refactor: V536/V569 实验同步为 V562 副本 (粗排截断 + 23 路白名单)

V536: 整类重写。之前是基于 V568 召回 + 自定义"极简 mergeAndSort"(只保留 rov 空兜底 +
流量池强插),  现替换为 V562 设计: 23 路召回白名单(7 个性化 + 17 非个性化) + 统一
按 alg_vid_recommend_exp_feature_20250212 粗排分截断, 不再 override mergeAndSort
(使用 Basic 完整 fusion 流程)。Apollo key 保持 v536, 弃用 item_recall_weight。

V569: PERSONAL 白名单新增 YearShareDkElementsRecallStrategy.PUSH_FROM (6 路 → 7 路),
对齐 V562 的 dk_elements 行为路实验。Apollo key 保持 v569。

V566/V564 当前已是语义等价副本 (仅名字标签不同), 本次未触碰。

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

recommend-server-service/src/main/java/com/tzld/piaoquan/recommend/server/service/rank/strategy/RankStrategy4RegionMergeModelV536.java

@@ -7,10 +7,7 @@ import com.tzld.piaoquan.recommend.server.common.base.RankItem;
 import com.tzld.piaoquan.recommend.server.model.MachineInfo;
 import com.tzld.piaoquan.recommend.server.model.Video;
 import com.tzld.piaoquan.recommend.server.service.FeatureService;
-import com.tzld.piaoquan.recommend.server.service.funnel.ColdStartAction;
-import com.tzld.piaoquan.recommend.server.service.funnel.FunnelContext;
 import com.tzld.piaoquan.recommend.server.service.rank.RankParam;
-import com.tzld.piaoquan.recommend.server.service.rank.RankResult;
 import com.tzld.piaoquan.recommend.server.service.rank.bo.UserShareReturnProfile;
 import com.tzld.piaoquan.recommend.server.service.rank.extractor.ExtractVideoMergeCate;
 import com.tzld.piaoquan.recommend.server.service.rank.tansform.FeatureV6;
@@ -20,7 +17,6 @@ import com.tzld.piaoquan.recommend.server.util.*;
 import lombok.extern.slf4j.Slf4j;
 import org.apache.commons.collections4.CollectionUtils;
 import org.apache.commons.collections4.MapUtils;
-import org.apache.commons.lang3.RandomUtils;
 import org.apache.commons.lang3.StringUtils;
 import org.springframework.beans.factory.annotation.Autowired;
 import org.springframework.stereotype.Service;
@@ -31,20 +27,6 @@ import java.util.concurrent.TimeUnit;
 import java.util.stream.Collectors;
 import java.util.stream.Stream;
 
-/**
- * V536 实验（2026-05-29 复写）：基于 V568 召回 + DNN 打分，fusion 只保留"流量池相关 + 兜底相关"
- *
- * 与 V568 唯一差异在 fusion 阶段（mergeAndSort）：
- *   - 保留：rov 空兜底 + 流量池按比例强插（topK 头部锁 + flowPoolP/newFlowPoolSelectRate 概率门 + 一侧用光兜底回填）
- *   - 删除：标签 filter / rov boost / 强插 / 品类降权 / 节日降权 / 密度控制（Basic 的段 2/3/4/5/6/8）
- *
- * Apollo key 保留 ${rank.score.merge.weightv536}，召回/打分逻辑与 V568 同源但参数独立可调，仅 fusion 策略不同。
- * 召回侧：与 V568 基线召回完全一致，RecallService 不需要 isHit536Exp 门控。
- *
- * 历史：
- *   - 原 V536（V569 + 4 路召回提权）在 2026-05-27 被复写为 V565 base + 极简 fusion
- *   - 2026-05-29 再次复写：base 从 V565 切到 V568（加回 5 路 oldSpecial + 3 路 prioriProvince，v1/cityRov 换回普通版）
- */
 @Service
 @Slf4j
 public class RankStrategy4RegionMergeModelV536 extends RankStrategy4RegionMergeModelBasic {
@@ -54,6 +36,61 @@ public class RankStrategy4RegionMergeModelV536 extends RankStrategy4RegionMergeM
     @Autowired
     private FeatureService featureService;
 
