fix(connector): preserve suffix stores on gpu prefix hits

daser/connector/scheduler/lifecycle.py

@@ -132,10 +132,6 @@ def get_num_new_matched_tokens(
             logger.debug("[CONNECTOR] cache miss req=%s", request.request_id[:8])
             return 0, False
 
-        extra_tokens = _contiguous_prefix_tokens(chunks, num_computed_tokens)
-        if extra_tokens <= 0:
-            return 0, False
-
         pending_store = (
             None
             if skip_save
@@ -148,6 +144,10 @@ def get_num_new_matched_tokens(
         if pending_store is not None:
             self._pending_alloc[request.request_id] = pending_store
 
+        extra_tokens = _contiguous_prefix_tokens(chunks, num_computed_tokens)
+        if extra_tokens <= 0:
+            return 0, False
+
         available = len(tokens) - num_computed_tokens
         if extra_tokens >= available:
             extra_tokens = available - 1

tests/connector/test_daser_connector.py

@@ -2163,7 +2163,7 @@ class Output:
 
 
 def test_prefix_mode_hit_tracks_store_from_first_missing_slot():
-    """Warm prefix hits should not rewrite loaded slots, only later misses."""
+    """Warm prefix hits track suffix stores even when GPU coverage is longer."""
 
     class MockIPCClient:
         def lookup(self, tokens, model_id):
@@ -2211,6 +2211,10 @@ class MockRequest:
     assert pending.rolling_key == "hit-0"
     assert pending.rolling_slot_index == 1
 
+    gpu_hit_connector = MockConnector()
+    assert gpu_hit_connector.get_num_new_matched_tokens(MockRequest(), 8) == (0, False)
+    assert gpu_hit_connector.pending_alloc["req"].start_slot_index == 1
+
 
 def test_prefix_mode_hit_still_allocates_missing_slot_stores_after_load():
     """Prefix-hit requests still store newly computed suffix slots."""