fix simulation reporting and research resilience

backend/app/services/llm.py

@@ -139,6 +139,21 @@ class LLMClient:
                     )
                 return content.strip()
             except Exception as e:
+                # Provider validation/policy responses are deterministic for
+                # the same payload. Retrying them only adds latency and can
+                # multiply the provider-side rejection count. Keep retries
+                # for transient network, timeout, and 5xx failures.
+                status_code = getattr(e, "status_code", None)
+                if status_code in {400, 401, 403, 404, 422}:
+                    await self._record_purpose(purpose, time.monotonic() - t0, ok=False)
+                    logger.error(
+                        "LLM 请求不可重试（用途=%s，HTTP %s，耗时=%.2f秒）：%s",
+                        purpose or "-",
+                        status_code,
+                        time.monotonic() - t0,
+                        e,
+                    )
+                    raise
                 if attempt == retries - 1:
                     await self._record_purpose(purpose, time.monotonic() - t0, ok=False)
                     logger.error(