Hi Eric and Samuel,

As I understand it, the problem is that the undo history gets scrambled
by the interleaving of user edits (in the indirect source-editing
buffer) and automatic changes introduced by org (un- and re-indenting
the source code).

I have the following patch, which seems to prevent the misbehavior
Samuel noticed.  It has the drawback of not keeping the fine-grained
undo information: after org-src-edit-exit, all changes made during the
edit are seen as only one change, and undone as a unit.

I think the problem of interleaving the automatic and user-driven
changes in a sensible way is tricky.  We don’t want the first invocation
of undo after org-src-edit-exit to remove the contents of the code
block, which is what a naive approach gives (since org-src-edit-exit
deletes then reinserts the code block contents).

I’ve been running with this patch for a while and not noticed any ill
effects.  But I haven’t made a concerted attempt to test undo around
code blocks, which is why I’ve held off on pushing it.  If it fixes
Samuel’s problem and looks good, perhaps it is ready to go.