fix(chat): surface silent generations instead of empty bubbles

Qwen3:14b sometimes burns output tokens on tool-calling attempts whose
emission doesn't parse into any field we read — eval_count > 0 but no
thinking/content/tool_calls ever stream to the caller. Generation
completes "successfully," the user sees an empty assistant bubble, and
no error is logged. Seen in conv 220 today.

Two safety rails:

- stream_chat_with_tools now tracks whether it yielded anything; when
  Ollama's done frame reports eval_count > 0 with zero yields, log a
  warning including the last ~5 raw frames so the next occurrence leaves
  breadcrumbs for diagnosis.

- run_generation checks the same post-condition after the tool loop
  exits and, if content is empty with no tool calls but output_tokens
  > 0, substitutes a visible fallback message and streams it as a chunk
  so the user gets something readable instead of a blank bubble.

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

src/fabledassistant/services/generation_task.py

@@ -542,6 +542,29 @@ async def run_generation(
         # Strip model artifacts from final content
         buf.content_so_far = _TOOL_CALL_MARKER.sub("", buf.content_so_far)
 
+        # Silent-generation safety net: the model burned output tokens but
+        # nothing landed in content or tool_calls (seen with qwen3:14b when
+        # its tool-call emission doesn't parse). Show a visible fallback so
+        # the user isn't staring at an empty bubble.
+        if (
+            not cancelled
+            and not buf.content_so_far.strip()
+            and not all_tool_calls
+            and (timing.get("output_tokens") or 0) > 0
+        ):
+            logger.warning(
+                "Silent generation for conv %d: output_tokens=%s but empty content "
+                "and no tool calls (model=%s)",
+                conv_id, timing.get("output_tokens"), model,
+            )
+            fallback = (
+                "I wasn't able to produce a usable response — the model generated "
+                "tokens that couldn't be parsed as content or a tool call. "
+                "Please try rephrasing, or try again."
+            )
+            buf.content_so_far = fallback
+            buf.append_event("chunk", {"chunk": fallback})
+
         # Final save
         logger.info("Generation complete for conv %d: content_length=%d, tool_calls=%d",
                      conv_id, len(buf.content_so_far), len(all_tool_calls))

src/fabledassistant/services/llm.py

@@ -265,20 +265,31 @@ async def stream_chat_with_tools(
         ) as resp:
             await _raise_ollama_error(resp, model)
             accumulated_tool_calls: list[dict] = []
+            # Silent-generation diagnostic: if Ollama reports non-zero eval_count
+            # but we never yielded any thinking/content/tool_calls, something
+            # in the frames isn't landing in a field we read. Capture the last
+            # few frames so we can see what Ollama actually sent.
+            yielded_anything = False
+            recent_frames: list[str] = []
             async for line in resp.aiter_lines():
                 if not line.strip():
                     continue
+                if len(recent_frames) >= 5:
+                    recent_frames.pop(0)
+                recent_frames.append(line[:500])
                 data = json.loads(line)
                 msg = data.get("message", {})
 
                 # Thinking chunks (qwen3 chain-of-thought, only when think=True)
                 thinking = msg.get("thinking", "")
                 if thinking:
+                    yielded_anything = True
                     yield ChatChunk(type="thinking", content=thinking)
 
                 # Content chunks
                 chunk = msg.get("content", "")
                 if chunk:
+                    yielded_anything = True
                     yield ChatChunk(type="content", content=chunk)
 
                 # Collect tool calls from any message (some models
@@ -294,13 +305,21 @@ async def stream_chat_with_tools(
                             len(accumulated_tool_calls),
                             json.dumps(accumulated_tool_calls)[:500],
                         )
+                        yielded_anything = True
                         yield ChatChunk(type="tool_calls", tool_calls=accumulated_tool_calls)
                     else:
                         logger.debug("Ollama done with no tool calls")
+                    eval_count = data.get("eval_count") or 0
+                    if not yielded_anything and eval_count > 0:
+                        logger.warning(
+                            "Ollama silent generation: model=%s eval_count=%d but no "
+                            "thinking/content/tool_calls were yielded. Last frames: %s",
+                            model, eval_count, recent_frames,
+                        )
                     yield ChatChunk(
                         type="done",
                         prompt_tokens=data.get("prompt_eval_count"),
-                        output_tokens=data.get("eval_count"),
+                        output_tokens=eval_count,
                     )
                     break