Consider the dateTime of the last leap second:

1998-12-31T23:59:60Z (P)

This instant in time can also have the lexical representation of, for example,

1998-12-31T22:59:60-01:00 (Q)

Section 3.2.7.3 defines the algoritm for comparing two dateTimes as follows:

"A.Normalize P and Q. That is, if there is a timezone present, but it is not Z, 
convert it to Z using the addition operation defined in Adding durations to 
dateTimes (E)"

Now in our example P has a Z timezone, but Q doesn't, so we need to normalize Q 
to Z using Appendix E. But E.1 says:

"Leap seconds are handled by the computation by treating them as overflows. 
Essentially, a value of 60 seconds in S is treated as if it were a duration of 
60 seconds added to S (with a zero seconds field). All calculations thereafter 
use 60 seconds per minute."

This implies that Q is first mapped into:

1998-12-31T23:00:00-01:00

Then following the rest of algorithm in Appendix E, this will map into: 

1999-01-01T00:00:00Z 

Now comparing: 

1998-12-31T23:59:60Z 

and 

1999-01-01T00:00:00Z 

we find that 

P < Q 

But P and Q represent the same value. So we have a contradiction: two lexical 
representations represent the same value, but the value represented by one 
lexical representation is less than the value represented by the other lexical 
representation.

See:
http://lists.w3.org/Archives/Public/www-xml-schema-comments/2002AprJun/0043.html