Logged In: YES 
user_id=31435

Thanks!  I checked in some code building on this.  Changes 
included:

+ Adjusted whitespace to meet the standard (spaces 
after "if" and "for", flanking binary operators, etc).

+ The refcount fiddling in x_mul caused assorted system 
crashes if KeyboardInterrupt was raised during a multiply.  
Repaired that.

+ More comments and asserts.

+ Removed k_join and built "the answer" piecemeal into the 
result object in k_mul.  This allows to free more chunks of 
memory sooner, reducing highwater mark and the probable 
size of the working set.

Since the cache behavior is quite different now, it would be 
cool if you could run your tuning tests again.  The cutoff 
value is now a #define, KARATSUBA_CUTOFF near the top 
of longobject.c.

Until I can make time for more thorough testing, k_mul isn't 
called by default:  multiplication invokes k_mul if and only if 
an environment variable named KARAT exists (its value is 
irrelevant; just its existence matters).