+    /**
+     * V536 个性化召回白名单 (7 路: V566 基础 6 路 + 1 路 dk_elements 行为路实验)：召回 key 含 mid/uid，
+     * 依赖该用户行为信号。
+     * V536 实验路径: YearShareDkElements (用户近期 share 行为 join dk_elements)
+     * 注：YearReturnCate2 因线上效果不佳, 2026-06-04 起移到非个性化白名单。
+     */
+    private static final Set<String> PERSONAL_RECALL_PUSH_FROMS = new HashSet<>(Arrays.asList(
+            UserCate1RecallStrategy.PUSH_FORM,
+            UserCate2RecallStrategy.PUSH_FORM,
+            Return1Cate2RosRecallStrategy.PUSH_FORM,
+            Return1Cate2StrRecallStrategy.PUSH_FORM,
+            YearShareCate1RecallStrategy.PUSH_FROM,
+            YearShareCate2RecallStrategy.PUSH_FROM,
+            YearShareDkElementsRecallStrategy.PUSH_FROM
+    ));
+
+    /**
+     * V536 非个性化召回白名单 (17 路)：只依赖 headVid + 地域/品类/相似度（vid-vid CF 也归此类）。
+     * 含 5 路旧地域、新地域、城市、head province/cate、先验省份、return 相似、scene CF、YearReturnCate2。
+     */
+    private static final Set<String> NON_PERSONAL_RECALL_PUSH_FROMS = new HashSet<>(Arrays.asList(
+            RegionHRecallStrategy.PUSH_FORM,
+            RegionHDupRecallStrategy.PUSH_FORM,
+            Region24HRecallStrategy.PUSH_FORM,
+            RegionRelative24HRecallStrategy.PUSH_FORM,
+            RegionRelative24HDupRecallStrategy.PUSH_FORM,
+            RegionRealtimeRecallStrategyV1.PUSH_FORM,
+            CityRovnRecallStrategy.PUSH_FROM,
+            HeadProvinceCate1RecallStrategy.PUSH_FORM,
+            HeadProvinceCate2RecallStrategy.PUSH_FORM,
+            HeadCate2RovRecallStrategy.PUSH_FROM,
+            PrioriProvinceRovnRecallStrategy.PUSH_FROM,
+            PrioriProvinceStrRecallStrategy.PUSH_FROM,
+            PrioriProvinceRosRecallStrategy.PUSH_FROM,
+            ReturnVideoRecallStrategy.PUSH_FORM,
+            SceneCFRovnRecallStrategy.PUSH_FORM,
+            SceneCFRosnRecallStrategy.PUSH_FORM,
+            YearReturnCate2RecallStrategy.PUSH_FROM
+    ));
+
+    /** PERSONAL ∪ NON_PERSONAL = 23 路。用于 fetchCoarseRankScores 跳过流量池等不参与截断的 vid。 */
+    private static final Set<String> ALL_ROV_PUSH_FROMS;
+    static {
+        Set<String> all = new HashSet<>(PERSONAL_RECALL_PUSH_FROMS);
+        all.addAll(NON_PERSONAL_RECALL_PUSH_FROMS);
+        ALL_ROV_PUSH_FROMS = Collections.unmodifiableSet(all);
+    }
+
+    /*
+     * 设计要点：
+     *   - fail-closed 白名单：RecallService 未来加新路不会自动进 V536，避免污染 vs V568 AB 对比
+     *   - 流量池 3 路 (flow_pool / quick_flow_pool / recall_strategy_hotspot) 不在任何名单——独立通道
+     *   - 调用顺序 = 个性化优先：同 vid 双类命中时归个性化，保护用户兴趣信号
+     */
+
     @Override
     public List<Video> mergeAndRankRovRecall(RankParam param) {
         Map<String, Double> mergeWeight = this.mergeWeight != null ? this.mergeWeight : new HashMap<>(0);
@@ -67,42 +104,30 @@ public class RankStrategy4RegionMergeModelV536 extends RankStrategy4RegionMergeM
         Set<Long> setVideo = new HashSet<>();
         setVideo.add(param.getHeadVid());
         List<Video> rovRecallRank = new ArrayList<>();
-        // -------------------5路特殊旧召回------------------
-        RecallUtils.extractOldSpecialRecall(mergeWeight.getOrDefault("oldSpecialN", (double) param.getSize()).intValue(), param, setVideo, rovRecallRank);
-        //-------------------return相似召回------------------
-        RecallUtils.extractRecall(mergeWeight.getOrDefault("v6", 5.0).intValue(), param, ReturnVideoRecallStrategy.PUSH_FORM, setVideo, rovRecallRank);
-        //-------------------新地域召回------------------
-        RecallUtils.extractRecall(mergeWeight.getOrDefault("v1", 5.0).intValue(), param, RegionRealtimeRecallStrategyV1.PUSH_FORM, setVideo, rovRecallRank);
-        //-------------------scene cf rovn------------------
-        RecallUtils.extractRecall(mergeWeight.getOrDefault("sceneCFRovn", 5.0).intValue(), param, SceneCFRovnRecallStrategy.PUSH_FORM, setVideo, rovRecallRank);
-        //-------------------scene cf rosn------------------
-        RecallUtils.extractRecall(mergeWeight.getOrDefault("sceneCFRosn", 5.0).intValue(), param, SceneCFRosnRecallStrategy.PUSH_FORM, setVideo, rovRecallRank);
-        // -------------------user cate1------------------
-        RecallUtils.extractRecall(mergeWeight.getOrDefault("cate1RecallN", 5.0).intValue(), param, UserCate1RecallStrategy.PUSH_FORM, setVideo, rovRecallRank);
-        // -------------------user cate2------------------
-        RecallUtils.extractRecall(mergeWeight.getOrDefault("cate2RecallN", 5.0).intValue(), param, UserCate2RecallStrategy.PUSH_FORM, setVideo, rovRecallRank);
