Performance matters and is sufficiently captured via working incremental updates. The single largest upside of a proposal like this is captured by using SQLAR over ZIP. That's what the Library of Congress does when SQLite claims them as a proponent. It's what Fossil does as others in this thread have pointed out. It's suggested as "first improvement" in the linked article. It's also the only part that should actually be considered for implementation.

You are right to point out the folly of deeper implementations like having and needing to understand table structures for things like slides. However, the current status quo involves Microsoft implementing a fairly esoteric "update the XML file's bytes as they would be encoded in a ZIP file" in their proprietary tool (where they have enough money to invest the engineering time) and all other tools use the slower "whole file in memory" approach.

User visible features like incremental fast saves (and shared editing) keep people on closed systems and give Microsoft the leverage to do the things you warn against. SQLite as a container format could have prevented that by giving everyone a shot at a lower cost but still fast implementation.