e.g.
+a----------------------+
|   +b---------+        |
|   |          |        |
|   |          |        |
|   |          |        |
|   |       +c-|------+ |    
|   |       |1 |      | |
|   +----------+      | |
|           |     2   | |
|           |         | |
|           +---------+ |
+-----------------------+
a is the parent of b. c is a sibling of a; it's under b and above a.
The hit-test data we have looks like:
root
  a
   b
  c
We can handle point 1 correctly because of z-order in hit-test data and we stop looking further after we found the target. However, for point 2, it's going to match window a instead of c - we go to a, then b (b doesn't contain it so we go back to a which contains it) and stops at a.