Pixels with alpha always seem to use the clamped light model. This includes textures with 1 bit alpha, as they will acquire some alpha after going after bilinear filtering. For these textures we only discard them if they are less than a specific threshold, but that leaves many opaque pixels that are technically opaque but with some alpha. This fixes some corner cases with the lighting in daytona.
https://github.com/trzy/Supermodel/commit/44280ae320a1ad4e1b8f72d714e97582d2b52989