You mentioned you have a parallel document that implies you do, in
fact, know the structure of the documents.

Not at design time. The structure must be infered by the XSLT script
from the XML Schema of the default and output document. So a key table
also should be constructed from the schema by the script.

I am using XSLT to walk through an XML Schema to construct an Xforms
(output) instance.
In parallel I am scanning an (input) instance (of a document defined
by the schema) to
include default values in the form.

Would this not reassure you that the default instance document has
the same document structure as the instance you are creating?

Yes they both have the same structure, although the Default document
 does not need to contain all elements.

Nevertheless you can modify my suggested key approach to use any
lookup value for the default values and then access those default
values from the loaded key table.