I've written a rough draft for a post import hook as discussed on the
python 3000 mailing list. The implementation is far from perfect. It
requires more unit tests, a code review and reference count checks.

It also needs to hold the import lock during both the "register" and 
"notify" operations.  In addition, the "notify" operation needs to be 
exposed for calling from Python (so that lazy module implementations 
can interop).  Finally, it's not clear to me whether there's any way 
one of the import APIs can exit with the module added to sys.modules, 
but *without* the notify operation being called.  It seems to me that 
those code paths were cleared up in 2.4 (i.e. failure to import 
causes the module to be removed from sys.modules), but it might be 
good to doublecheck that.

I've moved the result = PyImport_NotifyPostImport(result); inside the
protected block. It's now protected by the import lock. I've also added
the lock protection to the register function.

The notify method is now exposed through the imp module, too.

I've checked multiple code paths. They all end up in 
PyImport_ImportModuleLevel(): builtins.__import__(), PyImport_Import(),
PyImport_ImportModule(). Only PyImport_ImportFrozenModule() doesn't use
the code path in imp_init_frozen().

The new patch fixes some ref leaks, corner cases and adds some new unit
tests. All unit tests are passing but I'm leaking lots of references in
the register function.

UPDATES:

* no ref leaks
* PyModule_IsLazy(mod) checks mod.__lazy_import__ attribute (we can
argue about a nice name later)
* Added imp.is_lazy(mod_or_name)

Updates:
* Some minor cleanups
* First draft of lazy modules implemented in C. It works but it needs
some cleanup and more error checking.

The new patch applies against the latest svn revision. Several functions
were renamed.

The post import hook patch is in my pep 369 branch.