更新部署工作流，将分支从 prod 修改为 main，并相应调整文档说明

- Added new sections on missing data handling and experimental variables in data and signal semantics (2026-09-25).
- Included clarifications on clustering representatives and candidate evaluations in portfolio and correlation optimization (2026-09-25).
- Updated README to reflect the latest forum synchronization and new post counts, including significant additions and revisions across multiple topics (2026-09-25).

.gitea/workflows/deploy-production.yaml

@@ -6,8 +6,10 @@ on:
 
 jobs:
   deploy:
-    # Reuse the runner that deploys zhixing-system to the same Docker host.
+    # Manual runs must also select main before receiving production credentials.
-    runs-on: ubuntu-latest
+    if: ${{ github.ref == 'refs/heads/main' }}
+    # Match the label of tencent-prod-runner on server B, not its runner name.
+    runs-on: tencent-prod
     steps:
       - name: Checkout
         uses: https://github.com/actions/checkout@v4
@@ -22,7 +24,6 @@ jobs:
           WQ_PASSWORD: ${{ secrets.WQ_PASSWORD }}
           ENCRYPTION_KEY: ${{ secrets.ENCRYPTION_KEY }}
           ADMIN_USERNAME: ${{ vars.ADMIN_USERNAME }}
-          DATABASE_NETWORK: ${{ vars.DATABASE_NETWORK }}
           PUBLIC_ORIGIN: ${{ vars.PUBLIC_ORIGIN }}
           MCP_ENABLED: ${{ vars.MCP_ENABLED }}
         run: bash scripts/deploy-production.sh

.scratch/template-backtest-flow/issues/01-simplify.md

@@ -0,0 +1,22 @@
+# 简化模板到批量回测流程
+Status: ready-for-agent
+
+## 范围
+模板详情保持编辑、保存、新增版本；回测准备选择数据准备并展开候选，候选集合即用户确认界面，点击回测直接启动批量任务并导航到回测研究。
+模板生成仅保留表达式语法与数据准备/回测参数组合一致性检查；不持久化逐候选校验状态，不依赖算子/字段可用性缓存。字段类型用于候选域选择，不作为表达式类型检查。
+模板侧不提供评估研究结果、变体关系；保留回测来源关联。其他研究生产者的行为不扩大修改。
+
+## 验证
+覆盖语法/组合失败不保存候选、无元数据仍可生成、旧记录不被旧校验状态阻塞、启动幂等与来源、候选分页/选择/直接启动导航。
+
+## 完成结果
+已实现模板独立候选确认界面及单次回测接口，内部在同一事务创建执行快照与批量任务；按账户锁和请求键保证重试幂等，提交后唤醒执行器。
+新模板候选不保存 validation；历史行状态不再参与模板资格判定。旧快照按语法和组合契约读取，其他生产者保留各自检查。
+模板参数选项同步仅辅助选择，不阻塞生成。修改准备参数会移除旧候选；迟到的生成响应不覆盖新的准备状态。
+
+## 验证结果
+- Ruff、前端 typecheck、git diff --check 通过。
+- 后端研究工作区、MCP模板、研究流水线、批量回测：97项通过。
+- 现有设置、研究导航、研究结果浏览器回归：6项通过。
+- Playwright CLI 在隔离模拟环境验证51条候选、跨页选择、40px行高、窄屏滚动、参数变化使候选失效、取消1条后单击回测创建50条任务并导航；无额外预览/评估/变体关系入口，保存来源关联。
+- 未连接真实平台执行回测，未执行生产数据库变更，未提交或推送代码。

.scratch/template-backtest-flow/issues/02-tools.md

@@ -0,0 +1,14 @@
+# 对齐 MCP 与内置 bot 的模板流程
+Status: ready-for-agent
+
+## 范围
+模板来源回测证据改为可选，提供时继续校验真实完整性。MCP 补齐模板版本+数据准备生成候选、分页读取固定候选、按候选 ID 批量回测，沿用权限、幂等、原子事务及审计。
+内置 bot 可直接对模板候选集合请求一次确认，不再单独准备预览；统一模板语法/组合规则与来源。通用回测、SUPER、变体与结果评估工具保持既有职责。
+
+## 验证
+覆盖无来源创建与新增版本、来源错误、分页不截断执行、生成不启动、执行权限隔离、重试幂等及内容冲突、确认前零回测、拒绝/变更确认、提交后唤醒和关联追溯。
+
+## 完成记录
+已实现可选来源证据、MCP 候选生成/分页读取/按集合执行，以及内置 bot 的一次确认回测。模板共用语法与组合一致性规则；完整候选摘要在确认时固定，执行前核对，保持模板来源与调用方审计关联。
+
+验证：相关后端测试共 132 项通过（MCP 模板与研究集成 33 项；MCP/模板工作区/AI/能力回归 99 项，其中 SDK 工具数量断言更新后单独复跑通过）。Ruff 和 git diff --check 通过。使用隔离数据库、模拟模型和平台；未执行真实平台回测。未提交或推送代码。

.scratch/template-detail/issues/01-detail.md

@@ -0,0 +1,14 @@
+# 模板详情与回测准备迭代
+Type: task
+Status: ready-for-human
+
+## 范围
+模板详情移除 AI、假设、导入、删除、历史选择等交互；仅保留名称、类别、研究解释、表达式及占位符类型/描述，操作为保存和新增版本。回测准备按表达式、数据准备、展开选项、回测参数、生成候选集合排列。允许无取值模板保存，展开时按字段类型绑定固定输入；非字段参数不推测。保留现有不可变版本与并发保护，提供 bot 创建及新增版本能力。
+
+## 验证
+类型检查、后端模板保存/展开/版本/bot 回归、浏览器实际交互。
+
+## Answer
+已完成前后端实现。保存修改沿用不可变版本；新增版本允许内容不变时显式创建下一版。模板工坊仅配置字段类型与描述，已有显式 values 仍作为候选限制保留；空 field 从所选固定输入按类型绑定，非字段空参数返回明确 422。新增内置 bot 模板创建/版本能力和外部 MCP 版本、查询能力。
+
+验证：前端 typecheck、改动文件 Ruff、git diff --check 通过。模板与工作空间 50 项测试通过；AI/MCP 回归 39 项通过，1 项因新增工具导致总数断言变化，更新断言后单独复跑通过。两条既有浏览器回归通过。Playwright CLI 实测字段去重/删除、描述保存、类型切换、保存 v1/新增 v2/编辑保存 v3、历史不变、选择数据准备后生成候选且未启动回测。截图位于 output/playwright/template-detail-editor.png 与 template-detail-prepare.png。验证使用隔离数据库和模拟平台。

backend/app/backtests/ai_tools.py

@@ -96,7 +96,7 @@ async def wake_backtests(runner, result):
     runner.backtests.wake.set()
 
 
-INSTRUCTIONS = "回测先读取能力再准备固定候选预览，每次运行确认一次；后续候选新建预览。停止生成不取消回测。\n回测结果追问用 get_backtest/get_backtest_results。上下文或历史没有运行 ID 时，可用 list_backtests 按 source=chatbox 和会话 reference 找回；不得把启动返回当作结果。"
+INSTRUCTIONS = "模板集合使用 start_template_backtest 直接请求确认；其他回测先读取能力再准备固定候选预览。每次运行确认一次；后续候选新建预览。停止生成不取消回测。\n回测结果追问用 get_backtest/get_backtest_results。上下文或历史没有运行 ID 时，可用 list_backtests 按 source=chatbox 和会话 reference 找回；不得把启动返回当作结果。"
 
 
 CAPABILITIES = (

backend/app/mcp_api/server.py

@@ -33,7 +33,13 @@ TOOLS = {
     "get_pyramid_distribution": (c.PyramidQuery, "pyramid_distribution", "research:read", "实时读取指定 region（如 USA、GLB）和 delay（0/1）的个人 Pyramid Alpha 分布；必传 current_date（YYYY-MM-DD），自动按自然年四季度取完整起止日（如2026-09-13对应2026-07-01至2026-09-30），传给平台 startDate/endDate，不使用默认周期。按用户约定 alphaCount>=3 为 lit（已点亮），1–2 为 in_progress，0 为 unlit；每项含 category、alpha_count、距3条的 remaining。复用平台认证，未连接时先调用 authenticate_worldquant；缺失数据不当作0。不回测、不提交。"),
     "search_data_preparations": (c.PreparationSearch, "preparations", "research:read", "分页查询数据准备集合，返回固定范围、字段数及版本。研究可使用多个集合，各集合范围独立。"),
     "get_data_preparation": (c.PreparationRead, "preparation", "research:read", "按集合 ID 与版本分页预览字段、类型、描述和数据集归属。提交回测时携带 preparation_refs，由服务端核对版本并固定独立输入快照；空集合不能用于研究。"),
-    "create_research_template": (c.CreateTemplate, "create_template", "research:write", "将调用方大模型研究后自行总结的参数化模板保存到模板工坊，供用户后续批量回测。先用 get_backtest_results 阅读实际指标和检查，选择 1–20 个已完成采集的 source_item_ids，并说明 hypothesis；不要把 completed 当作检查通过。template 使用 {name} 占位符及逐一对应的 variables，字段变量须声明 MATRIX/VECTOR/GROUP，VECTOR 聚合须明确写入表达式。提供唯一名称和 idempotency_key，可附 reference。返回模板 ID、版本和理论组合数；仅核验结构及来源，不验证所有参数组合，不再次调用模型、不执行回测、不覆盖已有模板。"),
+    "search_research_templates": (c.TemplateSearch, "templates", "research:read", "分页搜索模板工坊的模板与最新版本，不执行研究。"),
+    "get_research_template": (c.TemplateRead, "template", "research:read", "读取模板内容、字段定义和来源；新增版本前核对最新版本。"),
+    "create_research_template_version": (c.CreateTemplateVersion, "create_template_version", "research:write", "为已有模板新增不可变版本。先读取模板，携带 template_id、expected_version、完整 template、hypothesis 和 idempotency_key。source_item_ids 可选，提供时须为真实、已完成采集的回测项。版本冲突须重新读取；不调用模型、不回测。"),
+    "create_research_template": (c.CreateTemplate, "create_template", "research:write", "保存调用方编写的模板，不要求已有回测结果。提供 template、hypothesis、唯一名称和 idempotency_key；source_item_ids 可选，提供时须为真实、已完成采集的回测项。template 使用 {name} 占位符及对应 variables，字段定义只需类型和描述；空字段 values 由展开时的数据准备绑定，其他空参数须补充 values 或直接写入表达式。仅保存模板，不调用模型、不展开、不执行回测。"),
+    "expand_research_template": (c.TemplateExpansion, "expand_template", "research:write", "按 template_id/version、preparation_refs 和完整 settings 生成固定候选集合，支持全组合或固定 seed 随机采样。仅检查语法和数据准备/回测参数组合一致性，无逐行校验状态，不回测。返回首25项和 experiment_id，更多候选用 get_template_candidates 分页读取；用户授权后按明确 candidate_ids 调用 start_template_backtest。重试复用 idempotency_key。"),
+    "get_template_candidates": (c.TemplateCandidates, "template_candidates", "research:read", "分页读取固定模板候选集合的表达式、参数、候选 ID、模板版本、数据准备及关联回测；每页最多100条，total 不是当前页数量。"),
+    "start_template_backtest": (c.SubmitTemplateBacktest, "start_template_backtest", "backtests:execute", "对用户已授权的模板候选集合执行批量回测。提供 experiment_id、明确的 candidate_ids 和 idempotency_key；服务端使用保存的表达式和参数并保留来源，不接受重写输入。不需要另建预览，不重新校验类型或平台可用性。相同请求重试返回同一运行；新幂等键会创建新的回测，包括重复候选。立即返回运行 ID，结果另行查询。"),
     "get_submission_check": (c.SelfCorrelationReference, "submission_check_context", "research:read", "读取已导入 Alpha 的表达式、Description、snapshot 和缓存检查结果；check_summary 分离 Alpha 检查和 REGULAR_SUBMISSION 提交限制，原始 checks 保留；限制不代表当前额度。不发起检查。先核对或生成三段 Description，再调用 check_submission。"),
     "check_submission": (c.SubmissionCheck, "check_submission", "research:refresh", "对单个待提交 Alpha 写回已获用户授权的 Description 并调用平台 GET /check，返回 job_id。须先用 get_submission_check 获取 snapshot；保留本地自相关门槛和冲突保护。通过 get_refresh_job 查进度、get_submission_check 读结果。无论检查结果如何，都不会调用 /submit 或正式提交 Alpha。"),
     "get_worldquant_connection": (c.ConnectionReference, "connection", "research:read", "读取 WorldQuant 连接状态及可选认证 job_id 的进度，不发起认证；人工验证在网页完成。"),
@@ -67,7 +73,7 @@ class MCPResearchServer:
         self.mutation_lock = asyncio.Lock()
         self.server = Server("wq-alpha-research", version="1.0.0", on_list_tools=self.list_tools,
                              on_call_tool=self.call_tool,
-                             instructions="自由探索，直接固定候选回测，无需先建研究资产。工具不安排定时研究；结果按运行 ID 查询。")
+                             instructions="自由探索，直接固定候选回测，无需先建研究资产。使用已有模板时，expand_research_template 生成候选，get_template_candidates 分页核对，获得用户授权后 start_template_backtest 执行；无需额外预览或结果评估步骤。工具不安排定时研究；结果按运行 ID 查询。")
         from urllib.parse import urlsplit
 
