I think I misunderstood what becomes the context node. Take the first
one I mentioned:

id('bodyContent')/h2[1]/following-sibling::*[position()<count(following::p[1]/preceding-sibling::*)
- count(preceding-sibling::*)]

I thought that while the position() filter was running, 'h2[1]' would
be the context node. Instead, it seems the result of
'following-sibling::*' is the context node? What, does it change the
context node to each item in that set as it applies the filter? I
don't suppose there's a way to refer to the previous context node in a
relative way, is there?

following has context node * for position()
a/b/c/*[position() = 2]

following has context node c for position()
a/b/c[position() = 2]/*