Slate keeps track of the user's selection in the editor in an immutable data store called a Selection. By doing this, it lets Slate manipulate the selection with transforms, but still update it in the DOM on render.
One of the constraints of the Slate document model is that leaf nodes are always text nodes. This constraint exists to make selection logic simpler.
A selection always refers to text nodes. Such that even if you set a selection relative to a non-text node, Slate will automatically correct it for you.
This makes selections easier to reason about, while still giving us the benefits of a recursive document tree, and it makes for a lot less boilerplate.
When a selection is used to compute a set of Block nodes, by convention those nodes are always the leaf-most Block nodes (ie. the lowest Block nodes in the tree at their location). This is important, because the nested document model allows for nested Block nodes.
This convention makes it much simpler to implement selection and transformation logic, since the user's actions are very often supposed to effect the leaf blocks.
Unline Block nodes, when a selection is used to compute a set of Inline nodes, the trunk-most nodes are used (ie. the highest Inline nodes in the tree at their location). This is done for the same reason, that most user actions are supposed to act at the highest level of inline nodes.