         host = urlsplit(settings.public_origin).netloc
@@ -82,8 +88,8 @@ class MCPResearchServer:
         return types.ListToolsResult(tools=[types.Tool(name=name, description=description,
             inputSchema=schema.model_json_schema(), annotations=types.ToolAnnotations(
                 readOnlyHint=scope == "research:read", destructiveHint=method == "control",
-                idempotentHint=method in {"submit", "control", "create_template", "save_super_plan", "build_super_candidates"} or scope == "research:read",
+                idempotentHint=method in {"submit", "control", "create_template", "create_template_version", "save_super_plan", "build_super_candidates", "expand_template", "start_template_backtest"} or scope == "research:read",
-                openWorldHint=method in {"refresh", "submit", "metadata", "check_self_correlation", "check_submission", "authenticate", "pyramid_distribution", "preview_super_selection"}))
+                openWorldHint=method in {"refresh", "submit", "metadata", "check_self_correlation", "check_submission", "authenticate", "pyramid_distribution", "preview_super_selection", "start_template_backtest"}))
             for name, (schema, method, scope, description) in TOOLS.items()
             if scope in principal.scopes and "research:read" in principal.scopes])
 
@@ -136,7 +142,7 @@ class MCPResearchServer:
                         data = {"error": ResearchError(code, "研究操作失败；可使用原幂等键重试或查询历史", retryable=True).data}
                     db.add(MCPAudit(id=str(uuid4()), token_id=principal.token_id, tool=name,
                         request_id=fingerprint({"request_id": request_id}), input_digest=digest,
-                        business_id=data.get("backtest_run_id", data.get("job_id", data.get("template_id", data.get("id")))),
+                        business_id=data.get("backtest_run_id", data.get("job_id", data.get("experiment_id", data.get("template_id", data.get("id"))))),
                         result_code=code, elapsed_ms=int((time.monotonic()-started)*1000)))
                 if not error:
                     if access.wake == "backtests":

backend/app/research/experiments.py

@@ -7,14 +7,22 @@ from collections import defaultdict
 from fastapi import HTTPException
 from sqlalchemy import func, select, update
 
-from ..backtests.contracts import Candidate, DraftInput, PreviewInput, SimulationSettings, Source
+from ..backtests.contracts import (
+    Candidate,
+    DraftInput,
+    PreviewInput,
+    SimulationSettings,
+    Source,
+    StartInput,
+    fingerprint,
+)
 from ..backtests.service import Backtests, uid
 from ..catalog.research_metadata import ResearchMetadata
 from ..catalog.service import Catalog
-from ..models import Alpha, BacktestRun, CatalogResource, ResearchExperiment
+from ..models import Account, Alpha, BacktestPreview, BacktestRun, CatalogResource, ResearchExperiment
 from ..preparations.service import Preparations
 from .assets import Assets
-from .expressions import GROUPS, analyze, expand
+from .expressions import GROUPS, ExpressionError, Parser, analyze, expand
 from .serialization import encode_snapshot as jsonable_encoder
 from .workspace_contracts import TemplateSpec
 
@@ -45,7 +53,7 @@ class Experiments:
         self.catalog = Catalog(db)
         self.assets = Assets(db)
 
-    async def inputs(self, ids, scope=None):
+    async def inputs(self, ids, scope=None, *, check_types=True):
         if len(set(ids)) != len(ids):
             raise HTTPException(422, "输入快照重复")
         snapshots = [await self.catalog.input(input_id) for input_id in ids]
@@ -56,7 +64,7 @@ class Experiments:
             for name, kind in item["field_types"].items():
                 if name not in item["field_ids"]:
                     continue
-                if name in fields and fields[name] != kind:
+                if check_types and name in fields and fields[name] != kind:
                     raise HTTPException(422, f"字段 {name} 在不同快照中类型不一致")
                 fields[name] = kind
         return snapshots, fields
@@ -138,9 +146,9 @@ class Experiments:
         asset = await self.assets.get(body.asset_id, body.version, "template") if body.asset_id else None
         template = TemplateSpec.model_validate(asset["content"]) if asset else body.template
         scope = scope_of(body.settings)
-        if template.scope and template.scope.model_dump() != scope:
+        snapshots, fields = await self.inputs(body.input_ids, scope, check_types=kind != "template")
-            raise HTTPException(422, "模板适用范围与候选设置不同")
+        if kind == "template" and not snapshots:
-        snapshots, fields = await self.inputs(body.input_ids, scope)
+            raise HTTPException(422, "请先选择数据准备")
         parents = (
             parent_snapshots
             if parent_snapshots is not None
@@ -148,31 +156,53 @@ class Experiments:
         )
         variables = {}
         for name, variable in template.variables.items():
+            values = variable.values
+            # Empty field definitions bind only to the selected immutable input scope.
+            # Existing explicit domains remain restrictions and are never silently widened.
             if variable.kind == "field":
-                for value in variable.values:
+                if not values:
-                    if fields.get(str(value)) != variable.field_type:
+                    values = sorted(field for field, kind in fields.items() if kind == variable.field_type)
+                if not values:
+                    raise HTTPException(422, f"变量 {name} 没有匹配的 {variable.field_type} 字段，请调整数据准备")
+                for value in values:
+                    if kind != "template" and fields.get(str(value)) != variable.field_type:
                         raise HTTPException(422, f"变量 {name} 的字段 {value} 不在固定输入中或类型不符")
-            if variable.kind == "group" and any(
+            if kind != "template" and variable.kind == "group" and any(
                 str(v) not in GROUPS and fields.get(str(v)) != "GROUP" for v in variable.values
             ):
                 raise HTTPException(422, f"分组变量 {name} 未在固定输入中核实")
+            if not values:
+                raise HTTPException(422, f"变量 {name} 缺少候选取值，请通过模板接口补充，或将固定参数直接写入表达式")
             variables[name] = [
-                json.dumps(v, ensure_ascii=False) if variable.kind == "string" else v for v in variable.values
+                json.dumps(v, ensure_ascii=False) if variable.kind == "string" else v for v in values
             ]
         try:
             expanded = expand(template.expression, variables, body.mode, body.limit, body.seed)
         except ValueError as exc:
             raise HTTPException(422, str(exc)) from None
+        validation_evidence = {}
+        if kind != "template":
             operators_snapshot = await ResearchMetadata(self.db).get("operators")
             operators = {item["name"] for item in operators_snapshot["content"].get("items", [])}
             setting_errors, settings_snapshot = await self.settings_check(body.settings)
             availability = await self.field_evidence(scope, fields)
+            validation_evidence = {
+                "field_availability": availability,
+                "availability_basis": "各输入的已发布范围目录；若另有字段级证据，须同时满足",
+                "operators_snapshot": operators_snapshot,
+                "settings_snapshot": settings_snapshot,
+            }
         candidates = []
         for index, item in enumerate(expanded["items"]):
+            findings = {}
+            if kind == "template":
+                self.check_syntax(item["expression"], f"候选 {index + 1}")
+            else:
                 validation = self.validate(item["expression"], fields, operators, scope, availability)
                 validation["availability"].extend(setting_errors)
                 if setting_errors and validation["status"] == "valid":
                     validation["status"] = "needs_review"
+                findings["validation"] = validation
             candidates.append(
                 {
                     **Candidate(
@@ -180,7 +210,7 @@ class Experiments:
                     ).model_dump(mode="json"),
                     "bindings": item["bindings"],
                     "input_ids": list(body.input_ids),
-                    "validation": validation,
+                    **findings,
                     "changes": [
                         self.diff(parent.get("expression", ""), item["expression"])
                         for parent in parents
@@ -190,16 +220,21 @@ class Experiments:
             )
         evidence = {
             "template": asset or {"content": template.model_dump(mode="json")},
-            "field_availability": availability,
-            "availability_basis": "各输入的已发布范围目录；若另有字段级证据，须同时满足",
             "combination_count": expanded["combination_count"],
             "seed": expanded["seed"],
-            "operators_snapshot": operators_snapshot,
+            **validation_evidence,
-            "settings_snapshot": settings_snapshot,
             **(extra_evidence or {}),
         }
         return await self.save(template.name, kind, body.hypothesis, snapshots, parents, candidates, evidence)
 
+    @staticmethod
+    def check_syntax(expression, label="表达式"):
+        """Reject unsupported syntax before persistence; platform semantics are not inferred."""
+        try:
+            Parser(expression).parse()
+        except (ExpressionError, RecursionError) as exc:
+            raise HTTPException(422, f"{label}语法错误：{exc}") from None
+
     @staticmethod
     def diff(before, after):
         return [
@@ -284,6 +319,26 @@ class Experiments:
             }
         )
 
+    async def template_candidates(self, experiment_id, limit=25, offset=0):
+        """Read a bounded page of stored template candidates and their immutable references."""
+        experiment = await self.get(experiment_id)
+        if experiment["kind"] != "template":
+            raise HTTPException(422, "此入口仅用于模板候选集合")
+        candidates = experiment["candidates"]
+        template = experiment["evidence"].get("template", {})
+        return {
+            "id": experiment_id, "experiment_id": experiment_id,
+            "name": experiment["name"], "archived": experiment["archived"],
+            "template": {k: template.get(k) for k in ("id", "version", "name")},
+            "inputs": [{k: item.get(k) for k in ("id", "preparation_id", "preparation_version", "scope")}
+                       for item in experiment["inputs"]],
+            "items": [{k: c[k] for k in ("client_item_id", "expression", "settings", "alpha_type", "bindings") if k in c}
+                      for c in candidates[offset:offset + limit]],
+            "total": len(candidates), "limit": limit, "offset": offset,
+            "has_more": offset + limit < len(candidates),
+            "backtest_run_ids": experiment["backtest_run_ids"],
+        }
+
     async def archive(self, experiment_id):
         """Hide an immutable experiment; backtests and lineage must still resolve it.
 
