Apparently, wl_layer_shell does not set size based on pixels, but on rescaled units. That caused a mismatch where I expected it to be on the same abstraction level as wl_output. Not seeing a reason why it should prevent itself from half+ the available pixel sizes, yet here we are.