diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 9c7a6d0..c057859 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -16,8 +16,20 @@ jobs:
         with:
           python-version: '3.11'
           cache: pip
-      - run: python -m pip install -e '.[test]' build
-      - run: python -m pytest -q tests
+      - run: python -m pip install -e '.[test]' build
+      - name: Download recommendation CLI
+        env:
+          GH_TOKEN: ${{ github.token }}
+        run: |
+          mkdir -p "$RUNNER_TEMP/deck/bin"
+          gh release download v0.0.3 --repo empty-sekai/ournotes-deck \
+            --pattern ournotes-deck-v0.0.3-x86_64-unknown-linux-musl.tar.gz \
+            --pattern SHA256SUMS --dir "$RUNNER_TEMP/deck"
+          cd "$RUNNER_TEMP/deck"
+          sha256sum --check --ignore-missing SHA256SUMS
+          tar -xzf ournotes-deck-v0.0.3-x86_64-unknown-linux-musl.tar.gz --strip-components=1 -C bin
+          echo "OURNOTES_DECK_BIN=$RUNNER_TEMP/deck/bin/ournotes-deck" >> "$GITHUB_ENV"
+      - run: python -m pytest -q -rs tests
       - run: python scripts/validate_contract.py scan-request schemas/examples/scan-request.json
       - run: python scripts/validate_contract.py scan-response schemas/examples/scan-response.json
       - run: python -m build
diff --git a/MANIFEST.in b/MANIFEST.in
index 60506f1..90fc20e 100644
--- a/MANIFEST.in
+++ b/MANIFEST.in
@@ -1,2 +1,3 @@
 include docs/DECK_PIPELINE.md
 include CHANGELOG.md
+recursive-include tests/fixtures *.json
diff --git a/README.md b/README.md
index 418dd2f..d011736 100644
--- a/README.md
+++ b/README.md
@@ -48,10 +48,10 @@ python boxvision.py scan --data DATA --threads 2 --output ./result screenshot-a.
 可选的正式 `deck-ui` 入口将截图识别、显式人工补齐和 [OurNotes Deck](https://github.com/empty-sekai/ournotes-deck) 的真实搜索连在一起：
 
 ```powershell
-python boxvision.py deck-ui --data DATA --deck-data deck-data.json --solver-bin ournotes-recommend.exe --workspace ./local/deck-ui --port 18790 --box-port 18793
+python boxvision.py deck-ui --data DATA --deck-data deck-data.json --solver-bin ournotes-deck.exe --workspace ./local/deck-ui --port 18790 --box-port 18793
 ```
 
-打开 `http://127.0.0.1:18790/`。需要与图库匹配的 DeckData 和支持 `ournotes-deck.recommendation-request/1` 的推荐程序；源码不附带游戏资源。未知字段不会默认填为等级／Rank 1，完整持有集合须由使用者声明。详细补齐契约、mock 复现、版本校验和目标语义见 [组卡输入管线](docs/DECK_PIPELINE.md)。
+打开 `http://127.0.0.1:18790/`。需要与图库匹配的 DeckData 和 ournotes-deck v0.0.3（`recommend` 子命令、`ournotes-deck.search-request/1` 请求）；源码不附带游戏资源。未知字段不会默认填为等级／Rank 1，完整持有集合须由使用者声明。详细补齐契约、mock 复现、版本校验和目标语义见 [组卡输入管线](docs/DECK_PIPELINE.md)。
 
 ## 数据和模型
 
diff --git a/bdon_vision/deck/recommend.py b/bdon_vision/deck/recommend.py
index 88c8d3d..035db24 100644
--- a/bdon_vision/deck/recommend.py
+++ b/bdon_vision/deck/recommend.py
@@ -21,14 +21,12 @@ def scene_matrix(deck_path):
     shortest=min(d['charts'],key=lambda c:len(c['notes']['id']))
     sid=shortest['scoreId']
     base=next(r for r in rows(d,'MasterLiveMusic') if sid in [r.get('_easyID'),r.get('_normalID'),r.get('_hardID'),r.get('_expertID')])
-    seed={'atoms':[[1,1],[-1,1],[42,1]],'provenance':'Declared illustrative mock finite native root law; not the real TickCount/server distribution'}
     specs=[]
     def add(id_, execution, scenario=None, metric=None, expected='supported'):
-        req={'format':'ournotes-deck.recommendation-request/1','execution':execution,'metric':metric or {'kind':'score'},
+        req={'format':'ournotes-deck.search-request/1','execution':execution,'metric':metric or {'kind':'score'},
             'constraints':{'noSnaps':False},'k':3,'strategy':{'kind':'exhaustive'},