@@ -297,7 +352,8 @@ class Experiments:
             raise HTTPException(404, "研究实验不存在")
         return {"ok": True}
 
-    async def preview(self, experiment_id, candidate_ids=None, source_kind=None, reference=None):
+    async def backtest_input(self, experiment_id, candidate_ids=None, source_kind=None, reference=None):
+        """Build the complete fixed selection and enforce its domain checks."""
         experiment = await self.get(experiment_id)
         candidates = experiment["candidates"]
         if candidate_ids is not None:
@@ -307,14 +363,21 @@ class Experiments:
             candidates = [item for item in candidates if item["client_item_id"] in chosen]
             if len(candidates) != len(chosen):
                 raise HTTPException(422, "选择包含未知候选")
-        else:
+        elif experiment["kind"] != "template":
             candidates = [item for item in candidates if item["validation"]["status"] == "valid"]
-        if not candidates or any(item["validation"]["status"] != "valid" for item in candidates):
+        if not candidates:
+            raise HTTPException(422, "请至少选择一条候选")
+        if experiment["kind"] != "template" and any(item["validation"]["status"] != "valid" for item in candidates):
             raise HTTPException(422, "候选存在语法、类型或可用性问题，请先解决；至少保留一条已核实候选")
         inputs = experiment["inputs"]
-        return await Backtests(self.db).preview(
+        if experiment["kind"] == "template":
-            PreviewInput(
+            # Historical collections follow the same syntax/scope contract; old row findings are irrelevant.
-                inline=DraftInput(
+            for candidate in candidates:
+                self.check_syntax(candidate["expression"], candidate["client_item_id"])
+                scope = scope_of(SimulationSettings.model_validate(candidate["settings"]))
+                if not inputs or any(item["scope"] != scope for item in inputs):
+                    raise HTTPException(422, "数据准备与回测参数组合不一致，请重新生成候选集合")
+        return DraftInput(
             name=experiment["name"],
             source=Source(
                 kind=source_kind or experiment["kind"],
@@ -334,9 +397,55 @@ class Experiments:
                 for item in candidates
             ],
         )
-            ),
+
-            preserve_source=True,
+    async def preview(self, experiment_id, candidate_ids=None, source_kind=None, reference=None, *, backtests=None):
-        )
+        draft = await self.backtest_input(experiment_id, candidate_ids, source_kind, reference)
+        return await (backtests or Backtests(self.db)).preview(PreviewInput(inline=draft), preserve_source=True)
+
+    async def start_template_backtest(self, experiment_id, body, *, backtests=None, confirmed_preview=None):
+        """Start the explicitly selected immutable collection in the caller's transaction.
+
+        Account locking covers preview creation as well as run creation, so concurrent
+        retries share one run. Reusing a key for another collection/selection raises 409.
+        The caller must wake the runner only after committing this transaction.
+        """
+        chosen = set(body.candidate_ids)
+        if len(chosen) != len(body.candidate_ids):
+            raise HTTPException(422, "候选选择包含重复项")
+        await self.db.scalar(select(Account).where(Account.id == 1).with_for_update())
+        previous = await self.db.scalar(select(BacktestRun).where(BacktestRun.idempotency_key == body.idempotency_key))
+        if previous:
+            saved = await self.db.get(BacktestPreview, previous.preview_id)
+            if previous.source.get("research_id") != experiment_id or chosen != {
+                c["client_item_id"] for c in saved.candidates
+            }:
+                raise HTTPException(409, "幂等键已用于另一候选集合或选择")
+            return await Backtests(self.db).run(previous.id)
+        experiment = await self.get(experiment_id)
+        if experiment["kind"] != "template":
+            raise HTTPException(422, "此入口仅用于模板候选集合")
+        if experiment["archived"]:
+            raise HTTPException(409, "候选集合已删除")
+        service = backtests or Backtests(self.db)
+        if confirmed_preview is None:
+            preview = await self.preview(experiment_id, body.candidate_ids, backtests=service)
+        else:
+            # Approval authorizes all persisted candidates, not the first display page.
+            draft = (await self.backtest_input(experiment_id, body.candidate_ids)).model_dump(mode="json")
+            expected = fingerprint({"candidates": draft["candidates"], "source": draft["source"]})
+            saved = await self.db.get(BacktestPreview, confirmed_preview["preview_id"])
+            if (
+                saved is None
+                or saved.version != confirmed_preview["version"]
+                or saved.digest != confirmed_preview["digest"]
+                or saved.digest != expected
+                or fingerprint({"candidates": saved.candidates, "source": saved.source}) != expected
+            ):
+                raise HTTPException(409, "回测候选与确认内容不匹配，请重新确认")
+            preview = confirmed_preview
+        return await service.start(StartInput(
+            preview_id=preview["preview_id"], version=preview["version"], idempotency_key=body.idempotency_key,
+        ))
 
     async def setting_variants(self, body, *, parent_snapshot=None, extra_evidence=None, kind="variant"):
         parents = (

backend/app/research/routes.py

@@ -2,6 +2,7 @@
 
 from fastapi import APIRouter, Depends, HTTPException, Query, Request
 
+from ..backtests.contracts import RunOutput
 from ..security import require_auth
 from .assets import Assets
 from .comparisons import compare
@@ -22,6 +23,7 @@ from .workspace_contracts import (
     ImportCommit,
     ImportPreview,
     SettingVariants,
+    TemplateBacktest,
     WorkflowSpec,
 )
 
@@ -149,6 +151,14 @@ async def preview(experiment_id: str, body: ExperimentPreview, request: Request)
         return await Experiments(db).preview(experiment_id, body.candidate_ids)
 
 
+@router.post("/experiments/{experiment_id}/backtest", status_code=202, response_model=RunOutput)
+async def template_backtest(experiment_id: str, body: TemplateBacktest, request: Request):
+    async with request.app.state.sessions.begin() as db:
+        result = await Experiments(db).start_template_backtest(experiment_id, body)
+    request.app.state.runner.backtests.wake.set()
+    return result
+
+
 @router.post("/variants/settings", status_code=201)
 async def settings_variants(body: SettingVariants, request: Request):
     async with request.app.state.sessions.begin() as db:

backend/app/research/runtime.py

@@ -522,7 +522,8 @@ class ResearchRuntime:
             "type": "candidates",
             "experiment_id": experiment["id"],
             "candidate_ids": [
-                c["client_item_id"] for c in experiment["candidates"] if c["validation"]["status"] == "valid"
+                c["client_item_id"] for c in experiment["candidates"]
+                if experiment["kind"] == "template" or c["validation"]["status"] == "valid"
             ],
         }
         ids = step.output["candidate_ids"]

backend/app/research/workspace_contracts.py

@@ -16,7 +16,8 @@ AssetKind = Literal["template", "feature", "view", "workflow", "superalpha_plan"
 
 class Variable(Contract):
     kind: Literal["field", "operator", "integer", "number", "group", "string", "fragment"]
-    values: list[str | int | float] = Field(min_length=1, max_length=10000)
+    values: list[str | int | float] = Field(default_factory=list, max_length=10000)
+    description: str = Field(default="", max_length=3000)
     field_type: Literal["MATRIX", "VECTOR", "GROUP"] | None = None
 
     @model_validator(mode="after")
@@ -44,7 +45,6 @@ class TemplateSpec(Contract):
     description: str = Field(default="", max_length=10000)
     expression: str = Field(min_length=1, max_length=20000)
     variables: dict[str, Variable] = Field(default_factory=dict, max_length=100)
-    scope: Scope | None = None
     category: Literal["template", "fragment"] = "template"
 
     @field_validator("expression")
@@ -144,6 +144,11 @@ class ExperimentPreview(Contract):
     candidate_ids: list[str] | None = Field(default=None, min_length=1, max_length=10000)
 
 
+class TemplateBacktest(Contract):
+    candidate_ids: list[str] = Field(min_length=1, max_length=10000)
+    idempotency_key: str = Field(min_length=1, max_length=100)
+
+
 class EvaluationRules(Contract):
     version: Literal["research-v1"] = "research-v1"
     sharpe_min: float = Field(default=1.0, allow_inf_nan=False)

backend/app/research/workspace_tools.py

@@ -1,15 +1,25 @@
 """Research capabilities use the same versioned assets and experiment services as HTTP."""
 
+from fastapi import HTTPException
 from pydantic import Field
 
 from ..ai.capabilities import Capability
+from ..backtests.ai_tools import wake_backtests
 from ..catalog.research_metadata import ResearchMetadata
 from ..schemas import Contract
 from .assets import Assets
 from .evaluations import Evaluations
 from .experiments import Experiments
 from .features import Features
-from .workspace_contracts import AssetWrite, EvaluateInput, Expansion, FeatureSpec, SettingVariants
+from .workspace_contracts import (
+    AssetWrite,
+    EvaluateInput,
+    Expansion,
+    FeatureSpec,
+    SettingVariants,
+    TemplateBacktest,
+    TemplateSpec,
+)
 
 
 class AssetQuery(Contract):
@@ -33,6 +43,10 @@ class FeatureWrite(Contract):
     version: int | None = Field(default=None, ge=1)
 
 
+class TemplateVersionWrite(FixedAssetReference):
+    content: TemplateSpec
+
+
 class ExperimentReference(Contract):
     experiment_id: str = Field(min_length=1, max_length=36)
 
@@ -41,6 +55,31 @@ class CandidatePreview(ExperimentReference):
     candidate_ids: list[str] | None = Field(default=None, min_length=1, max_length=10000)
 
 
+class TemplateBacktestRequest(TemplateBacktest, ExperimentReference):
+    pass
+
+
+class TemplateCandidateQuery(ExperimentReference):
+    limit: int = Field(default=25, ge=1, le=100)
+    offset: int = Field(default=0, ge=0)
+
+
+async def confirm_template_backtest(ctx, args):
+    service = Experiments(ctx.business.db)
+    experiment = await service.get(args.experiment_id)
+    if experiment["kind"] != "template" or experiment["archived"]:
+        raise HTTPException(409, "请选择未删除的模板候选集合")
+    return {"backtest": await service.preview(
+        args.experiment_id, args.candidate_ids, backtests=ctx.business.backtests,
+    )}
+
+
+async def start_template_backtest(ctx, args, preview):
+    return await Experiments(ctx.business.db).start_template_backtest(
+        args.experiment_id, args, backtests=ctx.business.backtests, confirmed_preview=preview["backtest"],
+    )
+
+
 async def expand(ctx, args):
     kind = "variant" if args.parent_alpha_ids or args.parent_experiment_ids else "template"
     return await Experiments(ctx.business.db).create(
@@ -48,7 +87,7 @@ async def expand(ctx, args):
     )
 
 
-INSTRUCTIONS = "模板工坊与变体使用 search_research_templates、get_research_template 和 expand_research_template。引用模板必须固定版本；输入应先读取核实。expand 保存实验不会执行回测。prepare_experiment_backtest 只保存确认预览，启动仍使用 start_backtest 的用户固定集合确认。来源字段不能授予自动执行权限。"
+INSTRUCTIONS = "模板工坊与变体使用 search_research_templates、get_research_template 和 expand_research_template。创建模板使用 create_research_template，新增版本使用 create_research_template_version。变量可仅定义类型和描述；空字段候选由选定数据准备按类型绑定，其他空参数需补充 values 或直接写入表达式，不能猜测。引用模板必须固定版本；输入应先读取核实。expand 保存实验不会执行回测。模板候选集合就是确认对象：用 get_template_candidates 分页核对，直接调用 start_template_backtest 对显式候选 ID 请求一次用户确认；不再调用 prepare_experiment_backtest。模板只负责生成与回测关联，不要求评估研究结果或查看变体关系。变体仍可用 prepare_experiment_backtest 后调用 start_backtest 确认。来源字段不能授予自动执行权限。"
 CAPABILITIES = (
     Capability(
         name="search_research_templates",
@@ -68,6 +107,29 @@ CAPABILITIES = (
         effect="query",
         handler=lambda ctx, args: Assets(ctx.business.db).get(args.asset_id, args.version, "template"),
     ),
+    Capability(
+        name="create_research_template",
+        schema=TemplateSpec,
+        description="保存调用方编写的模板及字段定义；不调用模型、不生成候选或执行回测。",
+        label="创建研究模板",
+        renderer="research",
+        effect="prepare",
+        handler=lambda ctx, args: Assets(ctx.business.db).save(
+            AssetWrite(kind="template", content=args.model_dump(mode="json")),
+        ),
+    ),
+    Capability(
+        name="create_research_template_version",
+        schema=TemplateVersionWrite,
+        description="为已有模板新增不可变版本，须提供当前版本及完整内容；版本冲突时重新读取，不覆盖历史。",
+        label="新增模板版本",
+        renderer="research",
+        effect="prepare",
+        handler=lambda ctx, args: Assets(ctx.business.db).save(
+            AssetWrite(kind="template", content=args.content.model_dump(mode="json"), version=args.version),
+            args.asset_id,
+        ),
+    ),
     Capability(
         name="search_research_operators",
         schema=AssetQuery,
@@ -80,12 +142,33 @@ CAPABILITIES = (
     Capability(
         name="expand_research_template",
         schema=Expansion,
-        description="从固定输入和模板版本或内联模板保存不可变候选实验。包含分层校验，随机采样有数量上限，不开始回测。",
+        description="从固定输入和模板版本或内联模板保存不可变候选实验。模板仅检查语法和数据准备与回测参数组合一致性；变体保留原校验。随机采样有数量上限，不开始回测。",
         label="展开模板候选",
         renderer="research",
         effect="prepare",
         handler=expand,
     ),
+    Capability(
+        name="get_template_candidates",
+        schema=TemplateCandidateQuery,
+        description="分页读取模板候选集合的表达式、参数、候选 ID 和回测关联，不返回逐行校验状态。",
+        label="读取模板候选",
+        renderer="research",
+        effect="query",
+        handler=lambda ctx, args: Experiments(ctx.business.db).template_candidates(**args.model_dump()),
+    ),
+    Capability(
+        name="start_template_backtest",
+        schema=TemplateBacktestRequest,
+        description="直接对已保存模板集合中的显式候选 ID 请求一次用户确认，确认后批量回测；无需准备额外预览。重试复用幂等键。",
+        label="回测模板候选",
+        renderer="backtest",
+        effect="confirm",
+        preview=confirm_template_backtest,
+        execute=start_template_backtest,
+        after_commit=wake_backtests,
+        refresh=("backtests",),
+    ),
     Capability(
         name="prepare_setting_variants",
         schema=SettingVariants,
@@ -107,7 +190,7 @@ CAPABILITIES = (
     Capability(
         name="prepare_experiment_backtest",
         schema=CandidatePreview,
-        description="从实验内已校验的固定候选保存回测确认预览，不启动模拟。",
+        description="为变体等研究实验保存回测预览，不启动模拟；模板直接使用 start_template_backtest。",
         label="准备研究回测",
         renderer="backtest",
         effect="prepare",

