fix: close relay loop from prefixed dedup mismatch

The dispatcher deduped on inbound text but relayed a prefixed copy
("[net/sender] text"), so a re-heard relayed message hashed differently
and was re-relayed, stacking another prefix each hop — an unbounded loop,
exactly what the dedup was meant to prevent.

Add MessageBus.mark_seen() and record the on-air relayed string before
sending, so the bridge recognises and drops its own relayed copies.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

host-bridge/bridge_core.py

@@ -56,5 +56,21 @@ class MessageBus:
             self._seen[h] = now + self._dedup_ttl
             return False
 
+    def mark_seen(self, text: str) -> None:
+        """Record text as already-seen without reporting duplicate status.
+
+        The dispatcher relays a *prefixed* copy of each message (e.g.
+        "[meshtastic/x] hello"), not the original text. Dedup is content-hash
+        based, so unless the exact on-air relayed string is recorded here, the
+        bridge won't recognise its own relayed copy when it echoes back from
+        another network — it would re-relay it, stacking another prefix each
+        hop (an unbounded loop). Marking every outbound string seen closes
+        that loop.
+        """
+        h = self._content_hash(text)
+        now = time.monotonic()
+        with self._lock:
+            self._seen[h] = now + self._dedup_ttl
+
     def get(self, timeout: float | None = None) -> BridgeMessage:
         return self._queue.get(timeout=timeout)