             'limits':{'timeLimitMs':30000,'maxCandidates':20000,'cacheEntries':2048}}
         if scenario: req['scenario']=scenario
-        if execution['kind']=='live': req['seedLaw']=seed
         specs.append({'id':id_,'request':req,'expected':expected})
     add('songless-power',{'kind':'power'},metric={'kind':'power'})
     for mode in ['free','mission','battle']:
@@ -114,7 +112,7 @@ def ticks(value):
             entry['request']['strategy']={'kind':'candidate','powerSeeds':2,'proposals':600,'proposalSeed':20261001}
             entry['request']['limits']={'timeLimitMs':10000,'maxCandidates':1000,'cacheEntries':1024}
     return {'schema':'ournotes.integration-scene-matrix/1','owner':'BoxLens optional Deck recommendation workflow',
-        'deckDataSha256':digest(deck_path),'entries':specs,'liveIsSeparate':True,'capabilitySource':'ournotes-deck.recommendation-request/1','randomScope':'Explicit finite native root law and theoretical play; unknown real native TickCount distribution','unsupportedScope':'Battle/Arena require explicitly provided aggregate peer confirmations. Conditional mock Battle cases are separate; current JP has no Arena master rows'}
+        'deckDataSha256':digest(deck_path),'entries':specs,'liveIsSeparate':True,'capabilitySource':'ournotes-deck.search-request/1','randomScope':'Solver-reported probability law and theoretical-best play','unsupportedScope':'Battle/Arena require explicitly provided aggregate peer confirmations; available scenes depend on the bound regional master'}
 
 def recommend(binary, deck_path, roster_path, matrix, out):
     out=Path(out);out.mkdir(parents=True,exist_ok=True)
@@ -141,7 +139,7 @@ def recommend(binary, deck_path, roster_path, matrix, out):
         if 'request' not in entry:
             result.append({'id':id_,'state':'Unavailable','reason':entry.get('reason'),'complete':False});continue
         request=entry['request'];reqfile=out/(id_+'-request.json');write_request(entry,reqfile)
-        command=[str(binary),'--data',str(frozen_deck),'--roster',str(frozen_roster),'--request',str(reqfile)]
+        command=[str(binary),'recommend','--data',str(frozen_deck),'--roster',str(frozen_roster),'--request',str(reqfile)]
         t=time.perf_counter()
         try:
             p=subprocess.run(command,capture_output=True,encoding='utf-8',errors='replace',timeout=min(600,max(180,(request.get('limits',{}).get('timeLimitMs') or 165000)/1000+15)))
diff --git a/bdon_vision/deck/web.html b/bdon_vision/deck/web.html
index 1efebbf..4c86cdb 100644
--- a/bdon_vision/deck/web.html
+++ b/bdon_vision/deck/web.html
@@ -9,7 +9,7 @@
 <section class="panel active" id="input"><div class="grid"><div class="card"><h2>截图输入与可选 mock 样本</h2><p>复用原始原生卡片 UI 生成器，含跨页重复、隐藏字段、降质与等级冲突。</p><div class="gallery" id="gallery"></div></div><div><div class="card"><h2>运行图片识别</h2><p>图片送到本机现有 BoxLens。识别器只接收图片、图库与模型。</p><input id="images" type="file" accept="image/*" multiple><label><input id="mock" type="checkbox"> 输入是声明的 mock 截图</label><div><button class="action" id="scan">识别所选截图</button></div><p><a id="boxlenslink" href="http://127.0.0.1:18781/" target="_blank">打开 BoxLens 原始校对界面 ↗</a></p><div class="note">截图只观察可见卡片；技能等级、角色 Rank、道具、VIP、回忆和完整持有集合需要单独提供。</div></div><div class="card"><h2>来源与版本</h2><div id="binding"></div><p><a href="/files/fixtures/fixture-manifest.json" target="_blank">生成器与截图真值清单 ↗</a></p></div></div></div></section>
