Framebuffer attachments should always be matched with the same image:

RenderGraph::new()
    .begin_pass("My Graphic Pass")
    .bind_pipeline(&my_graphic_pipeline)
    .clear_color(0, swapchain_image)
    .store_color(0, swapchain_image)

And in the case of merged render passes they will always having matching formats and sample counts because Attachment::are_identical() checks that they are in fact the same image.

With this logic it is safe to set the format, sample count, and initial layout on framebuffer load operations and then also set the format and sample count in addition to final layout on store operations. In cases where the user does not specify a load or store operation we can infer the format. This fixes the issue of depth images having the wrong layout due to it being set based on the load operation but ignoring the store.

In cases where the user simply attaches images (the DONT_CARE ops) there is no change.