This was discovered by Keping while working on Rocchio. As a workaround, she sets "flattenCombine=false" in her query parameters, but this traversal should maybe give up if it hits negative weights or norm=false, since it applies a norm as it goes (usually fine, since norming will be ranksafe) but explicitly not what the user wants when the have:
#combine(a b #combine:norm=false(c d))
It will take the weights on c and d and norm them before sticking them into the upper-level combine. Basically, if a child has a norm=false, it should give up.
Small change, but will need tests.