diff --git a/backend/core/restream.py b/backend/core/restream.py
index be38bb8..08b38c8 100644
--- a/backend/core/restream.py
+++ b/backend/core/restream.py
@@ -72,6 +72,34 @@ _HEALTH_WARN_SECS  = 12.0  # > 12 s without data → warning (normal reconnects
 _HEALTH_ERROR_SECS = 25.0  # > 25 s → error, trigger correction
 
 
+def _is_provider_session_end(last_stderr: str) -> bool:
+    """True when ffmpeg's last stderr line indicates the provider closed the TCP connection.
+
+    "Error during demuxing: Connection timed out" is the provider's session boundary
+    signal regardless of how long that session lasted.  It is NOT an error — the
+    provider's CDN simply ended the HTTP/TCP session (common every 30-300 s).
+    """
+    low = last_stderr.lower()
+    return "connection timed out" in low
+
+
+def _is_permanent_domain_failure(exc: Exception) -> bool:
+    """True only for failures that won't self-heal in a few seconds.
+
+    DNS NXDOMAIN and connection-refused are structural — the domain is unreachable.
+    HTTP 4xx (458, 401, etc.) and EOF are ephemeral: often a provider token-refresh
+    race condition that resolves on the next attempt or within seconds.
+    """
+    msg = str(exc).lower()
+    return (
+        "name or service not known" in msg
+        or "could not resolve" in msg
+        or "getaddrinfo" in msg
+        or "connection refused" in msg
+        or "errno 111" in msg
+    )
+
+
 def _classify_error(exc: Exception, pump_duration: float) -> str:
     """Return a human-readable Spanish cause for a stream failure."""
     msg = str(exc).lower()
@@ -87,10 +115,12 @@ def _classify_error(exc: Exception, pump_duration: float) -> str:
         return "Canal no encontrado en el proveedor (404)"
     if "401" in msg:
         return "Autenticación rechazada por el proveedor (401)"
+    if "458" in msg:
+        return "Token de sesión no disponible en el proveedor (HTTP 458)"
+    if "end of file" in msg:
+        return "Proveedor cerró la conexión (EOF)"
     if "rc=1" in msg or "rc=-1" in msg:
         return f"Error del proveedor (ffmpeg rc≠0) tras {int(pump_duration)}s"
-    if "rc=0" in msg:
-        return f"Sesión del proveedor cerrada tras {int(pump_duration)}s"
     return f"Error en {int(pump_duration)}s: {str(exc)[:200]}"
 
 
@@ -179,6 +209,9 @@ class BroadcastGroup:
         self._corrections:     int = 0      # how many times health monitor killed stuck ffmpeg
         self._health_task:     asyncio.Task | None = None
 
+        # Last stderr line from ffmpeg — used to classify pump exit reason
+        self._last_ffmpeg_stderr: str = ""
+
     # ------------------------------------------------------------------
     # Public API
     # ------------------------------------------------------------------
@@ -471,6 +504,7 @@ class BroadcastGroup:
             while self._running:
                 current_url = self._stream_urls[url_idx % n]
                 self._active_url = current_url
+                self._last_ffmpeg_stderr = ""   # reset so exit classification is fresh
                 pump_start = time.time()
                 try:
                     await self._pump_once(current_url)
@@ -484,10 +518,15 @@ class BroadcastGroup:
                     pump_duration = time.time() - pump_start
                     self._reconnect_count += 1
 
-                    is_real_failure = pump_duration < PROVIDER_SESSION_MIN
+                    # Primary classification: if ffmpeg's last stderr says
+                    # "Connection timed out" it is always a provider session boundary —
+                    # the duration doesn't matter (sessions can be 20 s or 200 s).
+                    # Only fall back to the duration threshold for other error types.
+                    is_session_end = _is_provider_session_end(self._last_ffmpeg_stderr)
+                    is_real_failure = (not is_session_end) and (pump_duration < PROVIDER_SESSION_MIN)
 
                     if is_real_failure:
-                        # Quick exit → real error (DNS, refused, HTTP 4xx, stall)
+                        # Quick exit with a non-session-timeout error
                         self._failure_count += 1
                         failure_attempts += 1
                         cause = _classify_error(e, pump_duration)
@@ -505,8 +544,11 @@ class BroadcastGroup:
                                 f"{cause} (pump={pump_duration:.1f}s)"
                             )
 
-                        # Mark domain as error immediately so other streams skip it
-                        asyncio.create_task(self._mark_domain_down(current_url))
+                        # Only mark domain as permanently down for DNS/refused failures.
+                        # HTTP 458 / EOF are ephemeral (provider token refresh race) —
+                        # marking them as error causes unnecessary domain rotation.
+                        if _is_permanent_domain_failure(e):
+                            asyncio.create_task(self._mark_domain_down(current_url))
 
                         # Rotate to next domain
                         url_idx = (url_idx + 1) % n
@@ -523,10 +565,10 @@ class BroadcastGroup:
                         # else: try next domain immediately
 
                     else:
-                        # Normal provider session timeout (>25 s) — reconnect to same domain
+                        # Provider session boundary (TCP closed by provider, any duration)
                         self._session_count += 1
-                        preferred_idx = url_idx          # remember: this domain works
-                        url_idx       = preferred_idx    # stay on it
+                        preferred_idx = url_idx
+                        url_idx       = preferred_idx   # stay on same domain
 
                         if in_failure_mode:
                             outage_secs = int(time.time() - (failure_start or time.time()))
@@ -550,7 +592,9 @@ class BroadcastGroup:
                             )
 
                         self._status = "reconnecting"
-                        # Reconnect immediately — no backoff for normal session rotations
+                        # Brief pause before reconnect — gives provider token-refresh
+                        # system time to update before we open a new session.
+                        await asyncio.sleep(0.4)
 
         finally:
             self._status  = "stopped"
@@ -654,6 +698,7 @@ class BroadcastGroup:
             async for line in proc.stderr:
                 text = line.decode(errors="replace").rstrip()
                 if text:
+                    self._last_ffmpeg_stderr = text  # track for exit classification
                     logger.warning(f"[ch{self.channel_id}] ffmpeg: {text}")
                     if any(p in text for p in _AV_WARN_PATTERNS):
                         self._av_desync_count += 1