It's true someone snooping at either end of a conversation could over time correlate timing and sizes to show that two users are communicating, but that's the most they can do. Signal is not peer to peer so you're not connecting to your recipient, and signal itself has enough raw volume that simply correlating sizes and timing of a small number of messages wouldn't really be sufficient to know who is communicating with who.

I think they could make that significantly more difficult by adding csprng delays and padding to the messages. That way you can't really effectively correlate timing and sizes without direct access to signals inner workings. I'm not sure what signal's actual throughput is, but if think as a paid feature it could be economical.

Another crazier way would be to send every message to a large number random latched recipients. Good way to 1000x your bandwidth.

> The metadata is as valuable as the data.

This can be true if you are able to get ahold of a user's device and access their signal messages. It's not true in most other cases. I don't particularly care if you know that I am talking to someone specific as much as I care that you don't know what I'm saying.

> the surveillance networks can capture metadata - who talks to who and when

If this is part of your threat model then I would suggest a different tool such as SimpleX since it uses onion routing and can be configured to always use private routing/relays.