-        // -------------------head province cate1------------------
-        RecallUtils.extractRecall(mergeWeight.getOrDefault("headCate1RecallN", 3.0).intValue(), param, HeadProvinceCate1RecallStrategy.PUSH_FORM, setVideo, rovRecallRank);
-        // -------------------head province cate2------------------
-        RecallUtils.extractRecall(mergeWeight.getOrDefault("headCate2RecallN", 3.0).intValue(), param, HeadProvinceCate2RecallStrategy.PUSH_FORM, setVideo, rovRecallRank);
-        //-------------------head cate2 of rovn------------------
-        RecallUtils.extractRecall(mergeWeight.getOrDefault("headCate2Rov", 5.0).intValue(), param, HeadCate2RovRecallStrategy.PUSH_FROM, setVideo, rovRecallRank);
-        //-------------------city rovn------------------
-        RecallUtils.extractRecall(mergeWeight.getOrDefault("cityRov", 5.0).intValue(), param, CityRovnRecallStrategy.PUSH_FROM, setVideo, rovRecallRank);
-        //-------------------priori province rovn------------------
-        RecallUtils.extractRecall(mergeWeight.getOrDefault("prioriProvinceRov", 3.0).intValue(), param, PrioriProvinceRovnRecallStrategy.PUSH_FROM, setVideo, rovRecallRank);
-        //-------------------priori province str------------------
-        RecallUtils.extractRecall(mergeWeight.getOrDefault("prioriProvinceStr", 1.0).intValue(), param, PrioriProvinceStrRecallStrategy.PUSH_FROM, setVideo, rovRecallRank);
-        //-------------------priori province ros------------------
-        RecallUtils.extractRecall(mergeWeight.getOrDefault("prioriProvinceRos", 1.0).intValue(), param, PrioriProvinceRosRecallStrategy.PUSH_FROM, setVideo, rovRecallRank);
-        //-------------------return1 cate2 ros------------------
-        RecallUtils.extractRecall(mergeWeight.getOrDefault("return1Cate2Ros", 5.0).intValue(), param, Return1Cate2RosRecallStrategy.PUSH_FORM, setVideo, rovRecallRank);
-        //-------------------return1 cate2 str------------------
-        RecallUtils.extractRecall(mergeWeight.getOrDefault("return1Cate2Str", 5.0).intValue(), param, Return1Cate2StrRecallStrategy.PUSH_FORM, setVideo, rovRecallRank);
-
-        RecallUtils.extractRecall(mergeWeight.getOrDefault("yearShareCate1", 5.0).intValue(), param, YearShareCate1RecallStrategy.PUSH_FROM, setVideo, rovRecallRank);
-        RecallUtils.extractRecall(mergeWeight.getOrDefault("yearShareCate2", 5.0).intValue(), param, YearShareCate2RecallStrategy.PUSH_FROM, setVideo, rovRecallRank);
-        RecallUtils.extractRecall(mergeWeight.getOrDefault("yearReturnCate2", 5.0).intValue(), param, YearReturnCate2RecallStrategy.PUSH_FROM, setVideo, rovRecallRank);
+
+        // ============================================================
+        // V536 实验：统一粗排分截断 (个性化 / 非个性化 两配额, 动态补足)
+        // 总配额 coarseRankTopN，个性化占 personalRatio。先个性化按上限抢位，
+        // 个性化不足时剩余名额转给非个性化，保证精排算力满载。
+        //
+        // 粗排分 = alg_vid_recommend_exp_feature_20250212.rovn_1h / rovn_24h 平均
+        // ============================================================
+        int totalTopN = mergeWeight.getOrDefault("coarseRankTopN", 80.0).intValue();
+        double personalRatio = mergeWeight.getOrDefault("personalRatio", 0.4);
+        int personalTopN = (int) Math.round(totalTopN * personalRatio);
+        Map<Long, Double> coarseRankMap = fetchCoarseRankScores(param);
+
+        int personalCandidates = RecallUtils.countDistinctCandidates(param, setVideo, PERSONAL_RECALL_PUSH_FROMS);
+        int sizeBeforePersonal = rovRecallRank.size();
+        RecallUtils.extractAllAndTruncateByCoarseRank(personalTopN, param, setVideo, rovRecallRank, coarseRankMap, PERSONAL_RECALL_PUSH_FROMS);
+        int personalActual = rovRecallRank.size() - sizeBeforePersonal;
+        int nonPersonalBudget = totalTopN - personalActual;  // 个性化不足时, 名额转给非个性化
+        int nonPersonalCandidates = RecallUtils.countDistinctCandidates(param, setVideo, NON_PERSONAL_RECALL_PUSH_FROMS);
+        int sizeBeforeNonPersonal = rovRecallRank.size();
+        RecallUtils.extractAllAndTruncateByCoarseRank(nonPersonalBudget, param, setVideo, rovRecallRank, coarseRankMap, NON_PERSONAL_RECALL_PUSH_FROMS);
+        int nonPersonalActual = rovRecallRank.size() - sizeBeforeNonPersonal;
+        log.info("coarse_rank_summary exp=536 quota={} pc={} ps={} nc={} ns={}",
+                totalTopN, personalCandidates, personalActual, nonPersonalCandidates, nonPersonalActual);
 
