So suppose an attacker sees two packets:
Encrypted("Indian for dinner?"): Their username, your address.
Encrypted("Sure, sounds good."): Your username, their address.
From this they could be reasonably sure you two are talking but it's less data than they had before and now that attacker needs to either know you're using the source address or see both messages to get the full picture.
Suppose Alice is asking about Indian and Bob replies that it sounds good.
A passive attacker sees TCP/IP packets between Alice's IP address and Signal's and between Bob's IP address and Signal's. They get no other information from this beyond that Alice and Bob use Signal (and perhaps not even that if Signal shares IP addresses with other services)
The Sealed Sender feature makes no difference in that layer
If an attacker has control of Signal's servers (or perhaps Signal's server admins are secretly bad guys) ordinarily they would be able to see the sender and recipient of every message.
With Sealed Sender, the servers don't know the Sender any more, only that the Sender seems to be someone permitted to send messages to this recipient.
You could try to correlate IP addresses to Signal users, but you have absolutely no guarantee that they're correlated, much less that there's a nice 1:1 correspondence.
In a mundane example, Alice proposes Indian food then leaves her office, her phone disconnects from WiFi and goes to a 3G network. The reply from Bob is picked up by Alice using a completely different IP address, because she isn't in the office any more.
It's also odd that the tokens are 96 bits. Very odd, really.
(This is my imagination and it might be completely wrong so please feel free to correct me)
they mentioned spam problem which leads me to believe that if enabled, the sending client will encrypt the whole message using the recipient's public key and put all metadata other than recipient's identifier inside this bigger encrypted envelope. The receiver opens this envelope with its key and opens the smaller box inside which contains the sender's metadata.
What we lose now with this is the server does not have much insight into who is sending messages (by design). This means if you allow sealed sender from everyone, someone could send you a lot of messages which you may not like.
Yes, their servers can see your IP but if they discard it immediately you're safer if the NSA or another government agency reaches out to them.
I think it prevents their servers from correlating my identity and my IP address etc., but since I want replies and I'm asking the server about replies, doesn't that operation tell the server what my identity is anyway?
(There are some comments here talking about anonymous messages, but that doesn't sound right since the phone number is apparently kept in the encrypted, inner envelope, and also how would you route replies if you didn't have an identity of some sort for the sender?)