backend/app/research_access/contracts.py

@@ -116,7 +116,7 @@ class Control(Contract):
 class CreateTemplate(Contract):
     template: TemplateSpec
     hypothesis: str = Field(min_length=1, max_length=10000)
-    source_item_ids: list[RunId] = Field(min_length=1, max_length=20)
+    source_item_ids: list[RunId] = Field(default_factory=list, max_length=20)
     reference: str | None = Field(default=None, max_length=200)
     idempotency_key: Identifier
 
@@ -133,6 +133,41 @@ class CreateTemplate(Contract):
         return self
 
 
+class CreateTemplateVersion(CreateTemplate):
+    template_id: RunId
+    expected_version: int = Field(ge=1)
+
+
+class TemplateRead(Contract):
+    template_id: RunId
+    version: int | None = Field(default=None, ge=1)
+
+
+class TemplateSearch(Page):
+    q: str = Field(default="", max_length=200)
+
+
+class TemplateExpansion(Contract):
+    template_id: RunId
+    version: int = Field(ge=1)
+    preparation_refs: list[PreparationReference] = Field(min_length=1, max_length=20)
+    settings: CompleteSettings
+    mode: Literal["all", "random"] = "all"
+    limit: int = Field(default=100, ge=1, le=10000)
+    seed: int = 0
+    idempotency_key: Identifier
+
+
+class TemplateCandidates(Page):
+    experiment_id: RunId
+
+
+class SubmitTemplateBacktest(Contract):
+    experiment_id: RunId
+    candidate_ids: list[Identifier] = Field(min_length=1, max_length=10000)
+    idempotency_key: Identifier
+
+
 class CatalogSearch(Contract):
     filters: CatalogFilters
     dataset_id: str | None = Field(default=None, min_length=1, max_length=200)

backend/app/research_access/service.py

@@ -81,7 +81,13 @@ class ResearchAccess(SuperResearchAccess):
                 "metadata_only": True, "actual_platform_allowance": None,
                 "templates": {
                     "create_with": "create_research_template", "required_scope": "research:write",
-                    "authored_by": "caller", "max_source_items": 20,
+                    "version_with": "create_research_template_version",
+                    "read_with": "get_research_template", "search_with": "search_research_templates",
+                    "authored_by": "caller", "source_items_required": False, "max_source_items": 20,
+                    "expand_with": "expand_research_template", "candidates_with": "get_template_candidates",
+                    "backtest_with": "start_template_backtest", "execution_scope": "backtests:execute",
+                    "max_candidates": 10000, "page_size_max": 100,
+                    "validation": "syntax_and_preparation_settings_combination",
                     "source_items_with": "get_backtest_results", "starts_backtests": False,
                     "web_url": f"{self.public_origin}/#templates",
                 },
@@ -285,6 +291,76 @@ class ResearchAccess(SuperResearchAccess):
         return await self.remember("create_research_template", args, digest, result,
                                    business_id=result["template_id"])
 
+    async def template(self, args):
+        """Read the exact saved template revision without executing research."""
+        from ..research.assets import Assets
+
+        return await Assets(self.db).get(args.template_id, args.version, "template")
+
+    async def templates(self, args):
+        """Search the same template library used by the browser."""
+        from ..research.assets import Assets
+
+        return await Assets(self.db).list("template", **args.model_dump())
+
+    async def create_template_version(self, args):
+        """Append an idempotent, optimistic revision with refreshed source evidence."""
+        from .templates import create_template
+
+        operation = "create_research_template_version"
+        previous, digest = await self.previous(operation, args)
+        if previous:
+            return previous.response
+        result = await create_template(self.db, args, self.principal, asset_id=args.template_id)
+        result["web_url"] = f"{self.public_origin}/#templates"
+        return await self.remember(operation, args, digest, result, business_id=args.template_id)
+
+    async def expand_template(self, args):
+        """Idempotently freeze a version and preparations; never start execution."""
+        from ..research.experiments import Experiments
+        from ..research.workspace_contracts import Expansion
+
+        operation = "expand_research_template"
+        previous, digest = await self.previous(operation, args)
+        if previous:
+            return previous.response
+        service = Experiments(self.db)
+        asset = await service.assets.get(args.template_id, args.version, "template")
+        experiment = await service.create(Expansion(
+            asset_id=args.template_id, version=args.version, preparation_refs=args.preparation_refs,
+            settings=args.settings, mode=args.mode, limit=args.limit, seed=args.seed,
+            hypothesis=asset["content"].get("description", "").strip() or f"使用模板：{asset['name']}",
+        ), extra_evidence={"method": "template", "mcp_token_id": self.principal.token_id,
+                           "admin_id": self.principal.admin_id})
+        result = await service.template_candidates(experiment["id"])
+        result.update({"read_with": "get_template_candidates", "backtest_with": "start_template_backtest",
+                       "starts_backtests": False})
+        return await self.remember(operation, args, digest, result, business_id=experiment["id"])
+
+    async def template_candidates(self, args):
+        from ..research.experiments import Experiments
+
+        return await Experiments(self.db).template_candidates(args.experiment_id, args.limit, args.offset)
+
+    async def start_template_backtest(self, args):
+        """Execute only stored candidate IDs; authorization comes from the caller's execute scope."""
+        from ..research.experiments import Experiments
+        from ..research.workspace_contracts import TemplateBacktest
+
+        operation = "start_template_backtest"
+        previous, digest = await self.previous(operation, args)
+        if previous:
+            return previous.response
+        provenance = {"mcp_token_id": self.principal.token_id, "admin_id": self.principal.admin_id}
+        result = await Experiments(self.db).start_template_backtest(
+            args.experiment_id,
+            TemplateBacktest(candidate_ids=args.candidate_ids, idempotency_key="mcp-template-" + str(uuid4())),
+            backtests=Backtests(self.db, provenance),
+        )
+        result = {**result, "input_digest": digest, "web_url": self.run_url(result["backtest_run_id"])}
+        return await self.remember(operation, args, digest, result,
+                                   business_id=result["backtest_run_id"], wake="backtests")
+
     async def previous(self, operation, args):
         # PostgreSQL row lock is shared with HTTP start and catalog/job creation.
         account = await self.db.scalar(select(Account).where(Account.id == self.principal.account_id).with_for_update())

backend/app/research_access/templates.py

@@ -10,23 +10,27 @@ from ..research.workspace_contracts import AssetWrite
 from .queries import item_summary
 
 
-async def create_template(db, args, principal):
+async def create_template(db, args, principal, *, asset_id=None):
-    """Save a new asset inside the caller's account-locked, idempotent transaction.
+    """Save a new asset or revision inside the caller's account-locked, idempotent transaction.
 
     Args contain the external model's TemplateSpec and local source item IDs.
     Return the versioned asset and theoretical combination count. Raise
     ResearchError for name conflicts or missing/incomplete research evidence.
     Stored evidence proves provenance, not profitability or platform eligibility;
     schema validation does not validate every expanded FASTEXPR combination.
+    Optional asset_id selects a version update guarded by args.expected_version;
+    a stale version raises HTTP 409 and cannot overwrite a historical revision.
     """
     from .service import ResearchError
 
     existing = await db.scalar(select(ResearchAsset).where(
         ResearchAsset.kind == "template", ResearchAsset.name == args.template.name,
+        ResearchAsset.id != asset_id if asset_id else True,
     ).order_by(ResearchAsset.id).limit(1))
     if existing:
         raise ResearchError("TEMPLATE_NAME_CONFLICT", "模板名称已存在，请使用新名称；此工具不覆盖已有模板",
                             affected_items=[{"template_id": existing.id, "version": existing.version}])
+    previous = await Assets(db).get(asset_id, expected_kind="template") if asset_id else None
     rows = (await db.execute(select(BacktestItem, BacktestResult).outerjoin(
         BacktestResult, BacktestResult.item_id == BacktestItem.id,
     ).where(BacktestItem.id.in_(args.source_item_ids)))).all()
@@ -48,13 +52,17 @@ async def create_template(db, args, principal):
         "admin_id": principal.admin_id,
         "source_items": [item_summary(*found[item_id]) for item_id in args.source_item_ids],
     }
+    if previous:
+        provenance["parent_template"] = {"id": asset_id, "version": args.expected_version}
     asset = await Assets(db).save(AssetWrite(
         kind="template", content=args.template.model_dump(mode="json"),
-    ), provenance=provenance)
+        version=args.expected_version if asset_id else None,
+    ), asset_id=asset_id, provenance=provenance)
     return {
         **asset, "template_id": asset["id"],
-        "combination_count": str(math.prod(len(v.values) for v in args.template.variables.values())),
+        "combination_count": (str(math.prod(len(v.values) for v in args.template.variables.values()))
-        "validation": {"structure": "valid", "source_evidence": "recorded",
+                              if all(v.values for v in args.template.variables.values()) else None),
+        "validation": {"structure": "valid", "source_evidence": "recorded" if args.source_item_ids else "not_provided",
                        "expanded_candidates": "not_validated", "platform_semantics": "unknown"},
-        "next_step": "在模板工坊选择固定输入及模拟设置，展开并核验候选，再确认批量回测。",
+        "next_step": "选择数据准备和回测参数，用 expand_research_template 生成候选集合；核对候选后按已获授权范围调用 start_template_backtest。也可在模板工坊完成。",
     }