         // 记录召回源中的视频
         this.rankBeforePostProcessor(rovRecallRank);
@@ -308,80 +333,74 @@ public class RankStrategy4RegionMergeModelV536 extends RankStrategy4RegionMergeM
     }
 
     /**
-     * V536 fusion: 只保留"流量池相关 + 兜底相关"逻辑
-     *   1. rov 空兜底：rov 池为空时流量池直接顶上 (Basic 段 1)
-     *   7. 流量池按比例强插：topK 头部锁 rov + topK..size 按 flowPoolP / newFlowPoolSelectRate 概率门
-     *      混入 flowVideos / douHotFlowPoolVideos，否则用 rov 中段；一侧用光时另一侧兜底回填 (Basic 段 7)
+     * V536 实验：拉取粗排分（按 vid → score 返回）。
+     *
+     * 数据源：alg_vid_recommend_exp_feature_20250212。
+     * 表里没有现成 rovn 字段，需要从原子字段 (return_n_uv_*, exp_*) 用 plusSmooth 算出来。
+     * 公式 = FeatureV6.oneTypeStatFeature 同口径：rovn = plusSmooth(return_n_uv, exp, plus, 1)
+     * 默认 plus=30 与 FeatureV6.largerSmoothPlus 对齐，AB 对比不会因口径不同污染结论。
+     *
+     * Apollo 可调维度：
+     *   - coarseRovn1hW / coarseRovn24hW：1h 和 24h 的加权（默认 0.5/0.5）
+     *   - coarseRovn1hSmoothPlus / coarseRovn24hSmoothPlus：贝叶斯平滑系数（默认 30/30）
      *
-     * 删除（相对 Basic）：
-     *   - 段 2: Apollo 实验参数解析 (filterRules / rankReduceConfig) — 解析出来只给 boost/filter 用，不需要了
-     *   - 段 3: 标签读取 RankExtractorItemTags
-     *   - 段 4: 时间卡控 RankProcessorTagFilter
-     *   - 段 5: rov 池提权 RankProcessorBoost.boostByTag
-     *   - 段 6: rov 池强插 RankProcessorInsert.insertByTag + 品类降权 boostByMergeCate + 节日降权 boostByFestive
-     *   - 段 8: 密度控制 RankProcessorDensity.mergeDensityControl
+     * 缺失自动归一化：单值缺失时剩下的撑起全部权重；两值都缺失则 caller 兜底 RovScore。
      */
-    @Override
-    public RankResult mergeAndSort(RankParam param, List<Video> rovVideos, List<Video> flowVideos, List<Video> douHotFlowPoolVideos) {
-
-        // 1 兜底策略，rov池子不足时，用冷启池填补。直接返回。
-        if (CollectionUtils.isEmpty(rovVideos)) {
-            if (param.getSize() < flowVideos.size()) {
-                return new RankResult(flowVideos.subList(0, param.getSize()));
-            } else {
-                return new RankResult(flowVideos);
-            }
+    private Map<Long, Double> fetchCoarseRankScores(RankParam param) {
+        if (param == null || param.getRecallResult() == null
+                || CollectionUtils.isEmpty(param.getRecallResult().getData())) {
+            return Collections.emptyMap();
         }
-
-        // 7 流量池按比例强插
-        FunnelContext funnelCtx = param.getFunnelContext();
-        List<Video> result = new ArrayList<>();
-        for (int i = 0; i < param.getTopK() && i < rovVideos.size(); i++) {
-            result.add(rovVideos.get(i));
-        }
-        double flowPoolP = getFlowPoolP(param);
-        int flowPoolIndex = 0;
-        int rovPoolIndex = param.getTopK();
-        for (int i = 0; i < param.getSize() - param.getTopK(); i++) {
-            double rand = RandomUtils.nextDouble(0, 1);
-            if (rand < flowPoolP) {
-                if (flowPoolIndex < flowVideos.size()) {
-                    Video v = flowVideos.get(flowPoolIndex++);
-                    result.add(v);
-                    markColdStartInserted(funnelCtx, v);
-                } else {
-                    break;
-                }
-            } else if (this.isInsertDouHotFlowPoolVideo()) {
-                if (flowPoolIndex < douHotFlowPoolVideos.size()) {
-                    Video v = douHotFlowPoolVideos.get(flowPoolIndex++);
-                    result.add(v);
-                    markColdStartInserted(funnelCtx, v);
-                } else {
-                    break;