 <section class="panel" id="inventory"><div class="card"><div class="row"><h2>截图实际读到的字段</h2><span class="badge">观察集合 · 保留 unknown</span></div><div id="rawtable"></div><div class="note">同一卡片的等级出现两个值时，合并值保留为未知，冲突证据完整保留。隐藏字段与降质拒识也保留未知。</div><details><summary>查看原始 BoxLens JSON</summary><pre id="rawjson"></pre></details><a href="/files/scan/observations.json" target="_blank">逐图观察与模型来源 ↗</a></div></section>
 <section class="panel" id="review"><div class="grid"><div class="card"><h2>已明确补齐的 Roster</h2><p>以下培养值来自识别或明确补充；技能与玩家状态的来源单独记录。</p><div id="rostertable"></div><div id="corrections"></div><details><summary>完整现有 Roster JSON</summary><pre id="rosterjson"></pre></details></div><div><div class="card"><h2>补充缺失信息</h2><p>演示真值与识别过程分开，点击后才用于补齐。真实账号请上传声明来源的人工补充文件。</p><button id="complete" class="action">采用声明的 mock 真值</button><input id="completionfile" type="file" accept="application/json"><button id="manual" class="action secondary">校验人工补充 JSON</button><p><a href="/files/fixtures/completion.json" target="_blank">查看补充契约样例 ↗</a></p><div id="player"></div><div class="note success">只有必要字段、持有集合和版本绑定全部校验通过，才导出 Roster 并进入搜索。</div></div></div></div></section>
-<section class="panel" id="results"><div class="card"><div class="row"><div><h2>推荐队伍</h2><p id="scope">声明有限根种子分布下的 native Live；技能顺序由原生五槽洗牌决定。</p></div><span class="badge status" id="completionstatus"></span></div><div class="row"><div><div class="caption" id="metriclabel">目标指标</div><div class="resultmetric" id="metric"></div></div><div id="resultmeta"></div></div><div class="slots" id="slots"></div><div id="unselected" class="note success"></div><div class="note" id="modelscope">Complete 的最优性限定于请求中的卡池、约束、打法、有限 seedLaw 和模型范围。真实 TickCount / 服务器根种子总体分布未知。</div><details><summary>各根种子的技能顺序、分数与精确目标值</summary><pre id="outcomes"></pre></details></div><div class="card"><h2>场景矩阵与状态</h2><p>Live、Skip 与综合力分别验收。不可用、未支持、超时和候选搜索状态按真实返回展示。</p><select id="caseSelect"></select><button id="recommend" class="action">运行当前场景</button><input id="requestfile" type="file" accept="application/json"><button id="customrecommend" class="action secondary">运行明确提供的场景请求</button><p>活动指标、外部排名时间线与打法可由请求文件明确提供；搜索会校验实际 master 与支持范围。</p><div id="casetable"></div><details><summary>查看完整搜索结果与请求</summary><pre id="recommendjson"></pre></details></div></section>
+<section class="panel" id="results"><div class="card"><div class="row"><div><h2>推荐队伍</h2><p id="scope">演出结果采用求解器返回的概率模型与请求中的打法。</p></div><span class="badge status" id="completionstatus"></span></div><div class="row"><div><div class="caption" id="metriclabel">目标指标</div><div class="resultmetric" id="metric"></div></div><div id="resultmeta"></div></div><div class="slots" id="slots"></div><div id="unselected" class="note success"></div><div class="note" id="modelscope">只有 Complete 且 optimality=proven 表示请求范围内的排名已经证明；概率模型与计算条件见完整结果。</div><details><summary>队伍分数与精确目标值</summary><pre id="outcomes"></pre></details></div><div class="card"><h2>场景矩阵与状态</h2><p>Live、Skip 与综合力分别验收。不可用、未支持、超时和候选搜索状态按真实返回展示。</p><select id="caseSelect"></select><button id="recommend" class="action">运行当前场景</button><input id="requestfile" type="file" accept="application/json"><button id="customrecommend" class="action secondary">运行明确提供的场景请求</button><p>活动指标、外部排名时间线与打法可由请求文件明确提供；搜索会校验实际 master 与支持范围。</p><div id="casetable"></div><details><summary>查看完整搜索结果与请求</summary><pre id="recommendjson"></pre></details></div></section>
 <div id="message" aria-live="polite"></div><footer>图片、识别、校对和搜索在本机运行 · <a id="downloadRoster" href="/files/roster.json">下载 Roster</a> · <a id="downloadAdaptation" href="/files/adaptation.json">审查补齐与纠错来源</a> · <a href="/files/baseline-live-run.json">查看实际命令与耗时</a></footer></main>
 <script>
