fix(MAT): skip redundant stream edit + retry empty responses with escalation model

1. Track last-sent text during streaming, skip final m.replace edit when
   content is identical — eliminates spurious '(bearbeitet)' indicator.
2. When base model (haiku) returns empty content + no tool calls, auto-retry
   with escalation model (sonnet) before giving up.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

bot.py

@@ -3217,11 +3217,12 @@ class Bot:
 
         try:
             reply = ""
+            last_sent_text = ""
             streamed_event_id: str | None = None  # set when streaming has already posted a message in Matrix
 
             # Agentic tool-calling loop: iterate up to MAX_TOOL_ITERATIONS
             for iteration in range(MAX_TOOL_ITERATIONS):
-                content, tool_calls, usage, streamed_event_id = await self._stream_chat_completion(
+                content, tool_calls, usage, streamed_event_id, last_sent_text = await self._stream_chat_completion(
                     room_id=room.room_id,
                     model=model,
                     messages=messages,
@@ -3242,6 +3243,12 @@ class Bot:
                     },
                 )
 
+                # Empty response with no tool calls — retry once with escalation model
+                if not content and not tool_calls and model != ESCALATION_MODEL:
+                    logger.warning("[empty-response] %s returned nothing, retrying with %s", model, ESCALATION_MODEL)
+                    model = ESCALATION_MODEL
+                    continue
+
                 if not tool_calls:
                     # No tool calls — final text response
                     break
@@ -3276,11 +3283,12 @@ class Bot:
             if iteration > 0:
                 sentry_sdk.set_tag("used_tools", "true")
 
-            # Send / finalize reply. If we streamed, just do a final edit so the
-            # Matrix message reflects the complete text (otherwise progressive
-            # throttling may have stopped short of the last tokens).
+            # Send / finalize reply. If we streamed, do a final edit only if
+            # the complete text differs from what was last sent (avoids the
+            # "(bearbeitet)" / "(edited)" indicator for unchanged messages).
             if reply:
                 if streamed_event_id:
+                    if reply != last_sent_text:
                         await self._send_stream_edit(room.room_id, streamed_event_id, reply, final=True)
                 else:
                     await self._send_text(room.room_id, reply)
@@ -3728,22 +3736,24 @@ class Bot:
         messages: list[dict],
         tools: list | None,
         prior_event_id: str | None = None,
-    ) -> tuple[str, list[dict] | None, dict | None, str | None]:
+    ) -> tuple[str, list[dict] | None, dict | None, str | None, str]:
         """Stream one chat completion turn.
 
         Progressively edits a Matrix message as content tokens arrive (unless
         tool_calls have started — those suppress visible streaming until the
         model settles on plain text on a later iteration).
 
-        Returns (content, tool_calls or None, usage dict or None, event_id).
+        Returns (content, tool_calls or None, usage dict or None, event_id, last_sent_text).
         `event_id` is the Matrix event we've been streaming into, or None if
         we didn't (yet) post a visible message this turn.
+        `last_sent_text` is the text last sent/edited to Matrix (for dedup).
         """
         content_parts: list[str] = []
         tool_calls_acc: dict[int, dict] = {}
         usage: dict | None = None
         event_id = prior_event_id
         last_edit = 0.0
+        last_sent_text: str = ""  # track what was last sent to Matrix to avoid redundant edits
         EDIT_THROTTLE = 0.6  # seconds — keep Matrix edit traffic reasonable
         MIN_CHARS_BEFORE_POST = 20  # avoid posting a single character first
 
@@ -3773,7 +3783,7 @@ class Bot:
                     "prompt_tokens": getattr(resp.usage, "prompt_tokens", 0),
                     "completion_tokens": getattr(resp.usage, "completion_tokens", 0),
                 }
-            return choice.message.content or "", tc_list, u, event_id
+            return choice.message.content or "", tc_list, u, event_id, ""
 
         async for chunk in stream:
             if not chunk.choices:
@@ -3813,6 +3823,7 @@ class Bot:
                                 event_id = await self._send_stream_start(room_id, text_so_far)
                             else:
                                 await self._send_stream_edit(room_id, event_id, text_so_far)
+                            last_sent_text = text_so_far
                             last_edit = now
 
             # Some providers attach usage to the last choice chunk
@@ -3874,7 +3885,7 @@ class Bot:
             "[stream] model=%s chars=%d tool_calls=%d streamed_to_matrix=%s",
             model, len(content), len(tc_list or []), event_id is not None,
         )
-        return content, tc_list, usage, event_id
+        return content, tc_list, usage, event_id, last_sent_text
 
     async def _get_call_encryption_key(self, room_id: str, sender: str, caller_device_id: str = "") -> bytes | None:
         """Read E2EE encryption key from call.member state (MSC4143) or timeline (legacy).