-                }
-            } else {
-                if (rovPoolIndex < rovVideos.size()) {
-                    result.add(rovVideos.get(rovPoolIndex++));
-                } else {
-                    break;
-                }
-            }
-        }
-        if (rovPoolIndex >= rovVideos.size()) {
-            for (int i = flowPoolIndex; i < flowVideos.size() && result.size() < param.getSize(); i++) {
-                Video v = flowVideos.get(i);
-                result.add(v);
-                markColdStartInserted(funnelCtx, v);
-            }
-        }
-        if (flowPoolIndex >= flowVideos.size()) {
-            for (int i = rovPoolIndex; i < rovVideos.size() && result.size() < param.getSize(); i++) {
-                result.add(rovVideos.get(i));
-            }
+        Map<String, Double> mergeWeight = this.mergeWeight != null ? this.mergeWeight : Collections.emptyMap();
+        double w1h = mergeWeight.getOrDefault("coarseRovn1hW", 0.5);
+        double w24h = mergeWeight.getOrDefault("coarseRovn24hW", 0.5);
+        double plus1h = mergeWeight.getOrDefault("coarseRovn1hSmoothPlus", 30.0);
+        double plus24h = mergeWeight.getOrDefault("coarseRovn24hSmoothPlus", 30.0);
+        // 只对参与统一截断的 23 路 vid 拉粗排分（跳过流量池 3 路，省 proto + RPC 延迟）
+        List<String> vids = param.getRecallResult().getData().stream()
+                .filter(d -> d != null && CollectionUtils.isNotEmpty(d.getVideos()))
+                .filter(d -> ALL_ROV_PUSH_FROMS.contains(d.getPushFrom()))
+                .flatMap(d -> d.getVideos().stream())
+                .map(v -> String.valueOf(v.getVideoId()))
+                .distinct()
+                .collect(Collectors.toList());
+        if (vids.isEmpty()) return Collections.emptyMap();
+
+        Map<String, Map<String, Map<String, String>>> feats = featureService.getVideoCoarseRankFeature(vids);
+        Map<Long, Double> result = new HashMap<>(vids.size());
+        for (String vid : vids) {
+            Map<String, String> row = feats.getOrDefault(vid, Collections.emptyMap())
+                    .getOrDefault("alg_vid_recommend_exp_feature_20250212", Collections.emptyMap());
+            Double rovn1h = computeRovn(row, "1h", plus1h);
+            Double rovn24h = computeRovn(row, "24h", plus24h);
+            // 加权平均，缺失自动归一化
+            double sumW = (rovn1h != null ? w1h : 0) + (rovn24h != null ? w24h : 0);
+            if (sumW <= 0) continue;
+            double sumWS = (rovn1h != null ? rovn1h * w1h : 0) + (rovn24h != null ? rovn24h * w24h : 0);
+            try {
+                result.put(Long.parseLong(vid), sumWS / sumW);
+            } catch (NumberFormatException ignore) { }
         }
+        return result;
+    }
+
+    /**
+     * 与 FeatureV6.oneTypeStatFeature 同口径：rovn = plusSmooth(return_n_uv, exp, plus, 1)
+     *
+     * 字段语义（区分 0 vs null）：
+     *   - exp 是 period 有效性 anchor：null 或 ≤0 → 整个 period 无效（return null）
+     *   - return_n_uv 缺失视为 0（真实信号"无回访"）：rovn=0，参与加权（不会让另一时段兜底）
+     */
+    private static Double computeRovn(Map<String, String> row, String period, double smoothPlus) {
+        Double exp = parseDoubleOrNull(row.get("exp_" + period));
+        if (exp == null || exp <= 0) return null;
+        Double returnNuv = parseDoubleOrNull(row.get("return_n_uv_" + period));
+        return FeatureUtils.plusSmooth(returnNuv != null ? returnNuv : 0, exp, smoothPlus, 1);
+    }
 