 const $=s=>document.querySelector(s),esc=s=>String(s??'—').replace(/[&<>"']/g,c=>({'&':'&amp;','<':'&lt;','>':'&gt;','"':'&quot;',"'":'&#39;'}[c]));
@@ -24,7 +24,7 @@
 function stamp(){return {input:inputVersion,request:requestVersion,scene:sceneVersion,session:sessionId};}
 function current(t,scene=false){return t.input===inputVersion&&t.request===requestVersion&&t.session===sessionId&&(!scene||t.scene===sceneVersion);}
 function fractionText(f,percent=false){try{let n=BigInt(f.numerator),d=BigInt(f.denominator);if(d===0n)return '—';if(percent)n*=100n;let neg=n<0n;n=neg?-n:n;let cents=(n*100n+d/2n)/d;return (neg?'-':'')+(cents/100n).toLocaleString('zh-CN')+'.'+String(cents%100n).padStart(2,'0');}catch{return '见精确值';}}
-function metricSemantic(kind,execution,request){const live=execution?.kind==='live',skip=execution?.kind==='skip',play=request?.execution?.play?.kind==='stream'?'明确提供的判定序列':'理论最优打法';if(kind==='power')return {label:'综合力',unit:'综合力',scope:'按当前成员、培养状态、留影与场景加成计算综合力。',conditional:false};if(kind==='scoreAtLeast'&&skip)return {label:'跳过分数达标结果 · 分数 ≥ '+(request?.metric?.threshold??'请求阈值'),unit:'%',scope:'按指定跳过分数和阈值计算确定性达标结果（0% 或 100%）。',conditional:false};if(kind==='scoreAtLeast')return {label:'演出达标概率 · 分数 ≥ '+(request?.metric?.threshold??'请求阈值'),unit:'%',scope:'在声明的有限根种子分布与打法下，估计演出分数达到阈值的概率。',conditional:true};if(kind==='clientEventPoints')return {label:live?'期望活动收益':'跳过活动收益',unit:'活动 pt',scope:'根据明确提供的活动时间、计数与当前 master 计算客户端活动收益'+(live?'，演出随机性限定于声明的有限 seedLaw。':'。'),conditional:live};if(kind==='conditionalClientEventItems')return {label:live?'期望目标道具收益':'跳过目标道具收益',unit:'个目标道具',scope:'以明确选择的奖励行、资源身份及活动计数为条件'+(live?'，演出随机性限定于声明的有限 seedLaw。':'计算客户端道具数量。'),conditional:live};return {label:skip?'跳过分数':'期望演出分数',unit:'分',scope:skip?'按当前编成与指定谱面计算跳过分数。':'在声明的有限根种子分布与'+play+'下计算演出期望分数；技能顺序采用原生五槽洗牌。',conditional:live};}
+function metricSemantic(kind,execution,request){const live=execution?.kind==='live',skip=execution?.kind==='skip',play=request?.execution?.play?.kind==='stream'?'明确提供的判定序列':'理论最优打法';if(kind==='power')return {label:'综合力',unit:'综合力',scope:'按当前成员、培养状态、留影与场景加成计算综合力。',conditional:false};if(kind==='scoreAtLeast'&&skip)return {label:'跳过分数达标结果 · 分数 ≥ '+(request?.metric?.threshold??'请求阈值'),unit:'%',scope:'按指定跳过分数和阈值计算确定性达标结果（0% 或 100%）。',conditional:false};if(kind==='scoreAtLeast')return {label:'演出达标概率 · 分数 ≥ '+(request?.metric?.threshold??'请求阈值'),unit:'%',scope:'按求解器的概率模型与声明打法计算演出分数达到阈值的概率。',conditional:true};if(kind==='clientEventPoints')return {label:live?'期望活动收益':'跳过活动收益',unit:'活动 pt',scope:'根据明确提供的活动时间、计数与当前 master 计算客户端活动收益'+(live?'，演出随机性采用求解器返回的概率模型。':'。'),conditional:live};if(kind==='conditionalClientEventItems')return {label:live?'期望目标道具收益':'跳过目标道具收益',unit:'个目标道具',scope:'以明确选择的奖励行、资源身份及活动计数为条件'+(live?'，演出随机性采用求解器返回的概率模型。':'计算客户端道具数量。'),conditional:live};return {label:skip?'跳过分数':'期望演出分数',unit:'分',scope:skip?'按当前编成与指定谱面计算跳过分数。':'按求解器的概率模型与'+play+'下计算演出期望分数；技能顺序采用原生五槽洗牌。',conditional:live};}
 function errorText(c){const e=c?.raw?.error;if(typeof e==='string')return e;if(e)return [e.code||e.kind,e.message].filter(Boolean).join('：');return c?.reason||c?.state||'当前请求没有可展示队伍';}
 function render(){$('#complete').disabled=!state.hasMockTruth;$('#binding').innerHTML='';if(state.boxLensUrl)$('#boxlenslink').href=state.boxLensUrl;let b=state.scan||state.inventory?.box,r=state.roster,m=state.manifest,a=state.adaptation;