backend/migrations/versions/0023_remove_template_scope.py

@@ -0,0 +1,80 @@
+"""Remove retired template scope metadata without changing execution scopes."""
+
+import sqlalchemy as sa
+from alembic import op
+
+revision = "0023"
+down_revision = "0022"
+branch_labels = None
+depends_on = None
+
+# Only structured research documents carry reusable template definitions. Do not
+# rewrite opaque model messages, audit digests, or already-authorized tool calls.
+DOCUMENTS = (
+    ("research_revisions", ("asset_id", "version"), ("content", "provenance")),
+    ("research_experiments", ("id",), ("evidence", "parents")),
+    ("research_flow_runs", ("id",), ("definition", "authorization")),
+    ("research_step_runs", ("id",), ("output",)),
+    ("research_requests", ("id",), ("response",)),
+)
+
+
+def remove_template_scope(value):
+    """Copy structured templates without scope; preserve input scopes and bindings.
+
+    Match the frozen persisted TemplateSpec shape, including nested feature
+    templates and asset snapshots. Never remove an arbitrary key named scope:
+    input snapshots and scope-named expression variables still require it.
+    """
+    if isinstance(value, list):
+        return [remove_template_scope(item) for item in value]
+    if not isinstance(value, dict):
+        return value
+    is_template = (
+        isinstance(value.get("name"), str)
+        and isinstance(value.get("expression"), str)
+        and isinstance(value.get("variables"), dict)
+    )
+    return {
+        key: remove_template_scope(item)
+        for key, item in value.items()
+        if not (is_template and key == "scope")
+    }
+
+
+def upgrade():
+    """Clean template documents in bounded batches, retaining IDs and versions."""
+    connection = op.get_bind()
+    for name, keys, documents in DOCUMENTS:
+        table = sa.table(name, *(
+            [sa.column(key, sa.Integer() if key == "version" else sa.String()) for key in keys]
+            + [sa.column(column, sa.JSON()) for column in documents]
+        ))
+        last = None
+        while True:
+            query = sa.select(table).order_by(*(table.c[key] for key in keys)).limit(500)
+            if last is not None:
+                query = query.where(sa.or_(*(
+                    sa.and_(*(table.c[keys[j]] == last[j] for j in range(i)), table.c[key] > last[i])
+                    for i, key in enumerate(keys)
+                )))
+            rows = connection.execute(query).mappings().all()
+            if not rows:
+                break
+            for row in rows:
+                changes = {}
+                for column in documents:
+                    cleaned = remove_template_scope(row[column])
+                    if cleaned != row[column]:
+                        changes[column] = cleaned
+                if changes:
+                    connection.execute(table.update().where(
+                        *(table.c[key] == row[key] for key in keys)
+                    ).values(**changes))
+            last = tuple(rows[-1][key] for key in keys)
+
+
+def downgrade():
+    # Previous schemas allow absent template scope; deleted metadata cannot be
+    # reconstructed. Execution scopes and all other research data are unchanged.
+    pass

backend/tests/test_mcp.py

@@ -162,7 +162,9 @@ async def test_official_sdk_client_and_error_contract(mcp_app):
             async with ClientSession(streams[0], streams[1]) as client:
                 await client.initialize()
                 listed = await client.list_tools()
-                assert len(listed.tools) == 29
+                assert len(listed.tools) == 35
+                assert {"expand_research_template", "get_template_candidates", "start_template_backtest"} <= {t.name for t in listed.tools}
+                assert {"search_research_templates", "get_research_template", "create_research_template_version"} <= {t.name for t in listed.tools}
                 assert any(tool.name == "get_pyramid_distribution" for tool in listed.tools)
                 assert {"search_data_preparations", "get_data_preparation"} <= {t.name for t in listed.tools}
                 caps = await client.call_tool("get_research_capabilities", {})

backend/tests/test_mcp_templates.py

@@ -100,7 +100,7 @@ async def test_sdk_template_creation_frozen_evidence_and_web_expansion(app, logg
                 assert "submit_backtests" not in listed
                 assert not tool.annotations.read_only_hint and not tool.annotations.destructive_hint
                 assert tool.annotations.idempotent_hint and not tool.annotations.open_world_hint
-                assert {"template", "hypothesis", "source_item_ids", "idempotency_key"} <= set(tool.input_schema["required"])
+                assert {"template", "hypothesis", "idempotency_key"} <= set(tool.input_schema["required"])
                 assert tool.input_schema["additionalProperties"] is False
                 caps = await client.call_tool("get_research_capabilities", {})
                 assert caps.structured_content["templates"]["create_with"] == TOOL
@@ -138,7 +138,7 @@ async def test_sdk_template_creation_frozen_evidence_and_web_expansion(app, logg
     assert {c["expression"] for c in experiment["candidates"]} == {
         f"rank({field}) + {offset}" for field in ["TEST_FIN_001", "TEST_FIN_002"] for offset in [0, 1, 5]
     }
-    assert all(c["validation"]["status"] == "valid" for c in experiment["candidates"])
+    assert all("validation" not in c for c in experiment["candidates"])
     assert experiment["evidence"]["template"]["provenance"] == stored["provenance"]
     async with app.state.sessions() as db:
         assert await db.scalar(select(func.count()).select_from(BacktestRun)) == 1
@@ -169,7 +169,7 @@ async def test_template_invalid_inputs_and_missing_sources_are_atomic(app, compl
     principal, _ = await credentials(app)
     valid = template_request(completed_source["id"])
     variants = [
-        {"source_item_ids": []}, {"source_item_ids": [completed_source["id"]] * 21},
+        {"source_item_ids": [completed_source["id"]] * 21},
         {"source_item_ids": [completed_source["id"]] * 2}, {"hypothesis": " "},
         {"force": True}, {"idempotency_key": ""},
         {"template": valid["template"] | {"expression": "rank({missing})"}},
@@ -236,3 +236,163 @@ async def test_browser_can_issue_template_only_and_all_permissions(app, logged_i
         async with app.state.sessions() as db:
             principal = await authenticate(db, response.json()["token"])
             assert principal.scopes == scopes
+
+
+async def test_mcp_template_version_is_idempotent_and_preserves_history(app, completed_source):
+    principal, _ = await credentials(app, {"research:read", "research:write"})
+    body = template_request(completed_source["id"])
+    created = await invoke(app, principal, TOOL, body)
+    content = deepcopy(body["template"])
+    content["variables"]["field"] = {"kind": "field", "field_type": "MATRIX", "description": "数据准备中的矩阵字段"}
+    update = {**body, "template": content, "template_id": created["id"],
+              "expected_version": 1, "idempotency_key": "version-2"}
+    tool = "create_research_template_version"
+    first = await invoke(app, principal, tool, update)
+    replay = await invoke(app, principal, tool, update)
+    assert first == replay and first["version"] == 2
+    assert first["combination_count"] is None
+    assert first["provenance"]["parent_template"] == {"id": created["id"], "version": 1}
+    read = await invoke(app, principal, "get_research_template", {"template_id": created["id"], "version": 1})
+    assert read["content"]["variables"]["field"]["values"] == ["TEST_FIN_001", "TEST_FIN_002"]
+    listed = await invoke(app, principal, "search_research_templates", {"q": content["name"]})
+    assert listed["items"][0]["version"] == 2
+    stale = await app.state.mcp.invoke(principal, tool, update | {"idempotency_key": "stale-version"})
+    assert stale.is_error and stale.structured_content["error"]["code"] == "CONFLICT"
+    async with app.state.sessions() as db:
+        assert await db.scalar(select(func.count()).select_from(ResearchRevision)) == 2
+    denied, _ = await credentials(app, {"research:read"})
+    from fastapi import HTTPException
+    with pytest.raises(HTTPException) as forbidden:
+        await app.state.mcp.invoke(denied, tool, update)
+    assert forbidden.value.status_code == 403
+
+
+async def test_template_without_result_sources_can_be_created_and_versioned(app):
+    principal, _ = await credentials(app, {"research:read", "research:write"})
+    body = template_request("unused")
+    body.pop("source_item_ids")
+    created = await invoke(app, principal, TOOL, body)
+    assert created["provenance"]["source_items"] == []
+    assert created["validation"]["source_evidence"] == "not_provided"
+    revised = await invoke(app, principal, "create_research_template_version", {
+        **body, "template_id": created["id"], "expected_version": 1, "idempotency_key": "no-source-v2",
+        "source_item_ids": [],
+    })
+    assert revised["version"] == 2 and revised["provenance"]["source_items"] == []
+    async with app.state.sessions() as db:
+        assert await db.scalar(select(func.count()).select_from(BacktestRun)) == 0
+
+
+async def template_collection(app, principal, research_input, count=130):
+    body = template_request("unused")
+    body.pop("source_item_ids")
+    body["template"]["variables"]["field"]["values"] = ["TEST_FIN_001"]
+    body["template"]["variables"]["offset"]["values"] = list(range(count))
+    saved = await invoke(app, principal, TOOL, body)
+    args = {
+        "template_id": saved["id"], "version": saved["version"],
+        "preparation_refs": [{"id": research_input["preparation_id"], "version": research_input["preparation_version"]}],
+        "settings": candidate()["settings"], "limit": count, "idempotency_key": "expand-collection",
+    }
+    return saved, args
+
+
+async def test_template_collection_mcp_paging_execution_and_replay(app, logged_in, research_input):
+    from sqlalchemy import delete
+
+    from app.models import BacktestPreview, CatalogResource, ResearchExperiment
+
+    principal, _ = await credentials(app)
+    saved, args = await template_collection(app, principal, research_input)
+    async with app.state.sessions.begin() as db:
+        await db.execute(delete(CatalogResource))
+    app.state.runner.backtests.wake.clear()
+    first, retry = await asyncio.gather(*[invoke(app, principal, "expand_research_template", args) for _ in range(2)])
+    assert first == retry and first["total"] == 130 and len(first["items"]) == 25
+    assert first["has_more"] and not first["starts_backtests"]
+    assert first["template"]["id"] == saved["id"] and first["template"]["version"] == 1
+    assert not app.state.runner.backtests.wake.is_set()
+    collected = []
+    for offset in (0, 100):
+        result = await invoke(app, principal, "get_template_candidates", {
+            "experiment_id": first["experiment_id"], "limit": 100, "offset": offset,
+        })
+        collected += result["items"]
+    assert len(collected) == 130 and not result["has_more"]
+    assert all("validation" not in item for item in collected)
+    async with app.state.sessions() as db:
+        assert await db.scalar(select(func.count()).select_from(ResearchExperiment)) == 1
+        assert await db.scalar(select(func.count()).select_from(BacktestRun)) == 0
+    args = {"experiment_id": first["experiment_id"], "candidate_ids": [c["client_item_id"] for c in collected],
+            "idempotency_key": "execute-collection"}
+    run, replay = await asyncio.gather(*[invoke(app, principal, "start_template_backtest", args) for _ in range(2)])
+    assert run == replay and run["total"] == 130
+    assert run["source"]["kind"] == "template" and run["source"]["research_id"] == first["experiment_id"]
+    assert run["source"]["input_snapshot_ids"]
+    assert app.state.runner.backtests.wake.is_set()
+    rotated, _ = await credentials(app)
+    assert await invoke(app, rotated, "start_template_backtest", args) == run
+    conflict = await app.state.mcp.invoke(principal, "start_template_backtest", args | {"candidate_ids": ["c1"]})
+    assert conflict.is_error and conflict.structured_content["error"]["code"] == "IDEMPOTENCY_CONFLICT"
+    async with app.state.sessions() as db:
+        assert await db.scalar(select(func.count()).select_from(BacktestRun)) == 1
+        assert await db.scalar(select(func.count()).select_from(BacktestPreview)) == 1
+        row = await db.get(BacktestRun, run["backtest_run_id"])
+        assert row.ai_context["mcp_token_id"] == principal.token_id
+        audit = await db.scalar(select(MCPAudit).where(MCPAudit.tool == "expand_research_template"))
+        assert audit.business_id == first["experiment_id"]
+    record = (await logged_in.get(f"/api/v1/research/experiments/{first['experiment_id']}")).json()
+    assert record["backtest_run_ids"] == [run["backtest_run_id"]]
+
+
+async def test_template_mcp_failures_are_atomic_and_do_not_consume_keys(app, research_input):
+    from app.models import BacktestPreview, ResearchExperiment
+
+    principal, _ = await credentials(app)
+    saved, args = await template_collection(app, principal, research_input, 2)
+    for changed in [args | {"settings": args["settings"] | {"region": "EUR"}},
+                    args | {"preparation_refs": [{**args["preparation_refs"][0], "version": 999}]}]:
+        result = await app.state.mcp.invoke(principal, "expand_research_template", changed)
+        assert result.is_error
+    async with app.state.sessions() as db:
+        assert await db.scalar(select(func.count()).select_from(ResearchExperiment)) == 0
+        assert not await db.scalar(select(ResearchRequest).where(ResearchRequest.operation == "expand_research_template"))
+    collection = await invoke(app, principal, "expand_research_template", args)
+    conflict = await app.state.mcp.invoke(principal, "expand_research_template", args | {"seed": 1})
+    assert conflict.structured_content["error"]["code"] == "IDEMPOTENCY_CONFLICT"
+    request = {"experiment_id": collection["experiment_id"], "candidate_ids": ["c1"], "idempotency_key": "execute"}
+    for ids in [[], ["unknown"], ["c1", "c1"]]:
+        result = await app.state.mcp.invoke(principal, "start_template_backtest", request | {"candidate_ids": ids})
+        assert result.is_error
+    result = await app.state.mcp.invoke(principal, "start_template_backtest", request | {"expression": "rank(other)"})
+    assert result.is_error and result.structured_content["error"]["code"] == "INVALID_INPUT"
+    async with app.state.sessions() as db:
+        assert await db.scalar(select(func.count()).select_from(BacktestRun)) == 0
+        assert await db.scalar(select(func.count()).select_from(BacktestPreview)) == 0
+        assert not await db.scalar(select(ResearchRequest).where(ResearchRequest.operation == "start_template_backtest"))
+    await invoke(app, principal, "start_template_backtest", request)
+
+
+async def test_template_tool_discovery_and_execute_permissions(app, research_input):
+    from fastapi import HTTPException
+
+    principal, secret = await credentials(app, {"research:read", "research:write"})
+    async with httpx.AsyncClient(transport=httpx.ASGITransport(app=app), base_url="http://testserver",
+                               headers={"Authorization": f"Bearer {secret}", "Accept": "application/json, text/event-stream"}) as http:
+        listed = (await http.post(ENDPOINT, json={"jsonrpc": "2.0", "id": 1, "method": "tools/list"})).json()
+        tools = {t["name"]: t for t in listed["result"]["tools"]}
+        assert "expand_research_template" in tools and "get_template_candidates" in tools
+        assert "start_template_backtest" not in tools
+        assert "source_item_ids" not in tools[TOOL]["inputSchema"]["required"]
+        assert tools["expand_research_template"]["annotations"]["idempotentHint"]
+    _, args = await template_collection(app, principal, research_input, 2)
+    collection = await invoke(app, principal, "expand_research_template", args)
+    with pytest.raises(HTTPException) as denied:
+        await app.state.mcp.invoke(principal, "start_template_backtest", {
+            "experiment_id": collection["experiment_id"], "candidate_ids": ["c1"], "idempotency_key": "denied",
+        })
+    assert denied.value.status_code == 403
+    reader, _ = await credentials(app, {"research:read"})
+    await invoke(app, reader, "get_template_candidates", {"experiment_id": collection["experiment_id"]})
+    with pytest.raises(HTTPException):
+        await app.state.mcp.invoke(reader, "expand_research_template", args)

