fix: fall back to secondary LLM when the primary call raises

Primary model errors (HTTP 4xx/5xx, connection failures) raised out of
_call_llm uncaught, skipping the fallback branch entirely — fallback
only ran when the primary returned an empty/"can't answer" string.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

AI/talk_handler.py

@@ -614,6 +614,7 @@ async def _call_llm(
 ) -> str:
     disable_thinking = FORCE_DISABLE_THINKING if disable_thinking is None else disable_thinking
 
+    try:
         result = await _call_single_model(
             messages,
             api_url=LLAMA_API_URL,
@@ -625,6 +626,9 @@ async def _call_llm(
             disable_thinking=disable_thinking,
             reasoning_budget=reasoning_budget,
         )
+    except Exception:
+        logger.exception("Primary model call failed")
+        result = ""
 
     # Если основная модель не смогла ответить — пробуем fallback
     if LLAMA_FALLBACK_API_URL and (not result or _CANT_ANSWER_RE.search(result)):