 if(b){$('#stats').innerHTML=[['观察身份',b.unique_count],['逐图观察',b.observations],['重复观察',b.duplicate_observations],['冲突卡片',b.cards.filter(x=>Object.keys(x.conflicts).length).length]].map(([label,n])=>'<div class="stat"><strong>'+n+'</strong><span>'+label+'</span></div>').join('');$('#rawtable').innerHTML=table(['身份','等级','卡阶','特训','证据'],b.cards.map(c=>[esc(c.kind+' #'+c.id),...['level','card_rank','awake_count'].map(f=>c.conflicts[f]?'<span class="warn">未知 · '+esc(Object.keys(c.conflicts[f]).join(' / '))+'</span>':esc(c[f].value??'unknown')),esc(c.sources.length+' 个来源')]));$('#rawjson').textContent=JSON.stringify(b,null,2);}else{$('#stats').textContent='等待新截图识别';$('#rawtable').textContent='';$('#rawjson').textContent='';}
diff --git a/docs/DECK_PIPELINE.md b/docs/DECK_PIPELINE.md
index ad0373b..586d74f 100644
--- a/docs/DECK_PIPELINE.md
+++ b/docs/DECK_PIPELINE.md
@@ -4,10 +4,10 @@ BoxLens 的可选 `deck-ui` 入口调用现有识别器和独立的 ournotes-dec
 
 ## 启动
 
-安装本仓库后，准备自己的完整视觉 `DATA`、同区服 DeckData 和支持 v1 推荐接口的程序：
+安装本仓库后，准备自己的完整视觉 `DATA`、同区服 DeckData 和 ournotes-deck v0.0.3 推荐程序：
 
 ```powershell
-python boxvision.py deck-ui --data DATA --deck-data deck-data.json --solver-bin ournotes-recommend.exe --workspace ./local/deck-ui --port 18790 --box-port 18793
+python boxvision.py deck-ui --data DATA --deck-data deck-data.json --solver-bin ournotes-deck.exe --workspace ./local/deck-ui --port 18790 --box-port 18793
 ```
 
 浏览器打开 `http://127.0.0.1:18790/`。两项服务仅绑定回环地址，端口占用会拒绝启动，可改为其他独立端口。空 workspace 可直接上传自己的截图；也可使用明确生成的 mock workspace。使用 Ctrl+C 关闭时，界面会终止自己启动的识别子进程。
@@ -65,21 +65,21 @@ CLI 和 HTTP 补齐共用工作区审阅状态。补齐开始即写入 `needs-re
 python boxvision.py deck-mock --data DATA --deck-data deck-data.json --members 7 --snaps 3 --output ./local/mock/fixtures
 python boxvision.py scan --data DATA --threads 1 --output ./local/mock/scan ./local/mock/fixtures/01-level-a.png ./local/mock/fixtures/02-level-overlap.png ./local/mock/fixtures/03-training.png ./local/mock/fixtures/04-hidden.png ./local/mock/fixtures/05-snap.png ./local/mock/fixtures/06-conflict.png ./local/mock/fixtures/07-degraded-fields.png ./local/mock/fixtures/08-duplicate.png
 python boxvision.py deck-adapt --data DATA --deck-data deck-data.json --region jp --mock --box ./local/mock/scan/box.json --completion ./local/mock/fixtures/completion.json --output ./local/mock
-python boxvision.py deck-recommend --deck-data deck-data.json --solver-bin ournotes-recommend.exe --roster ./local/mock/roster.json --request request.json --output ./local/recommend-new
-python boxvision.py deck-ui --data DATA --deck-data deck-data.json --solver-bin ournotes-recommend.exe --workspace ./local/mock
+python boxvision.py deck-recommend --deck-data deck-data.json --solver-bin ournotes-deck.exe --roster ./local/mock/roster.json --request request.json --output ./local/recommend-new
+python boxvision.py deck-ui --data DATA --deck-data deck-data.json --solver-bin ournotes-deck.exe --workspace ./local/mock
 ```
 
-每次生成和推荐使用新的输出目录，避免覆盖证据。三留影布局包含可见边界断言。推荐需要与 Roster 相邻的 `adaptation.json`，拒绝未经审阅的库存。请求文件遵循 solver 的 `ournotes-deck.recommendation-request/1`；可由界面选择预置场景或上传明确提供的请求。
+每次生成和推荐使用新的输出目录，避免覆盖证据。三留影布局包含可见边界断言。推荐需要与 Roster 相邻的 `adaptation.json`，拒绝未经审阅的库存。请求文件遵循 solver 的 `ournotes-deck.search-request/1`；可由界面选择预置场景或上传明确提供的请求。
 
 上传 JSON 在浏览器使用严格 UTF-8 解码并保留原文本传入 HTTP；服务拒绝顶层／嵌套／转义后重复键、NaN、Infinity、非法 UTF-8 和畸形 JSON。合法大整数、小数 token 与 CRLF 原样保存至实际 CLI 请求，浏览器不会先 JSON.parse 再 stringify 将它们丢失。
 
 ## 目标与完成状态
 
-综合力按综合力单位显示，不附加随机根种子条件。演出期望、阈值概率和活动指标显示各自语义与有限 seedLaw／声明打法／模型条件。Complete 表示程序确实穷尽了请求范围，最优性限定于该卡池、约束、外部输入、打法与声明根分布；未知现实根种子分布下的无条件最优不在结论中。Skip、综合力和 Live 分别展示，不能相互代替演出验收。
+综合力按综合力单位显示。演出期望、阈值概率和活动指标采用求解器返回的 `probabilityLaw`、请求中的打法和场景条件。推荐结果保留 `completion`、`optimality` 和 `exitReason`；只有 `Complete` 且 `optimality=proven` 表示请求范围内的排名已经证明。候选搜索和未完成搜索按实际返回状态展示。
 
