P: Blur/Sharpen Adjustment Layers

You can do that with a smart object because it has an explicit update step, and you don't try to apply the filter "live" with each change to the child object. Without the explicit update step, the performance is horrid (we tried it once, and, well, there isn't enough profanity to describe how bad it was).

Think about this: when you change one pixel under an adjustment layer, we have to update just that one pixel to recomposite the image. And that count doesn't change even with a large number of adjustment layers.
When you change one pixel under a radius 2.0 gaussian blur, we have to update about 300 pixels to recomposite the image. And that number grows geometrically with each added filter and it's support area -- and it would get insane with filters that do large warping (radial blur, distort, twirl, etc.). You could literally change one pixel and have to re-run every filter on the entire image area.

Why shouldn't I be able to non-destructively blur a shape, as though it's a smart object? It would be great if I could apply effects to shapes/images etc without having to convert them into a smart object

Probably never going to happen - the performance impact is much bigger than you think, and gets worse with multiple "filter layers".
Adjustment layers only work well because they don't have a support region (area) to adjust -- they can work on a single pixel at a time.

This is a tough nut to crack (performance, usability wise):

The Secret Life of Smart Filters