backend/tests/test_ppac_candidate_migration.py

@@ -51,5 +51,6 @@ def test_ppac_backfill_preserves_evidence_across_batches_and_downgrade(tmp_path,
             assert len(rows) == 503
             for i, row in enumerate(rows):
                 assert dict(row) == {**original[i], "check_type": patterns[i % len(patterns)][position]}
+    command.upgrade(config, "head")
     command.check(config)
     engine.dispose()

backend/tests/test_research_integration.py

@@ -274,3 +274,58 @@ async def test_new_interfaces_require_login_and_same_origin(client):
             json=construction("none"),
         )
     ).status_code == 403
+
+
+@pytest.mark.parametrize("decision", ["approve", "deny", "tamper", "archive"])
+async def test_template_collection_single_confirmation(app, logged_in, fixed_input, decision):
+    from app.models import ResearchExperiment
+    from tests.test_research_workspace import expansion
+
+    await setup(app)
+    await configure(app, logged_in)
+    body = expansion(fixed_input["id"])
+    body["template"]["expression"] = "rank({field}) + {offset}"
+    body["template"]["variables"]["offset"] = {"kind": "integer", "values": list(range(20))}
+    response = await logged_in.post("/api/v1/research/experiments", json=body)
+    assert response.status_code == 201, response.text
+    experiment = response.json()
+    assert len(experiment["candidates"]) == 40
+    app.state.ai.model_factory = single_tool_factory("start_template_backtest", {
+        "experiment_id": experiment["id"],
+        "candidate_ids": [c["client_item_id"] for c in experiment["candidates"]],
+        "idempotency_key": "template-confirm",
+    })
+    conversation = (await logged_in.post("/api/v1/ai/conversations")).json()["id"]
+    run = await ask(logged_in, conversation, "回测这个模板集合")
+    assert run["status"] == "waiting_approval", run
+    assert len(run["tools"]) == 1
+    approval = run["tools"][0]
+    preview = approval["preview"]["backtest"]
+    assert preview["total"] == 40 and len(preview["items"]) < 40
+    async with app.state.sessions.begin() as db:
+        assert await db.scalar(select(func.count()).select_from(BacktestRun)) == 0
+        assert await db.scalar(select(func.count()).select_from(BacktestPreview)) == 1
+        if decision == "tamper":
+            saved = await db.get(BacktestPreview, preview["preview_id"])
+            candidates = copy.deepcopy(saved.candidates)
+            candidates[-1]["expression"] = "rank(close)"
+            saved.candidates = candidates
+        elif decision == "archive":
+            saved = await db.get(ResearchExperiment, experiment["id"])
+            saved.archived = True
+    for _ in range(2):
+        result = await logged_in.post(f"/api/v1/ai/approvals/{approval['id']}/decision",
+                                     json={"approved": decision != "deny"})
+        assert result.status_code == 200, result.text
+    async with app.state.sessions() as db:
+        runs = (await db.scalars(select(BacktestRun))).all()
+        assert len(runs) == (1 if decision == "approve" else 0)
+        assert await db.scalar(select(func.count()).select_from(BacktestPreview)) == 1
+        if runs:
+            assert runs[0].preview_id == preview["preview_id"]
+            assert runs[0].source["kind"] == "template"
+            assert runs[0].source["research_id"] == experiment["id"]
+            assert runs[0].ai_context["conversation_id"] == conversation
+            assert runs[0].ai_context["ai_run_id"] == run["id"]
+            saved = await db.get(BacktestPreview, runs[0].preview_id)
+            assert len(saved.candidates) == 40

backend/tests/test_research_workspace.py

@@ -8,7 +8,7 @@ from app.catalog.research_metadata import ResearchMetadata
 from app.models import BacktestRun, CatalogResource, ResearchExperiment
 from app.research.expressions import analyze, expand
 from tests.conftest import alpha
-from tests.test_backtests import execute, setup, start
+from tests.test_backtests import execute, setup
 from tests.test_catalog import SCOPE, prepare, sync
 from tests.test_catalog import catalog as catalog_fixture
 
@@ -120,7 +120,7 @@ def test_bounded_sampling_and_repeated_placeholders():
         expand(expression, values, "all", 100)
 
 
-async def test_template_version_expansion_preview_and_backtest(app, logged_in, research_input):
+async def test_template_version_expansion_direct_backtest_and_idempotency(app, logged_in, research_input):
     saved = await logged_in.post("/api/v1/research/assets", json={"kind": "template", "content": template()})
     assert saved.status_code == 201, saved.text
     asset = saved.json()
@@ -130,7 +130,7 @@ async def test_template_version_expansion_preview_and_backtest(app, logged_in, r
     assert generated.status_code == 201, generated.text
     experiment = generated.json()
     assert len(experiment["candidates"]) == 2
-    assert all(c["validation"]["status"] == "valid" for c in experiment["candidates"])
+    assert all("validation" not in c for c in experiment["candidates"])
     modified = template()
     modified["expression"] = "-rank({field})"
     response = await logged_in.put(
@@ -147,10 +147,24 @@ async def test_template_version_expansion_preview_and_backtest(app, logged_in, r
         )
     ).status_code == 409
     platform, lane = await setup(app)
-    preview = await logged_in.post(f"/api/v1/research/experiments/{experiment['id']}/preview", json={})
+    from app.models import BacktestPreview
-    assert preview.status_code == 201, preview.text
+
-    assert not platform.posts
+    url = f"/api/v1/research/experiments/{experiment['id']}/backtest"
-    run = await start(logged_in, preview.json(), "research-stage-one")
+    request = {"candidate_ids": ["c2"], "idempotency_key": "template-confirmation"}
+    result = await logged_in.post(url, json=request)
+    assert result.status_code == 202, result.text
+    run = result.json()
+    assert run["total"] == 1
+    assert run["source"]["kind"] == "template"
+    assert run["source"]["input_snapshot_ids"] == [research_input["id"]]
+    retry = await logged_in.post(url, json=request)
+    assert retry.status_code == 202 and retry.json()["backtest_run_id"] == run["backtest_run_id"]
+    conflict = await logged_in.post(url, json={**request, "candidate_ids": ["c1"]})
+    assert conflict.status_code == 409
+    async with app.state.sessions() as db:
+        assert await db.scalar(select(func.count()).select_from(BacktestRun)) == 1
+        assert await db.scalar(select(func.count()).select_from(BacktestPreview)) == 1
+    assert lane.wake.is_set()
     await execute(app, lane, run["backtest_run_id"])
     results = (await logged_in.get(f"/api/v1/backtests/runs/{run['backtest_run_id']}/results")).json()
     aid = results["items"][0]["alpha_id"]
@@ -161,22 +175,34 @@ async def test_template_version_expansion_preview_and_backtest(app, logged_in, r
     assert old["backtest_run_ids"] == [run["backtest_run_id"]]
 
 
-async def test_invalid_fields_and_unknown_operators_never_start(app, logged_in, research_input):
+async def test_template_generation_does_not_require_field_operator_or_settings_evidence(app, logged_in, research_input):
+    from sqlalchemy import delete
+
+    async with app.state.sessions.begin() as db:
+        await db.execute(delete(CatalogResource))
     body = expansion(research_input["id"])
     body["template"]["variables"]["field"]["values"] = ["other_field"]
-    assert (await logged_in.post("/api/v1/research/experiments", json=body)).status_code == 422
+    body["template"]["expression"] = "made_up({field}) + vec_avg(TEST_FIN_001)"
-    async with app.state.sessions() as db:
-        assert await db.scalar(select(func.count()).select_from(ResearchExperiment)) == 0
-    body = expansion(research_input["id"])
-    body["template"]["expression"] = "made_up({field})"
     response = await logged_in.post("/api/v1/research/experiments", json=body)
     assert response.status_code == 201, response.text
-    eid = response.json()["id"]
+    experiment = response.json()
-    assert (await logged_in.post(f"/api/v1/research/experiments/{eid}/preview", json={})).status_code == 422
+    assert all("validation" not in c for c in experiment["candidates"])
+    assert not {"operators_snapshot", "settings_snapshot", "field_availability"} & experiment["evidence"].keys()
+    assert (await logged_in.post(f"/api/v1/research/experiments/{experiment['id']}/preview", json={})).status_code == 201
     async with app.state.sessions() as db:
         assert await db.scalar(select(func.count()).select_from(BacktestRun)) == 0
 
 
+@pytest.mark.parametrize("expression", ["rank({field}", "rank({field},,)", "x = {field}"])
+async def test_template_syntax_errors_reject_whole_collection(app, logged_in, research_input, expression):
+    body = expansion(research_input["id"])
+    body["template"]["expression"] = expression
+    response = await logged_in.post("/api/v1/research/experiments", json=body)
+    assert response.status_code == 422 and "语法错误" in response.text
+    async with app.state.sessions() as db:
+        assert await db.scalar(select(func.count()).select_from(ResearchExperiment)) == 0
+
+
 async def test_import_preview_conflict_and_explicit_commit(logged_in):
     legacy = {
         "name": "legacy",
@@ -367,7 +393,7 @@ def test_actual_cnhk_setting_choice_nesting_is_supported():
     )
 