 Mission／Battle／Arena 强制撃奏；Battle／Arena 多人结果需要显式外部排名确认到达时间线，不能自动视为 Solo 第一。活动收益使用声明的活动时钟和结果上下文，服务器所选奖励未知时不能声称现实收益保证。
 
-独立 QA 的最终场景矩阵为 **45 PASS / 0 FAIL / 2 BLOCKED**，两项 blocked 因冻结 JP master 没有 Arena 数据；不表示 Arena 已在当前 JP 原生场景实测。synthetic Arena 模型检查单独保留。采用的 R3 推荐程序 SHA256 为 `d645f10a2e23efbdf1ef722f49892a26829871a711e9bc79179006e4afd3a1ef`；已有 48,960 物理队伍 Top3、槽位、留影和逐根结果独立对拍证据，不在此次接入重跑长穷举。数值模型边界是离线 native 1.0.1-25，未据此声称 JP 1.0.4 运行时完全等价。
+CI 下载并校验 ournotes-deck v0.0.3 的发布包，用手写合成数据通过实际推荐入口检查综合力、Skip、Live 和结构化错误。自定义请求使用同一发布版的请求契约，求解器负责校验场景与计算条件。
 
 ## 仓库检查
 
@@ -88,4 +88,4 @@ python -m pip install -e ".[test]"
 python -m pytest -q tests
 ```
 
-公共回归使用手写 synthetic master 和无游戏图像的 JSON，不分发原始 master、chart、APK、截图、卡面、账号或内部协作文件。根目录 public-manifest.json 记录发布源码各文件的字节数和 SHA-256。
+公共回归使用手写 synthetic master 和无游戏图像的 JSON，不分发原始 master、chart、APK、截图、卡面、账号或内部协作文件。根目录 public-manifest.json 记录 v0.2.0 发布时的源码文件身份。
diff --git a/tests/fixtures/deck-synthetic.json b/tests/fixtures/deck-synthetic.json
new file mode 100644
index 0000000..b311979
--- /dev/null
+++ b/tests/fixtures/deck-synthetic.json
@@ -0,0 +1,67 @@
+{
+  "format": "nnnotes.deck-data/1",
+  "provenance": {"synthetic": true, "region": "synthetic", "master": {"version": "synthetic-1"}},
+  "master": {
+    "MasterMemberCard": {"columns": ["_cardType", "_characterID", "_id", "_memberCardAwakeGroup", "_memberCardLevelGroup", "_memberCardRankGroup", "_performancePowerMax", "_rarity", "_technicPowerMax", "_visualPowerMax"], "rows": [[1, 1, 1, 1, 1, 1, 100, 1, 100, 100], [1, 2, 2, 1, 1, 1, 100, 1, 100, 100], [1, 3, 3, 1, 1, 1, 100, 1, 100, 100], [1, 4, 4, 1, 1, 1, 100, 1, 100, 100], [1, 5, 5, 1, 1, 1, 100, 1, 100, 100]]},
+    "MasterSupportCard": {"columns": [], "rows": []},
+    "MasterMemberCardLevel": {"columns": ["_group", "_id", "_level", "_performanceRate", "_technicRate", "_visualRate"], "rows": [[1, 1, 1, 10000, 10000, 10000]]},
+    "MasterMemberCardLevelLimit": {"columns": [], "rows": []},
+    "MasterMemberCardAwake": {"columns": ["_awakeCount", "_group", "_id"], "rows": [[1, 1, 1]]},
+    "MasterMemberCardRank": {"columns": ["_group", "_id", "_rank"], "rows": [[1, 1, 1]]},
+    "MasterSupportCardLevel": {"columns": [], "rows": []},
+    "MasterSupportCardRank": {"columns": [], "rows": []},
+    "MasterCharacter": {"columns": ["_bandID", "_id"], "rows": [[1, 1], [1, 2], [1, 3], [1, 4], [1, 5]]},
+    "MasterCharacterRank": {"columns": ["_bonus", "_id", "_rank"], "rows": [[0, 1, 1]]},
+    "MasterCharacterTotalRank": {"columns": [], "rows": []},
+    "MasterBandItem": {"columns": [], "rows": []},
+    "MasterBandItemLevel": {"columns": [], "rows": []},
+    "MasterBandItemSkillEffect": {"columns": [], "rows": []},
+    "MasterSkillTarget": {"columns": [], "rows": []},
+    "MasterLeaderSkillEffect": {"columns": [], "rows": []},
+    "MasterSkillCondition": {"columns": [], "rows": []},
+    "MasterSkillConditionSet": {"columns": [], "rows": []},
+    "MasterSkillCumulativeCondition": {"columns": [], "rows": []},
+    "MasterVip": {"columns": [], "rows": []},
+    "MasterVipRankBonus": {"columns": [], "rows": []},
+    "MasterParameter": {"columns": ["_id", "_value"], "rows": [["music_type_base_bonus_rate", "400"], ["music_tag_base_bonus_rate", "300"], ["type_link_base_bonus_rate", "600"]]},
+    "MasterLiveSkill": {"columns": [], "rows": []},
+    "MasterGekisouSkill": {"columns": [], "rows": []},
+    "MasterLiveSkillEffect": {"columns": [], "rows": []},
+    "MasterLiveMusic": {"columns": ["_bestMusicTagIDs", "_expertID", "_gekisouMission1", "_gekisouMission2", "_gekisouMission3", "_id", "_musicType"], "rows": [[[], 1004, 1, 1, 1, 10, 1]]},
