Reproducible on a Mac debug build.
Reproducible on linux with the stable release from Oct.

Downloaded the bert Freesurfer test data set[1], made a scene as described.

When saving the scene, rh.pial and lh.pial are missing in the node list, but Slicer saved links to them in the .mrml file, and the display node has the active scalar set, and it has an overlay storage node.

When load the scene:
RROR: In /Users/nicole/Slicer4-svn/Slicer/Libs/MRML/Core/vtkMRMLFreeSurferModelOverlayStorageNode.cxx, line 79
vtkMRMLFreeSurferModelOverlayStorageNode (0x7fc3386a4550): ReadData: the model node doesn't have poly data yet, resetting the overlay read state to pending, try again later.

ERROR: In /Users/nicole/Slicer4-svn/Slicer/Libs/MRML/Core/vtkMRMLFreeSurferModelOverlayStorageNode.cxx, line 79
vtkMRMLFreeSurferModelOverlayStorageNode (0x7fc3386a84d0): ReadData: the model node doesn't have poly data yet, resetting the overlay read state to pending, try again later.

lh model has file name lh.thickness

[1] http://surfer.nmr.mgh.harvard.edu/fswiki/FsTutorial/Data

There are a few layers of bugs here, the first one appears when loading the scalar overlay:

After loading lh.pial, the model node has one storage node, the model storage node.

When you select the scalar file (lh.thickness) to load onto the surface lh.pial, the scalar overlay reader invokes the models logic AddScalar. That calls AddAndObserveStorageNodeID before it calls read data on the scalar overylay storage node. At this point stepping through the code there are two storage node references on the storable node and all is correct. But then ReadData at the storage node superclass calls the ReadDataInternal and then calls SetAndObserveStorageNodeID which over writes the model storage node reference with the scalar overlay storage node id.

Changing the vtkMRMLStorageNode to use AddAndObserve instead of SetAndObserve, that goes through the newer reference management system which determines how many storage node references there are and adds to the end, witout checking if the currently passed storage node id exists in the list already. So just reading in the lh.pial surface ends up with 2 storage node ids, both vtkMRMLFreeSurferModelStorageNode1 (due to the combo of Add in the vtkMRMLStorageNode and Set in ModelsLogic AddModel(filename)). Adding the scalar overlay after that results in four storage node ids, two of the model storage node, two of the overlay storge node.

One solution: not add the observation in the models logic AddScalar but make sure that the storage node does use add, and make sure that AddAndObserve checks for the id existing already. Adding a check on HasNodeReferenceID in AddAndObserve and returning null if it's found gets things further and exposes the save bugs.

A question directed to people who have a better grasp on the new reference management system: can HadNodeReferenceID return -1 on failure, and the Nth value on success that we can pass to UpdateNthNodeReference? Does it even need to be updated at that point?

I need to do further testing on if changing the Set to an Add in the storage node with the duplicate check will have any side effects, see this commit:
https://github.com/naucoin/Slicer/commit/256c78cc5c5568c6869071d617c9262c950ce1fa

When saving the correctly loaded scene, the FS model is set to save as a .vtk file but there are some missing pieces:

• the model node isn't flagged as modified since read, that needs to be done when adding a scalar overlay
• the storage nodes aren't updated for a regular model node (the .vtk will contain the scalar overlay information), I need to remove the FS scalar overlay node information for save (maybe flag them as not to save with scene?), make sure the .mrml file only has one storage node reference.
  It might be worth doing this on load since we don't have any FS surface writers.

FreeSurfer lh.* test data uploaded to:
http://slicer.kitware.com/midas3/folder/2718

Fixed in r26092.