I just received the 4th edition (it really is hardbound!)
so I could check if anything had changed, but on p. 602 :

 There are three kinds of expressions in XPath 2.0 whose result
 is guaranteed to be a sequence of nodes in document order...

  o Any axis step (even and axis step like <preceding-sibling::*> that
    uses a reverse axis delivers its results in forwards document order)

 In effect, the predicate operator <[]> has a higher precedence (it
 binds more tightly) than the path operator </>

That's a bit subtle, but OK, now I know, XPath has an evaluation
precedence to consider.

I believe the issue is (and I would like some clarification), is that
axis steps always return nodes in document order so

preceding-sibling::w:p[1]

will always return the first paragraph with its TableCaption child.

I believe the work around for this, to collect the nodes in search
order, is to make them a set,

(preceding-sibling::w:p)[1]

No this is wrong. preceding-sibling::w:p[1] selects the most recent w:p
element (the last in document order), whereas
(preceding-sibling::w:p)[1] selects the first in document order.

Michael Kay
Saxonica

--
Louis (Hank) Ratzesberger
sopac.ucsd.edu