I'm a newbie at this stuff, but here goes:
"single-rendering-pass order-independent transparency" - let's say I have three translucent objects at roughly the same depth, with parts of one in front of and behind parts of the others (and maybe the same is true for objects B and C as well). Figuring out the correct draw order is absolute fucking murder, and there still isn't a generalized approach for anybody but the most advanced of the most advanced (like Dual depth peeling or making convex hulls out of all translucent geo in the scene). Core API support for dealing with this issue would be a godsend and is about 10 years overdue for ALL graphics APIs.
Neat fact: the PowerVR-based GPU used by the iPhone/iPad uses a tile-based rendering method in which (I am told) this problem generally doesn't arise.
"Capturing GPU-tessellated geometry and drawing multiple instances of the result of a transform feedback to enable complex objects to be efficiently repositioned and replicated;" Easier to quickly render massive crowds, forests, and procedural cities.
"Modifying an arbitrary subset of a compressed texture, without having to re-download the whole texture to the GPU for significant performance improvements;" Shaders not requiring four fucking separate mask textures all dancing on the head of a pin to pull off a simple effect? Yeah, I'll take that. Could probably also have some nice gains in procedural content variation.
"Packing multiple 8 and 16 bit values into a single 32-bit value for efficient shader processing with significantly reduced memory storage and bandwidth, especially useful when transferring data between shader stages." Massive performance gains for any sort of post-processing work, basically.