-code interpreter, but it uses extremely fast algorithms. CMUCL does better with FP, though. But CLISP has arbitrary-precision floats, which is unusual (you have to toggle something to get them though).
Perhaps a nice compromise would be to have bignum library written in 'C' and import it into newLISP which is more or less what David is suggesting. This library then could have other math stuff too, used by people dealing with bignums.
Then it won't automatically switch between fixnums and bignums. This is extremely convenient. Most other languages are moving towards this feature, not away.