[optimizer] bound KVCM replay state with 24h TTL

Author: Tyndalllll

kv_cache_manager/optimizer/docs/online_event_subscription_code_review.md

@@ -298,6 +298,7 @@ KVCM capacity_bytes    -> 一个 capacity_gb 容量点
 eviction_policy        -> lru
 enable_prefix_hash     -> true
 theoretical max        -> true
+ttl_seconds            -> 24 hours
 ```
 
 ### Instance 映射
@@ -333,6 +334,7 @@ OnlineOptimizerManager::RegisterInstance
 - KVCM 自动接入暂不推断 linear state；多 group 的语义不明确时标记为 unsupported 并跳过。
 - Subscriber 收到已知 unsupported Instance 的事件时直接丢弃，不触发配置刷新。
 - KVCM Group quota 被每个 Instance 分别作为完整容量模拟，不是共享 quota。
+- KVCM 自动创建的 Group 固定使用 24 小时 TTL，限制 LiteHit 保留的历史工作集。
 - Group 已存在时不检查 capacity、prefix hash、theoretical flag。
 - Instance 已存在时不检查 block size 和 specs。
 

kv_cache_manager/optimizer/service/optimizer_service_impl.cc

@@ -20,6 +20,7 @@ namespace kv_cache_manager {
 namespace {
 
 constexpr long double kBytesPerGb = 1024.0L * 1024.0L * 1024.0L;
+constexpr int64_t kKvcmAutoGroupTtlSeconds = 24 * 60 * 60;
 
 void SetPbResponseHeader(proto::optimizer::CommonResponseHeader *header, ErrorCode ec) {
     auto *status = header->mutable_status();
@@ -330,6 +331,7 @@ ErrorCode OptimizerServiceImpl::ApplyKvcmConfiguration(const proto::optimizer::K
             group.set_eviction_policy("lru");
             group.set_enable_prefix_hash(true);
             group.set_enable_theoretical_max_cache(true);
+            group.set_ttl_seconds(kKvcmAutoGroupTtlSeconds);
 
             std::string invalid_fields;
             if (!group.ValidateRequiredFields(invalid_fields)) {

kv_cache_manager/optimizer/test/kvcm_event_subscriber_test.cc

@@ -266,6 +266,7 @@ TEST_F(KvcmEventSubscriberTest, DiscoversLeaderAndRegistersConfigurationBeforeCo
     EXPECT_DOUBLE_EQ(2.0, group->capacity_gb()[0]);
     EXPECT_TRUE(group->enable_prefix_hash());
     EXPECT_TRUE(group->enable_theoretical_max_cache());
+    EXPECT_EQ(24 * 60 * 60, group->ttl_seconds());
 
     leader.event_service_.Publish(MakeEvent("known", "normal", 1));
     auto short_prompt = MakeEvent("known", "short", 2);

kv_cache_manager/optimizer/test/optimizer_service_impl_test.cc

@@ -433,6 +433,7 @@ TEST_F(OptimizerServiceImplTest, ApplyKvcmConfigurationCreatesEnabledGroupAndIns
     EXPECT_DOUBLE_EQ(2.0, stored_group->capacity_gb().front());
     EXPECT_TRUE(stored_group->enable_prefix_hash());
     EXPECT_TRUE(stored_group->enable_theoretical_max_cache());
+    EXPECT_EQ(24 * 60 * 60, stored_group->ttl_seconds());
 
     ASSERT_EQ(EC_OK, manager_->GetInstanceState("kvcm-instance", [](const InstanceState &state) {
         EXPECT_EQ("kvcm-group", state.instance_info->instance_group_name());