Exactly. There is a flattened RGBA image at the end of a PSD file, which has dimensions corresponding to dimensions of a project.

Reading that flattened image from PSD takes 20 lines of code, while reading the structured content and producing final composition (blend modes, layer styles, masks, adjustment layers ...) takes at least 20 000 lines of code.

Row length encoding is used for storing channel data by default, so area with a constant color can be compressed much more efficiently, than area which has more "noisy" content.

If you're using psd filesize probably isn't your first priority.

I believe that's because it saves bytes when generating a flattened preview image which for high resolution PSDs can be quite expensive to do losslessly. Can anybody confirm?

This is likely what is going on. PSDs have a flattened, full-size version of the image embedded for compatibility's sake.

Instead of a white top-layer you could also disable the "Maximize compatibility" option when saving the PSD, which will strip out this full size rendition.

If you disable Maximize Compatibility you also lose compatibility with current programs from Adobe itself, such as Lightroom. So while it seems like a good thing to turn off old crufty stuff, it actually is horrible.

I didn't know that Lightroom had that restriction, actually. Is that tested with CC 2015 versions of both programs?

I do know that Maximize Compatibility doesn't affect usage within Illustrator and InDesign documents, though, which to me are the more common use cases for a PSD anyway, so I'll probably stick to unticking that box.

Or putting all layers inside a not visible group. Nothing to save (or it's very well compressed) in the flatenned preview. In large files it's a big save