connection with the stored data.  This is handy when multiple
threads or processes are accessing and updating data simultaneously
(e.g. When using the monitor client to poke at a running Zope process).

This at least makes errors visible during debugging,
but I need to figure out a better way to handle these.

I keep forgetting that PyDict_GetItem has it's own special
refcount rules.

dropped from the cache.

ExtensionClass deletion protocol.

Added DECREF for object id, which is now an object.

destination version to fail to detect the new data.

invalidated objects.

Waaa.  I'll probably need to add a separate lock file.