 
-async def test_published_input_does_not_override_conflicting_field_evidence(app, logged_in, research_input):
+async def test_template_candidates_ignore_conflicting_field_evidence(app, logged_in, research_input):
     async with app.state.sessions.begin() as db:
         await ResearchMetadata(db).publish(
             "availability-fixture",
@@ -382,12 +408,11 @@ async def test_published_input_does_not_override_conflicting_field_evidence(app,
     experiment = (
         await logged_in.post("/api/v1/research/experiments", json=expansion(research_input["id"]))
     ).json()
-    assert experiment["candidates"][0]["validation"]["status"] == "needs_review"
+    assert all("validation" not in c for c in experiment["candidates"])
-    assert experiment["candidates"][1]["validation"]["status"] == "valid"
     denied = await logged_in.post(
         f"/api/v1/research/experiments/{experiment['id']}/preview", json={"candidate_ids": ["c1"]}
     )
-    assert denied.status_code == 422
+    assert denied.status_code == 201
     duplicate = await logged_in.post(
         f"/api/v1/research/experiments/{experiment['id']}/preview", json={"candidate_ids": ["c2", "c2"]}
     )
@@ -526,3 +551,131 @@ async def test_model_cannot_append_preparation_references_to_fixed_inputs(app, l
     assert response.status_code == 422 and "不能改变" in response.text
     async with app.state.sessions() as db:
         assert not await db.scalar(select(ResearchAsset).where(ResearchAsset.name == "untrusted"))
+
+
+async def test_template_definitions_bind_selected_fields_without_mutating_asset(logged_in, research_input):
+    content = template()
+    content["expression"] = "rank({field}) + rank({field})"
+    content["variables"]["field"] = {"kind": "field", "field_type": "MATRIX", "description": "横截面字段"}
+    saved = await logged_in.post("/api/v1/research/assets", json={"kind": "template", "content": content})
+    assert saved.status_code == 201, saved.text
+    asset = saved.json()
+    assert asset["content"]["variables"]["field"] == {
+        "kind": "field", "field_type": "MATRIX", "description": "横截面字段", "values": [],
+    }
+    body = expansion(research_input["id"], asset_id=asset["id"], version=1, mode="random", limit=2, seed=19)
+    body.pop("template")
+    first = await logged_in.post("/api/v1/research/experiments", json=body)
+    second = await logged_in.post("/api/v1/research/experiments", json=body)
+    assert first.status_code == second.status_code == 201, first.text
+    assert first.json()["candidates"] == second.json()["candidates"]
+    assert len(first.json()["candidates"]) == 2
+    for candidate in first.json()["candidates"]:
+        field = candidate["bindings"]["field"]
+        assert research_input["field_types"][field] == "MATRIX"
+        assert candidate["expression"] == f"rank({field}) + rank({field})"
+    stored = (await logged_in.get(f"/api/v1/research/assets/{asset['id']}")).json()
+    assert stored["content"] == asset["content"]
+
+
+async def test_empty_template_domains_fail_expansion_with_actionable_errors(logged_in, research_input):
+    body = expansion(research_input["id"])
+    body["template"]["variables"]["field"] = {"kind": "field", "field_type": "GROUP"}
+    response = await logged_in.post("/api/v1/research/experiments", json=body)
+    assert response.status_code == 422 and "没有匹配的 GROUP 字段" in response.text
+    body["template"]["expression"] = "ts_mean(TEST_FIN_001, {window})"
+    body["template"]["variables"] = {"window": {"kind": "integer", "description": "时间窗口"}}
+    response = await logged_in.post("/api/v1/research/experiments", json=body)
+    assert response.status_code == 422 and "变量 window 缺少候选取值" in response.text
+
+
+async def test_native_bot_creates_template_and_immutable_version(app, logged_in):
+    from fastapi import HTTPException
+
+    from app.ai.capabilities import ToolContext
+    from app.ai.tools import CAPABILITIES
+    from app.business import Business
+
+    async with app.state.sessions.begin() as db:
+        ctx = ToolContext(Business(db))
+        content = template()
+        content["variables"]["field"].pop("values")
+        created = await CAPABILITIES["create_research_template"].invoke(ctx, content)
+        content["description"] = "修订后的研究解释"
+        args = {"asset_id": created["id"], "version": 1, "content": content}
+        updated = await CAPABILITIES["create_research_template_version"].invoke(ctx, args)
+        assert updated["version"] == 2
+        with pytest.raises(HTTPException) as conflict:
+            await CAPABILITIES["create_research_template_version"].invoke(ctx, args)
+        assert conflict.value.status_code == 409
+    old = (await logged_in.get(f"/api/v1/research/assets/{created['id']}?version=1")).json()
+    assert old["content"]["description"] == "测试经济假设"
+
+
+@pytest.mark.parametrize("candidate_settings", [{"region": "EUR"}, {"universe": "TOP1000"}, {"delay": 0}])
+async def test_preparation_scope_must_still_match_candidate_settings(app, logged_in, research_input, candidate_settings):
+    body = expansion(research_input["id"])
+    body["input_ids"] = []
+    body["preparation_refs"] = [{"id": research_input["preparation_id"], "version": research_input["preparation_version"]}]
+    body["settings"].update(candidate_settings)
+    response = await logged_in.post("/api/v1/research/experiments", json=body)
+    assert response.status_code == 422 and "输入快照与研究范围不一致" in response.text
+    async with app.state.sessions() as db:
+        assert await db.scalar(select(func.count()).select_from(ResearchExperiment)) == 0
+
+
+async def test_template_contract_rejects_removed_scope(logged_in):
+    from app.research.workspace_contracts import TemplateSpec
+
+    content = template() | {"scope": SCOPE}
+    response = await logged_in.post("/api/v1/research/assets", json={"kind": "template", "content": content})
+    assert response.status_code == 422, response.text
+    assert "scope" in response.text
+    assert "scope" not in TemplateSpec.model_json_schema()["properties"]
+
+
+async def test_template_start_checks_selection_and_ignores_old_row_validation(app, logged_in, research_input):
+    from app.models import BacktestPreview
+
+    experiment = (await logged_in.post("/api/v1/research/experiments", json=expansion(research_input["id"]))).json()
+    async with app.state.sessions.begin() as db:
+        row = await db.get(ResearchExperiment, experiment["id"])
+        row.candidates = [{**c, "validation": {"status": "needs_review", "syntax": [], "types": [],
+                                              "availability": ["历史字段未核实"]}} for c in row.candidates]
+    await setup(app)
+    url = f"/api/v1/research/experiments/{experiment['id']}/backtest"
+    for ids in [[], ["unknown"], ["c1", "c1"]]:
+        response = await logged_in.post(url, json={"candidate_ids": ids, "idempotency_key": "confirm-old"})
+        assert response.status_code == 422, response.text
+    async with app.state.sessions() as db:
+        assert await db.scalar(select(func.count()).select_from(BacktestPreview)) == 0
+        assert await db.scalar(select(func.count()).select_from(BacktestRun)) == 0
+    response = await logged_in.post(url, json={"candidate_ids": ["c1", "c2"], "idempotency_key": "confirm-old"})
+    assert response.status_code == 202 and response.json()["total"] == 2
+    retry = await logged_in.post(url, json={"candidate_ids": ["c2", "c1"], "idempotency_key": "confirm-old"})
+    assert retry.status_code == 202 and retry.json()["backtest_run_id"] == response.json()["backtest_run_id"]
+
+
+@pytest.mark.parametrize("change", ["syntax", "scope", "archived", "disconnected"])
+async def test_template_start_rejects_unusable_collection_without_partial_writes(app, logged_in, research_input, change):
+    from app.models import Account, BacktestPreview
+
+    experiment = (await logged_in.post("/api/v1/research/experiments", json=expansion(research_input["id"]))).json()
+    async with app.state.sessions.begin() as db:
+        row = await db.get(ResearchExperiment, experiment["id"])
+        if change == "syntax":
+            row.candidates = [{**c, "expression": "rank("} for c in row.candidates]
+        elif change == "scope":
+            row.candidates = [{**c, "settings": {**c["settings"], "region": "EUR"}} for c in row.candidates]
+        elif change == "archived":
+            row.archived = True
+        else:
+            (await db.get(Account, 1)).connection_status = "disconnected"
+    app.state.runner.backtests.wake.clear()
+    response = await logged_in.post(f"/api/v1/research/experiments/{experiment['id']}/backtest",
+                                    json={"candidate_ids": ["c1"], "idempotency_key": "invalid"})
+    assert response.status_code == (422 if change in ("syntax", "scope") else 409), response.text
+    assert not app.state.runner.backtests.wake.is_set()
+    async with app.state.sessions() as db:
+        assert await db.scalar(select(func.count()).select_from(BacktestPreview)) == 0
+        assert await db.scalar(select(func.count()).select_from(BacktestRun)) == 0

backend/tests/test_template_scope_migration.py

@@ -0,0 +1,87 @@
+"""Retire template scope once in storage, without a runtime compatibility path."""
+
+from pathlib import Path
+
+import sqlalchemy as sa
+from alembic import command
+from alembic.config import Config
+from cryptography.fernet import Fernet
+from sqlalchemy.orm import Session
+
+from app.models import (
+    Account,
+    ResearchAsset,
+    ResearchExperiment,
+    ResearchFlowRun,
+    ResearchRequest,
+    ResearchRevision,
+    ResearchStepRun,
+)
+from app.research.workspace_contracts import FeatureSpec, TemplateSpec
+
+
+def test_remove_template_scope_preserves_execution_scopes_and_all_versions(tmp_path, monkeypatch):
+    database = tmp_path / "template-scopes.db"
+    monkeypatch.setenv("DATABASE_URL", f"sqlite+aiosqlite:///{database}")
+    monkeypatch.setenv("ADMIN_PASSWORD", "migration-test-only")
+    monkeypatch.setenv("ENCRYPTION_KEY", Fernet.generate_key().decode())
+    monkeypatch.setenv("WQ_EMAIL", "")
+    monkeypatch.setenv("WQ_PASSWORD", "")
+    root = Path(__file__).resolve().parents[1]
+    config = Config(str(root / "alembic.ini"))
+    config.set_main_option("script_location", str(root / "migrations"))
+    command.upgrade(config, "0022")
+    engine = sa.create_engine(f"sqlite:///{database}")
+    scope = {"instrument_type": "EQUITY", "region": "EUR", "universe": "TOP2500", "delay": 1}
+    clean = {"name": "template", "expression": "rank({scope})", "description": "scope is also a variable name",
+             "variables": {"scope": {"kind": "field", "field_type": "MATRIX", "values": []}}}
+    old = {**clean, "scope": scope}
+    feature = {"name": "feature", "hypothesis": "test", "template": old}
+    clean_feature = {**feature, "template": clean}
+    payload = {"template": {"id": "a", "version": 1, "content": old}, "feature": {"content": feature},
+               "inputs": [{"scope": scope, "field_ids": ["close"]}], "scope": scope,
+               "bindings": {"scope": "close"}, "items": [old, None, "unchanged"]}
+    expected = {**payload, "template": {**payload["template"], "content": clean},
+                "feature": {"content": clean_feature}, "items": [clean, None, "unchanged"]}
+    with Session(engine) as db, db.begin():
+        db.add(Account(id=1))
+        db.add_all([ResearchAsset(id="a", kind="template", name="template", version=501),
+                    ResearchAsset(id="b", kind="feature", name="feature", version=2)])
+        db.flush()
+        db.add_all([ResearchRevision(asset_id="a", version=i, content={**old, "scope": scope if i % 2 else None},
+                                    provenance=payload) for i in range(1, 502)])
+        db.add_all([ResearchRevision(asset_id="b", version=i, content=feature, provenance=payload) for i in (1, 2)])
+        db.add(ResearchExperiment(id="experiment", name="test", kind="template", hypothesis="test",
+                                  inputs=payload["inputs"], parents=[payload], candidates=[{"settings": scope}], evidence=payload))
+        db.add(ResearchFlowRun(id="flow", request_id="request", name="test", definition=payload, authorization=payload))
+        db.flush()
+        db.add(ResearchStepRun(id="step", run_id="flow", node_id="expand", round=1, output=payload))
+        db.add(ResearchRequest(id="request", account_id=1, operation="create_research_template",
+                               idempotency_key="key", digest="a" * 64, business_id="a", response=payload))
+    # Include every stored column in the comparison, not just the removed key.
+    tables = {name: sa.Table(name, sa.MetaData(), autoload_with=engine) for name in (
+        "research_revisions", "research_experiments", "research_flow_runs", "research_step_runs", "research_requests",
+    )}
+    def snapshots():
+        with engine.connect() as db:
+            return {name: [dict(row) for row in db.execute(sa.select(table).order_by(*table.primary_key.columns)).mappings()]
+                    for name, table in tables.items()}
+    before = snapshots()
+    wanted = {name: [dict(row) for row in rows] for name, rows in before.items()}
+    for row in wanted["research_revisions"]:
+        row.update(content=clean if row["asset_id"] == "a" else clean_feature, provenance=expected)
+    wanted["research_experiments"][0].update(evidence=expected, parents=[expected])
+    wanted["research_flow_runs"][0].update(definition=expected, authorization=expected)
+    wanted["research_step_runs"][0].update(output=expected)
+    wanted["research_requests"][0].update(response=expected)
+    command.upgrade(config, "head")
+    assert snapshots() == wanted
+    for row in wanted["research_revisions"]:
+        (TemplateSpec if row["asset_id"] == "a" else FeatureSpec).model_validate(row["content"])
+    # Downgrade does not invent the deleted metadata; re-upgrade is idempotent.
+    command.downgrade(config, "0022")
+    assert snapshots() == wanted
+    command.upgrade(config, "head")
+    assert snapshots() == wanted
+    command.check(config)
+    engine.dispose()

compose.production.yaml

@@ -1,6 +1,17 @@
 # Independent production configuration; do not merge with compose.yaml.
 name: wq-alpha-production
 
