After some discussion with several interested parties, we've decided to re-work Xdmf a bit. Most of the changes involve the API, particularly moving to smart pointers and a more full featured and customizable API for producing Xdmf in parallel.
The existing XML format will most likely stay the same with a few additions. So if you're currently producing Xdmf data outside the API this will have minimal impact. Also if you're just reading Xdmf as a file format this will have minimal impact as well.
Smart pointers in C++ will help eliminate memory leaks and double free problems. Most likely the implementation will be intrusive smart pointers with reference counting. It is expected that moving to smart pointers will have minimal effect on performance while greatly increasing the efficiency of resource management.
API changes/additions for sharing and producing data in parallel
While reading Xdmf is well defined and straight forward in the current API, producing and sharing the underlying geometric and computed data can become complex (particularly in a parallel computation). Since there can be several ways to represent the same data (a single large uniform grid, collections, etc.) it is best to allow the application to determine the proper representation and the most efficient method to achieve the result (generate the heavy data and collect the associated XML). What is needed is a more flexible mechanism.
Will Dicharry (Stellar Science Ltd Co) has proposed an API that looks very promising. If you'd like to take a look, the code is available at: 
Remove libxml2 dependencies from base objects - use containers
Currently, Xdmf is tightly tied to libxml2, actually using the library to store state information. This leads to instances where data in the Xdmf objects can become out of sync with the data in the XML representation in libxml2. A better approach is probably to let the Xdmf objects store all necessary data internally using C++ containers and decouple the generation of the XML using a visitor pattern. This would also allow for various other representations of the light data as they become necessary.
Better Fortran Support
Xdmf is currently 'C'-centric in that array dimensions are assumed to be row-major and data is accessed in that fashion. To properly use Xdmf with arrays in Fortran, which is column-major, the arrays need to be transposed on one end or the other. Transposing in memory may not work for huge arrays and transposing during I/O may incur significant overhead. At very least there should be transposing support with options.
Additional Topology Support
Additional Topology types
Static Geometry, Dynamic Attribute Support
If the Topology and/or Geometry does not change over time, it would be more efficient to cache that information. The current implementation assumes that every <Grid> is independent and re-reads the Topo/Geo. One idea to use XPointers in XML and mark the <Grid> with an XML attribute that tells Xdmf that it's possible to cache. Another idea is to cache the CData of the <DataItem> (from Dominik Szczerba) to flag identical heavy data items.
Active Grid an/or Attribute
Some mechanism to mark a Grid or an Attribute as "active" so, by default, only that one is read. Very efficient, particularly during debugging.
Much more testing, example code, and example datasets