Matrix evolves as a protocol by folks proposing new APIs (e.g. MSC3575), implementing them, showing that they work, and making the case that they should be merged into the main spec.

Sliding Sync is an interesting case in that it clearly does work, and it kicks ass - but it turns out to be unnecessarily complicated for the subset of features that you actually need when implementing something like Element X (which is my fault entirely, fwiw). However, should we hold off on putting it in people's hands while we sort that? Nope. Is it a pain to temporarily run a proxy shim service while the protocol stabilises? Yes. It's like running a SPDY or QUIC-aware reverse proxy in front of a webapp that only speaks HTTP/1.1 while waiting for SPDY to be ratified as HTTP/2. Sure, it's another moving part to look after, but the spectacular improvement is likely worth the pain. YMMV though :)