+    "MasterChallengeMusic": {"columns": [], "rows": []},
+    "MasterArenaMusic": {"columns": [], "rows": []},
+    "MasterLiveMusicScore": {"columns": ["_fullComboCount", "_id", "_musicScoreLevel"], "rows": [[1, 1004, 5]]},
+    "MasterLiveNoteParameter": {"columns": ["_id", "_noteOperateType", "_scorePercent"], "rows": [[1, 1, 100]]},
+    "MasterLiveJudgementParameter": {"columns": ["_damage", "_id", "_noteSimulateJudgement", "_scorePercent"], "rows": [[0, 1, 5, 100], [0, 2, 4, 80]]},
+    "MasterLiveComboScoreBonus": {"columns": [], "rows": []},
+    "MasterLiveSettings": {"columns": ["_id", "_key", "_value"], "rows": [[1, "note_score_adjustment_factor", "3"], [2, "note_score_life_onus_factor", "0.5"], [3, "life_base", "1000"], [4, "life_denger", "300"], [5, "gekisou_luck_gauge_max", "10"], [6, "gekisou_luck_gauge_max_rush", "10"], [7, "gekisou_luck_rush_score_bonus_percent", "10"]]},
+    "MasterMemoryMusicGroup": {"columns": [], "rows": []},
+    "MasterMemoryMusic": {"columns": [], "rows": []},
+    "MasterMemoryMusicBonus": {"columns": [], "rows": []},
+    "MasterMemoryMemberLevel": {"columns": [], "rows": []},
+    "MasterMemorySupportLevel": {"columns": [], "rows": []},
+    "MasterEventEffect": {"columns": [], "rows": []},
+    "MasterBand": {"columns": [], "rows": []},
+    "MasterEvent": {"columns": [], "rows": []},
+    "MasterEventAchievementReward": {"columns": [], "rows": []},
+    "MasterEventAchievementLoopReward": {"columns": [], "rows": []},
+    "MasterLiveEventReward": {"columns": [], "rows": []},
+    "MasterChallengeLiveEventReward": {"columns": [], "rows": []},
+    "MasterLiveEventPoint": {"columns": [], "rows": []},
+    "MasterChallengeLiveEventPoint": {"columns": [], "rows": []},
+    "MasterLiveChallengePoint": {"columns": [], "rows": []},
+    "MasterLiveMusicBoostBonus": {"columns": [], "rows": []},
+    "MasterChallengeMusicBoostBonus": {"columns": [], "rows": []},
+    "MasterLiveScoreRank": {"columns": [], "rows": []},
+    "MasterSupportSkillEffect": {"columns": [], "rows": []},
+    "MasterSkillEffectSetting": {"columns": [], "rows": []},
+    "MasterLiveJudgementTiming": {"columns": ["_afterMs", "_beforeMs", "_id", "_noteJudgementType", "_noteSimulateJudgement"], "rows": [[0, 0, 1, 1, 5]]},
+    "MasterGekisouSkillEffect": {"columns": [], "rows": []},
+    "MasterGekisouSupportSkill": {"columns": [], "rows": []},
+    "MasterGekisouSupportSkillEffect": {"columns": [], "rows": []},
+    "MasterLiveGekisouLuckBasePoint": {"columns": [], "rows": []},
+    "MasterLiveGekisouLuckBonusLot": {"columns": [], "rows": []},
+    "MasterLiveGekisouRankingScoreBonus": {"columns": ["_count", "_id", "_missionPattern", "_rank", "_scoreBonusPercent"], "rows": [[1, 11, 1, 1, 10], [1, 12, 1, 2, 10], [1, 13, 1, 3, 10], [1, 14, 1, 4, 10], [2, 21, 1, 1, 10], [2, 22, 1, 2, 10], [2, 23, 1, 3, 10], [2, 24, 1, 4, 10], [3, 31, 1, 1, 10], [3, 32, 1, 2, 10], [3, 33, 1, 3, 10], [3, 34, 1, 4, 10]]}
+  },
+  "charts": [{"scoreId": 1004, "asset": {"key": "synthetic-one-note", "sha256": "0000000000000000000000000000000000000000000000000000000000000000"}, "notes": {"id": [1], "op": [1], "judgementType": [1], "timeMs": [100]}, "skillEvents": {"timeMs": []}, "fevers": {"startMs": [], "endMs": []}}]
+}
diff --git a/tests/test_deck_cli.py b/tests/test_deck_cli.py
new file mode 100644
index 0000000..be76b8b
--- /dev/null
+++ b/tests/test_deck_cli.py
@@ -0,0 +1,75 @@
+"""Exercise the reviewed-input adapter with a released recommendation CLI."""
