fix(planner): don't block agg decode scaling when max_num_batched_tokens is missing (#8196)

Signed-off-by: hongkuanz <hongkuanz@nvidia.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

components/src/dynamo/planner/core/load_scaling.py

@@ -243,10 +243,11 @@ class LoadScalingMixin:
         d_caps = self._capabilities.decode
         max_tokens = d_caps.max_num_batched_tokens if d_caps else None
         if not max_tokens or max_tokens <= 0:
-            logger.warning("max_num_batched_tokens not available, skipping agg scaling")
-            self._diag_load_reason = "insufficient_data"
-            return None
-
+            logger.warning(
+                "max_num_batched_tokens not available, skipping agg prefill scaling"
+            )
+            p_desired = None
+        else:
             p_desired = self._agg_prefill_scaling(fpm_stats, num_workers, max_tokens)
         d_desired = self._agg_decode_scaling(fpm_stats, num_workers)
 
@@ -258,6 +259,9 @@ class LoadScalingMixin:
             desired = p_desired
         elif d_desired is not None and d_desired > num_workers:
             desired = d_desired
+        elif p_desired is None and d_desired is not None and d_desired < num_workers:
+            # Prefill signal unavailable: allow decode-only scale-down.
+            desired = d_desired
         elif (
             p_desired is not None
             and p_desired < num_workers