While debugging some Shell code, I found that one of our enum
members was wrong in introspection land. The enum in question
looked like:

    typedef enum {
      SHELL_KEYBINDING_MODE_NONE          = 0,
      SHELL_KEYBINDING_MODE_NORMAL        = 1 << 0,
      SHELL_KEYBINDING_MODE_OVERVIEW      = 1 << 1,
      /* ... */
    
      SHELL_KEYBINDING_MODE_ALL = ~SHELL_KEYBINDING_MODE_NONE,
    } ShellKeyBindingMode;

Looking into it, I found that the scanner was dealing with
symbolic references in enum members badly, overwriting previous
values if they existed. I don't know how memory is managed in
the source scanner (but we seem to create new symbols and reuse
old ones without refing in the same codepath, so I think it's
already broken), so this patch may leak.