It turns out one silly causing slab slowness is just stock desktop thrashing:
calling gmenu_tree_get_root_directory does a huge thrash of work: some N^2 in the number of .desktop files:

It calls 'process_layout' once, which calls 'menu_layout_node_menu_get_app_dirs' 150 times - ~always with the same data; this patch adds a 1 item cache.

My callgrind numbers go from:

3.35 billion (!) instructions beforehand to:
3.08 billion (!) afterwards

only an 8% win it's true, but every cycle helps here I guess :-)