+import os
+from pathlib import Path
+
+import pytest
+
+from bdon_vision.deck.common import digest, write
+from bdon_vision.deck.recommend import recommend, scene_matrix
+
+
+DECK = Path(__file__).parent / 'fixtures' / 'deck-synthetic.json'
+
+
+def test_scene_requests_use_the_solver_contract():
+    matrix = scene_matrix(DECK)
+    requests = [entry['request'] for entry in matrix['entries'] if 'request' in entry]
+    assert {request['execution']['kind'] for request in requests} == {'power', 'skip', 'live'}
+    assert all(request['format'] == 'ournotes-deck.search-request/1' for request in requests)
+    assert all('seedLaw' not in request for request in requests)
+
+
+@pytest.fixture
+def solver():
+    binary = os.environ.get('OURNOTES_DECK_BIN')
+    if not binary:
+        pytest.skip('Set OURNOTES_DECK_BIN to run the released CLI integration tests')
+    path = Path(binary).resolve()
+    assert path.is_file(), f'Recommendation CLI not found: {path}'
+    return path
+
+
+@pytest.fixture
+def reviewed_roster(tmp_path):
+    roster = {'members': [{'id':i, 'level':1} for i in range(1,6)], 'snaps':[], 'player':{}}
+    inventory = {'mock':True, 'source':'synthetic', 'region':'synthetic',
+                 'masterVersion':'synthetic-1', 'deckDataSha256':digest(DECK), 'box':{'cards':[]}}
+    write(tmp_path / 'inventory.json', inventory)
+    write(tmp_path / 'adaptation.json', {**inventory, 'complete':True,
+          'originalBox':inventory['box'], 'roster':roster})
+    path = tmp_path / 'roster.json'
+    write(path, roster)
+    return path
+
+
+@pytest.mark.parametrize('scene', [
+    'songless-power', 'free-skip', 'free-live', 'free-live-gekisou', 'mission-live',
+])
+def test_released_cli_recommends_reviewed_roster(solver, reviewed_roster, tmp_path, scene):
+    matrix = scene_matrix(DECK)
+    entry = next(entry for entry in matrix['entries'] if entry['id'] == scene)
+    result = recommend(solver, DECK, reviewed_roster, {**matrix, 'entries':[entry]}, tmp_path / scene)
+    case = result['cases'][0]
+    assert case['exitCode'] == 0, case['rawJson']
+    raw = case['raw']
+    assert raw['format'] == 'ournotes-deck.recommendation-result/3'
+    assert raw['metric'] == entry['request']['metric']
+    assert raw['results']
+    assert all(sorted(deck['members']) == [1,2,3,4,5] for deck in raw['results'])
+    if scene == 'songless-power':
+        assert raw['completion'] == 'Complete'
+        assert raw['optimality'] == 'proven'
+        assert raw['results'][0]['power'] == 1500
+    else:
+        assert int(raw['results'][0]['expectedPayoff']['numerator']) > 0
+    assert result['deckDataSha256'] == digest(DECK)
+    assert result['rosterSha256'] == digest(reviewed_roster)
+
+
+def test_released_cli_preserves_request_errors(solver, reviewed_roster, tmp_path):
+    entry = scene_matrix(DECK)['entries'][0]
+    entry['request']['format'] = 'unsupported-request/1'
+    case = recommend(solver, DECK, reviewed_roster, {'entries':[entry]}, tmp_path / 'error')['cases'][0]
+    assert case['exitCode'] != 0
+    assert case['raw']['error']['code'] == 'Input'
+    assert 'unsupported recommendation format' in case['raw']['error']['message']
diff --git a/tests/test_deck_inputs.py b/tests/test_deck_inputs.py
index 22dd7a5..deb42fe 100644
--- a/tests/test_deck_inputs.py
+++ b/tests/test_deck_inputs.py
@@ -178,6 +178,7 @@ def forbidden(*args,**kwargs):raise AssertionError('solver must not run with fai
     assert not (out/'needs-review.json').exists()
     assert load_review(out)['roster']['members'][4]['id']==6
     def solver(command,**kwargs):
+        assert command[1] == 'recommend'
         assert read(command[command.index('--roster')+1])['members'][4]['id']==6
         return SimpleNamespace(stdout='{}',stderr='',returncode=0)
     monkeypatch.setattr('bdon_vision.deck.recommend.subprocess.run',solver)