-        return new RankResult(result);
+    private static Double parseDoubleOrNull(String s) {
+        if (StringUtils.isBlank(s)) return null;
+        try { return Double.parseDouble(s); } catch (NumberFormatException e) { return null; }
     }
 
     private UserShareReturnProfile parseUserProfile(Map<String, Map<String, String>> userOriginInfo) {

recommend-server-service/src/main/java/com/tzld/piaoquan/recommend/server/service/rank/strategy/RankStrategy4RegionMergeModelV569.java

@@ -37,7 +37,9 @@ public class RankStrategy4RegionMergeModelV569 extends RankStrategy4RegionMergeM
     private FeatureService featureService;
 
     /**
-     * V569 个性化召回白名单 (6 路)：召回 key 含 mid/uid，依赖该用户行为信号。
+     * V569 个性化召回白名单 (7 路: V566 基础 6 路 + 1 路 dk_elements 行为路实验)：召回 key 含 mid/uid，
+     * 依赖该用户行为信号。
+     * V569 实验路径: YearShareDkElements (用户近期 share 行为 join dk_elements)
      * 注：YearReturnCate2 因线上效果不佳, 2026-06-04 起移到非个性化白名单。
      */
     private static final Set<String> PERSONAL_RECALL_PUSH_FROMS = new HashSet<>(Arrays.asList(
@@ -46,7 +48,8 @@ public class RankStrategy4RegionMergeModelV569 extends RankStrategy4RegionMergeM
             Return1Cate2RosRecallStrategy.PUSH_FORM,
             Return1Cate2StrRecallStrategy.PUSH_FORM,
             YearShareCate1RecallStrategy.PUSH_FROM,
-            YearShareCate2RecallStrategy.PUSH_FROM
+            YearShareCate2RecallStrategy.PUSH_FROM,
+            YearShareDkElementsRecallStrategy.PUSH_FROM
     ));
 
     /**