The problem is not that it cannot be done in principle -- calibre's current metadata update code already does it (albeit for record 0 not n).
The problem is that there are lots and lots of corner cases, and the only way to test those corner cases is to laboriously upload a file to two actual kindles (one that uses the old mobi and one that uses the new) and see if it works. That makes the effort required to develop a robust solution too large for the gain, at least for me. And I'm afraid that 8192 bytes is not nearly sufficient, some people have rather a lot of metadata, I even remember one user that used put the entire book text into the comments field
