feat(ai): Upgrade to Gemini 3 Flash + add 503 fallback resilience

- Switch primary model from flash-lite (2.5) to 3-flash (Gemini 3 Flash Preview)
  for better reasoning and thinking mode across all AI features
- Add _is_retryable() method to handle 503 UNAVAILABLE (server overload)
  in addition to existing 429 rate limit fallback
- Fallback chain: 3-flash → 2.5-flash-lite → 2.5-flash

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

app.py

@@ -308,7 +308,7 @@ login_manager.login_message = 'Zaloguj się, aby uzyskać dostęp do tej strony.
 
 # Initialize Gemini service
 try:
-    gemini_service.init_gemini_service(model='flash-lite')  # Paid tier: Unlimited RPD, fallback: 3-flash (10K RPD) → 2.5-flash (10K RPD)
+    gemini_service.init_gemini_service(model='3-flash')  # Paid tier: 10K RPD, thinking mode, fallback: 2.5-flash-lite (Unlimited) → 2.5-flash (10K)
     logger.info("Gemini service initialized successfully")
 except Exception as e:
     logger.error(f"Failed to initialize Gemini service: {e}")

gemini_service.py

@@ -193,6 +193,13 @@ class GeminiService:
         error_str = str(error)
         return '429' in error_str or 'RESOURCE_EXHAUSTED' in error_str
 
+    @staticmethod
+    def _is_retryable(error: Exception) -> bool:
+        """Check if error is retryable (rate limit or server overload)."""
+        error_str = str(error)
+        return ('429' in error_str or 'RESOURCE_EXHAUSTED' in error_str or
+                '503' in error_str or 'UNAVAILABLE' in error_str)
+
     def _build_generation_config(self, model: str, temperature: float,
                                   max_tokens: Optional[int],
                                   thinking_level: Optional[str]) -> types.GenerateContentConfig:
@@ -324,12 +331,12 @@ class GeminiService:
                 return response_text
 
             except Exception as e:
-                if self._is_rate_limited(e) and model != models_to_try[-1]:
-                    logger.warning(f"Rate limited on {model}, trying next fallback...")
+                if self._is_retryable(e) and model != models_to_try[-1]:
+                    logger.warning(f"Retryable error on {model} ({type(e).__name__}), trying next fallback...")
                     last_error = e
                     continue
 
-                # Non-429 error or last model in chain — fail
+                # Non-retryable error or last model in chain — fail
                 latency_ms = int((time.time() - start_time) * 1000)
 
                 self._log_api_cost(