fix(telegram): back off on transient polling failures

Give getUpdates a wider HTTP timeout than the long-poll window, log transient SSL/network disconnects as warnings, apply exponential backoff with recovery logging, and only escalate to error after five consecutive failures.

reaction_board/notify/telegram.py

@@ -86,9 +86,17 @@ class TelegramClient:
         self.poll_timeout = poll_timeout
         self.logger = logger
 
-    def _call(self, method: str, params: Mapping[str, Any] | None = None) -> Any:
+    def _call(
+        self,
+        method: str,
+        params: Mapping[str, Any] | None = None,
+        *,
+        timeout: float | None = None,
+    ) -> Any:
         url = f"https://api.telegram.org/bot{self.token}/{method}"
-        payload = self.transport.post(url, params or {}, timeout=self.timeout)
+        payload = self.transport.post(
+            url, params or {}, timeout=self.timeout if timeout is None else timeout
+        )
         if not isinstance(payload, Mapping):
             raise TelegramError("Telegram 响应不是对象")
         if not payload.get("ok"):
@@ -145,7 +153,10 @@ class TelegramClient:
         }
         if offset is not None:
             params["offset"] = int(offset)
-        result = self._call("getUpdates", params)
+        # 长轮询本身会挂起 poll_timeout 秒，HTTP 读超时必须留出更大余量。
+        result = self._call(
+            "getUpdates", params, timeout=self.poll_timeout + 15
+        )
         return result if isinstance(result, list) else []
 
     def answer_callback(self, callback_id: str, text: str = "") -> Any:

reaction_board/services/telegram_poller.py

@@ -88,11 +88,41 @@ class TelegramPoller:
             self._log("warning", "telegram.answer_failed", error=str(exc))
             return False
 
-    def run_forever(self, stop_event=None) -> None:
+    def run_forever(
+        self,
+        stop_event=None,
+        *,
+        initial_delay: float = 3.0,
+        max_delay: float = 60.0,
+        max_iterations: int | None = None,
+    ) -> None:
         offset: int | None = None
+        failures = 0
+        delay = initial_delay
+        iterations = 0
         while stop_event is None or not stop_event.is_set():
             try:
                 offset = self.run_once(offset)
-            except Exception as exc:  # noqa: BLE001  长轮询失败后退避重试
-                self._log("error", "telegram.poll_failed", error=str(exc))
-                self._sleep(3)
+            except Exception as exc:  # noqa: BLE001  长轮询瞬时断连后退避重试
+                failures += 1
+                level = "error" if failures >= 5 else "warning"
+                self._log(
+                    level, "telegram.poll_failed",
+                    error=str(exc), failures=failures, retry_in=delay,
+                )
+                if stop_event is not None:
+                    if stop_event.wait(delay):
+                        break
+                else:
+                    self._sleep(delay)
+                delay = min(delay * 2, max_delay)
+            else:
+                if failures:
+                    self._log(
+                        "info", "telegram.poll_recovered", failures=failures
+                    )
+                failures = 0
+                delay = initial_delay
+            iterations += 1
+            if max_iterations is not None and iterations >= max_iterations:
+                break