fix: prevent replication stream from skipping local writes after other writer advance

In Stream.get_updates(), the early-return path (last_token >=
minimal_local_current_token) was advancing last_token to current_token.
In a multi-writer setup, current_token (via get_current_token_for_writer)
can be inflated by other writers' positions (via advance() ->
_add_persisted_position -> _max_position_of_local_instance). This caused
last_token to jump past the local writer's actual position, silently
dropping subsequent local writes.

Fix: advance last_token to minimal_local_current_token() instead of
current_token. This ensures last_token never exceeds the local
writer's actual persisted position, so subsequent local writes are
always correctly detected and sent as RDATA.

synapse/replication/tcp/streams/_base.py

@@ -178,8 +178,14 @@ class Stream:
         # If the minimum current token for the local instance is less than or
         # equal to the last thing we published, we know that there are no
         # updates.
+        #
+        # We only advance last_token to minimal_local_current_token() rather
+        # than current_token. current_token may be inflated by other writers'
+        # positions (via advance()), and jumping last_token past the local
+        # writer's actual position would cause subsequent local writes to be
+        # silently dropped by the early return on the next call.
         if self.last_token >= self.minimal_local_current_token():
-            self.last_token = current_token
+            self.last_token = self.minimal_local_current_token()
             return [], current_token, False
 
         updates, current_token, limited = await self.get_updates_since(