Overzealous garbage collection(?)

Not-so-minimal example (Lua 5.3):

local Gtk = require('lgi').Gtk
local GObject = require('lgi').GObject

local Foo = require('lgi').package('Foo')

local CrashModel = Foo:class('CrashModel', GObject.Object, { Gtk.TreeModel })
do local CrashModel = CrashModel._override['Gtk.TreeModel']
    function CrashModel:get_column_type(col)
        if col == 0 then
            return GObject.Type.STRING
        end
    end

    function CrashModel:get_iter(path)
        local i, i2 = table.unpack(path:get_indices())
        if i2 then
            return nil
        end
        return Gtk.TreeIter { stamp = i }
    end

    function CrashModel:iter_n_children(iter)
        if iter == nil then
            return 500
        end
        return 0
    end

    function CrashModel:get_value(iter, col)
        if col == 0 then
            return GObject.Value(GObject.Type.STRING, "Blah")
        end
    end

    function CrashModel:get_path(iter)
        return Gtk.TreePath.new_from_indices { iter.stamp }
    end

    function CrashModel:iter_next(iter, ...)
        if (iter.stamp + 1) >= 500 then
            return false
        end
        iter.stamp = iter.stamp + 1
        return true
    end
end

collectgarbage("stop") -- Thanks to this line, this code leaks memory. If I remove it, the program crashes instead.

local iv = Gtk.IconView()
iv:set_model(CrashModel())
iv:set_markup_column(0)

local wind = Gtk.Window()
wind:add(iv)
wind:show_all()
wind.on_delete_event = Gtk.main_quit

Gtk.main()

I did some investigating, and it turns out ffi_closure_free is called on a closure which is still referenced inside a GtkTreeModelIface structure used by Gtk.

Start by reproducing the Lua 5.3 example and tracing the lifetime of the closure referenced by the GtkTreeModelIface structure. Investigate the ffi_closure_free call and compare behavior with garbage collection stopped; done means the example no longer leaks or crashes during Gtk.main().