If you think WhatsApp leaves a lot of metadata on the table for analysis, try doing a Matrix chat. You get a plaintext view of which device used which key to send which message ID to which room/person. If the message is a reply, you get the message ID your new message is a reply to in plaintext as well.

Without even looking at things like HTTP headers, this is what the metadata an E2EE-encrypted message (with verified+cross-signed keys) looks like, with specific identifiers censored just in case:

    {
      "type": "m.room.encrypted",
      "sender": "@.......:jeroenhd.nl",
      "content": {
        "algorithm": "m.megolm.v1.aes-sha2",
        "ciphertext": "AwgAEqAC/..........",
        "device_id": "EDNM......",
        "sender_key": "+rKR.......",
        "session_id": "H3Oyob........",
        "m.relates_to": {
          "m.in_reply_to": {
            "event_id": "$5qFg........"
          }
        }
      },
      "origin_server_ts": 17507.......,
      "unsigned": {
        "membership": "join",
        "age": 127,
        "transaction_id": "m17507........."
      },
      "event_id": "$_KBk.......",
      "room_id": "!.........:jeroenhd.nl"
    }

Unlike on platforms like Whatsapp, these message envelopes are available to anyone with access to either a session token or the user's password. The E2EE keys require a bit of extra verification, but you don't need those to build a pretty solid who-talks-to-who-when network even in encrypted chatrooms.

I understand why they implemented some of the metadata this way, but the encryption-stapled-to-unencrypted-messaging approach just leaves a lot to be desired. Signal, on the other hand, leaks pretty much nothing.