Hi Tim,
I think Steven’s suggestion is the only feasible. Just remember that normal reconstruction is already half the resolution (a quarter of the pixels). If you are worried about the small details, then there is literally nothing you can do because no software can do magic. If you say downsample 3 (I think it means half of half, if I’m not mistaken), then there is no point trying preview, which can be set to “Use sparse point cloud = false” so that it actually calculates depth maps instead of using the tie points…