Currently Lux does not keep track of the materials the light has passed through (i.e. it only takes the current material into account). So to get the correct result you have to use "to_medium_ior"/"from_medium_ior" (where from is the side the normal points to) for the ior.
As this is obviously not the most natural way to describe a material, we should probably best change to normal behavior (i.e. each material just has its own ior and lux keeps track of the from medium, combined with some precedence-like system) sometime in the future.
So in the case where you have bubbles in water, you just model these bubbles "inside" the water (i.e. the water contains no "holes"), have all normals point outside and use for materials:
water mat-> "glass" with water ior (1.33 or so)
bubbles mat-> "glass" with air ior / water ior (1.0 / 1.33 = 0.75 or so)