fix: refill from expanded candidate history

modelhub_submmit_api/modelhub_client.py

@@ -321,7 +321,7 @@ class ModelHubClientPool:
         identity_hash = hashlib.sha256(identity.encode("utf-8")).hexdigest()
         self._selection_cursor = int(identity_hash[:12], 16) % len(clients)
         self._verify_cache: dict[str, tuple[float, dict[str, Any]]] = {}
-        self._verify_cache_ttl = max(1.0, float(os.getenv("MODELHUB_AGENT_VERIFY_CACHE_TTL_SECONDS", "30")))
+        self._verify_cache_ttl = max(1.0, float(os.getenv("MODELHUB_AGENT_VERIFY_CACHE_TTL_SECONDS", "900")))
         # Single reader client to avoid fanout on read operations
         self._reader = clients[0]
 
@@ -539,12 +539,15 @@ class ModelHubClientPool:
         self._verify_cache[model_id] = (time.monotonic(), payload)
 
     def begin_cycle(self) -> None:
-        """Drop model verification cache entries from the previous scan cycle."""
+        """Keep recent verification results across cycles and prune expired entries."""
         with self._state_lock:
-            self._verify_cache.clear()
+            now = time.monotonic()
+            for model_id, (cached_at, _payload) in list(self._verify_cache.items()):
+                if now - cached_at >= self._verify_cache_ttl:
+                    self._verify_cache.pop(model_id, None)
 
     def search_by_model_id(self, model_id: str) -> dict[str, Any]:
-        # Check cache first (per-cycle cache to avoid repeated API calls for the same model)
+        # Reuse recent model verification results across short poll cycles.
         with self._state_lock:
             cached = self._read_from_cache(model_id)
         if cached is not None: