[NVIDIA#12560][fix] Fix disaggserving hang on block reuse after eviction (NVIDIA#12667)

Signed-off-by: Iman Tabrizian <10105175+tabrizian@users.noreply.github.com>

Author: Tabrizian

cpp/tensorrt_llm/batch_manager/kvCacheManager.cpp

@@ -1688,6 +1688,15 @@ std::pair<SizeType32, std::vector<KVCacheBlock::IdType>> WindowBlockManager::sto
             }
             if (pinBlocks)
             {
+                // If the block has no refs it sits in the eviction policy's free
+                // queue. Claim it first so that the later unpinBlocksById /
+                // releaseBlock cycle does not create a duplicate queue entry.
+                // Pass the block's existing priority and duration so that
+                // claimBlock does not clear its retention/expiry metadata.
+                if (!searchRoot->hasRefs())
+                {
+                    mEvictionPolicy->claimBlock(searchRoot, searchRoot->getPriority(), searchRoot->getDurationMs());
+                }
                 searchRoot->incRefCount();
                 pinnedBlockIds.push_back(searchRoot->getBlockId());
             }