Keep bot operational when forecast model is unavailable

tools/train_torch_recurrent_forecaster.py

@@ -1020,12 +1020,22 @@ def _ensemble_candidate(members: list[dict[str, Any]], seeds: list[int]) -> dict
         "context_norm_weight",
         "context_norm_bias",
     )
-    result["ensemble_members"] = [
-        {name: member[name] for name in export_names if name in member}
-        | {"seed": seeds[index] if index < len(seeds) else index}
-        for index, member in enumerate(members)
-    ]
     result["ensemble_size"] = len(members)
+    result["ensemble_seeds"] = [seeds[index] if index < len(seeds) else index for index in range(len(members))]
+    if len(members) > 1:
+        result["ensemble_members"] = [
+            {name: member[name] for name in export_names if name in member}
+            | {"seed": seeds[index] if index < len(seeds) else index}
+            for index, member in enumerate(members)
+        ]
+        # Ensemble inference uses the member payloads. Keeping the first
+        # member at the top level duplicated a complete network in every
+        # exported symbol and could push an otherwise valid artifact over
+        # the server upload limit.
+        for name in export_names:
+            result.pop(name, None)
+    else:
+        result.pop("ensemble_members", None)
     symbol_names = sorted(
         {
             symbol