+x-observability-labels: &observability-labels
+  observability.logs: "true"
+  observability.project: "wq-alpha"
+  observability.env: "production"
+
+x-container-logging: &container-logging
+  driver: json-file
+  options:
+    max-size: "20m"
+    max-file: "5"
+
 x-backend: &backend
   image: wq-alpha-production-backend:${DEPLOY_TAG:-local}
   build:
@@ -21,13 +32,15 @@ x-backend: &backend
     AI_OUTPUT_TOKENS: ${AI_OUTPUT_TOKENS:-4096}
     AI_TIMEOUT: ${AI_TIMEOUT:-180}
     WQ_BASE_URL: ${WQ_BASE_URL:-https://api.worldquantbrain.com}
-  networks:
+  # Reach the remote PostgreSQL endpoint through the default network's egress.
-    - default
-    - database
 
 services:
   backend:
     <<: *backend
+    labels:
+      <<: *observability-labels
+      observability.service: "backend"
+    logging: *container-logging
     # Migration is run separately with all application writers stopped.
     command: [uvicorn, 'app.main:create_app', --factory, --host, 0.0.0.0, --port, '8000', --workers, '1', --proxy-headers]
     healthcheck:
@@ -40,10 +53,18 @@ services:
     restart: unless-stopped
   migrate:
     <<: *backend
+    labels:
+      <<: *observability-labels
+      observability.service: "migrate"
+    logging: *container-logging
     profiles: [jobs]
     command: [alembic, upgrade, head]
     restart: 'no'
   web:
+    labels:
+      <<: *observability-labels
+      observability.service: "web"
+    logging: *container-logging
     image: wq-alpha-production-web:${DEPLOY_TAG:-local}
     build:
       context: .
@@ -66,10 +87,6 @@ services:
       start_period: 10s
     restart: unless-stopped
 
-networks:
-  database:
-    external: true
-    name: ${DATABASE_NETWORK:?Set the existing PostgreSQL Docker network}
 volumes:
   caddy_data:
   caddy_config:

docs/deployment-gitea.md

@@ -1,6 +1,8 @@
 # Gitea 生产部署
 
-本方案在目标 Linux 服务器上由已有 Gitea Runner 构建并启动 Docker Compose，复用已有 PostgreSQL。入口绑定 `127.0.0.1:8112`，由宿主机反向代理提供公网 HTTPS。现有 `compose.yaml`、`compose.public.yaml` 保持独立，不与生产文件叠加。
+本方案使用两台服务器：A 服务器托管 Gitea 和项目仓库；B 服务器运行 `tencent-prod-runner`，拉取 A 上的 `main` 分支代码，在 B 上构建镜像并启动 Docker Compose。工作流通过标签 `tencent-prod` 选择该运行器。
+
+PostgreSQL 位于独立数据库服务器或云数据库，B 上的应用通过 `DATABASE_URL` 访问它，不依赖同机数据库容器或外部 Docker 网络。应用入口绑定 **B 服务器**的 `127.0.0.1:8112`，由 B 的宿主机反向代理提供公网 HTTPS。现有 `compose.yaml`、`compose.public.yaml` 保持独立，不与生产文件叠加。
 
 ## 1. 填写配置和数据库账号密码
 
@@ -8,22 +10,23 @@
 
 | 位置 | 名称 | 填写内容 |
 | --- | --- | --- |
-| Secrets（必填） | `DATABASE_URL` | `postgresql+asyncpg://数据库账号:数据库密码@数据库网络别名:5432/数据库名` |
+| Secrets（必填） | `DATABASE_URL` | `postgresql+asyncpg://数据库账号:数据库密码@远程数据库域名或IP:端口/数据库名`，地址须能从 B 的应用容器访问 |
 | Secrets（必填） | `ADMIN_PASSWORD` | 系统初始管理员密码，至少 12 字符，与数据库密码独立 |
 | Secrets（必填） | `ENCRYPTION_KEY` | 下面命令生成的 Fernet 密钥，升级保持不变 |
 | Secrets（必填） | `WQ_EMAIL` | WorldQuant 登录邮箱 |
 | Secrets（必填） | `WQ_PASSWORD` | WorldQuant 登录密码，按原样填写，不加引号 |
-| Variables（必填） | `DATABASE_NETWORK` | PostgreSQL 所在的现有 Docker 网络，例如 `1panel-network` |
 | Variables（必填） | `PUBLIC_ORIGIN` | 实际 HTTPS 来源，例如 `https://alpha.your-domain.com`，不带路径或末尾 `/` |
 | Variables（可选） | `ADMIN_USERNAME` | 初始管理员账号，默认 `admin` |
 | Variables（可选） | `MCP_ENABLED` | 设置 `true` 启用 MCP；未配置时默认关闭 |
 
-端口与 AI 限制使用 `compose.production.yaml` 的默认值，不需要在 Gitea 配置：`WEB_PORT=8112`、`AI_REQUEST_LIMIT=12`、`AI_TOOL_LIMIT=12`、`AI_OUTPUT_TOKENS=4096`、`AI_TIMEOUT=180`。需要调整时修改 Compose 中对应默认值；工作流不再读取这些同名 Gitea Variables。
+端口与 AI 限制使用 `compose.production.yaml` 的默认值，不需要在 Gitea 配置：`WEB_PORT=8112`、`AI_REQUEST_LIMIT=12`、`AI_TOOL_LIMIT=12`、`AI_OUTPUT_TOKENS=4096`、`AI_TIMEOUT=180`。需要调整时修改 Compose 中对应默认值；工作流不再读取这些同名 Gitea Variables。原有 `DATABASE_NETWORK` 不再使用，无需在 B 上创建同名外部网络。
+
+已配置的 `PROD_SSH_KEY`、`PROD_HOST`、`PROD_PORT`、`PROD_USER`、`PROD_KNOWN_HOSTS` 可以继续保留供远程运维使用。本流程由 B 的 Runner 直接部署，不读取这些 SSH Secrets；镜像在 B 本机构建和使用，也不读取 `REGISTRY_USERNAME`、`REGISTRY_PASSWORD`。
 
 数据库连接串示例（实际填写时替换示例值）：
 
 ```text
-postgresql+asyncpg://wq_user:YOUR_PASSWORD@postgresql:5432/wq_alpha
+postgresql+asyncpg://wq_user:YOUR_PASSWORD@db.example.com:5432/wq_alpha
 ```
 
 生成加密密钥：
@@ -34,7 +37,9 @@ python3 -c 'import base64,secrets; print(base64.urlsafe_b64encode(secrets.token_
 
 `DATABASE_URL` 的用户名和密码中的特殊字符须做 URL 百分号编码，例如 `@` → `%40`、`#` → `%23`、`/` → `%2F`、`%` → `%25`。在 Gitea 输入框填写值本身，不添加包裹引号，不填写 `DATABASE_URL=` 前缀。管理员密码中的 `$`、引号等按原样填写，由环境变量传递，无需 shell 转义。不要在日志打印变量，或开启 `set -x`。
 
-在现有 PostgreSQL 管理界面先创建专用数据库与账号，账号须可连接并拥有该库中应用 schema 的建表和迁移权限。确认网络存在、数据库别名可解析、PostgreSQL 允许该 Docker 网络访问。应用使用异步驱动 `asyncpg`；远程数据库或强制 TLS 的实例需另外按该实例的 TLS 配置调整连接参数，此模板默认同机 Docker 网络。
+在远程 PostgreSQL 管理界面先创建专用数据库与账号，账号须可连接并拥有该库中应用 schema 的建表和迁移权限。数据库安全组、白名单或防火墙须允许 B 的实际出口地址访问数据库端口，且 B 的应用容器须能解析并访问数据库地址；使用私网地址时先确保 B 与数据库的私网路由连通。不要在连接串中沿用只在 A 上可解析的数据库容器名，也不要使用 `localhost` 或 `127.0.0.1` 指向远程数据库。
+
+应用使用异步驱动 `asyncpg`；强制 TLS 或自定义 CA 的实例需按该实例要求配置兼容的连接参数及证书，不应直接假定其控制台提供的其他驱动连接串可用。部署脚本会在停止旧服务前，从 B 的后端容器进行真实数据库连接预检。
 
 如果从已有安装迁移数据，需恢复完整数据库并沿用原 `ENCRYPTION_KEY`；修改管理员环境变量不会重置已有管理员密码。
 
@@ -42,19 +47,27 @@ WorldQuant 凭据由环境变量管理，启动时加密写入数据库；页面
 
 ## 2. 配置 Gitea Runner
 
-沿用已成功部署 `zhixing-system` 的运行器标签 **`ubuntu-latest`**，无需注册新的 `wq-production` 运行器。该运行器的执行环境须能通过 Docker CLI/Compose 操作同一台 1Panel 服务器的 Docker daemon；可以复用现有 Docker 连接方式，不强制 host 模式。
+使用 B 服务器上已注册的 **`tencent-prod-runner`**，工作流的 `runs-on` 填写其标签 **`tencent-prod`**。只有 B 上用于生产部署的运行器应持有此标签。
+
+运行器的执行环境须通过 Docker CLI/Compose 操作 **B 宿主机的 Docker daemon**。可以沿用现有 host 或 Docker 执行模式：host 模式需要执行用户有权访问 B 的 Docker；容器模式需要部署步骤所在的作业容器能够访问 B 的 Docker socket（通常是 `/var/run/docker.sock`），且具有 Docker CLI/Compose。只给 Runner 管理容器挂载 socket 并不能证明作业容器内也可访问；不要将部署命令连接到 A 的远程 Docker 或临时 Docker-in-Docker 实例。
+
+Runner 注册的 Gitea 地址及仓库克隆地址必须能从 B 和作业执行环境访问，应使用 A 的实际域名或可达 IP；不能使用指向 A 本机的 `localhost`、`127.0.0.1` 或仅在 A 的 Docker 网络内可解析的容器名。
 
 执行环境需要 Git、Bash、Node.js 20（checkout v4）以及支持 `up --wait --wait-timeout` 的 Docker Compose。本流程不需要 `flock` 或预建 `/opt/wq-alpha`。服务器需能访问 checkout action、基础镜像仓库和依赖源。
 
-在仓库启用 Actions，`main` push 或手动运行会部署。Secrets 通过部署步骤的环境变量注入，`--env-file /dev/null` 防止误读开发 `.env`。构建可并行；构建完成后，以 Docker 唯一容器名 `wq-alpha-production-deploy-lock` 互斥保护预检、迁移和服务切换。锁容器不启动、不携带凭据，正常结束或失败时删除；竞争失败的任务不会删除其他任务的锁。
+在仓库启用 Actions，向 **`main`** push 会部署；手动运行也须选择 **`main`**，选择其他分支时本版工作流会跳过部署任务。Checkout 使用本次触发的提交，构建和迁移均在 B 的执行环境完成，无需 A 通过 SSH 再复制代码到 B。
+
+发布统一使用 `main` 分支。`tencent-prod` 是 B 服务器的 Runner 标签，与 Git 分支名无关。
+
+Secrets 仍在 A 的 Gitea 仓库设置中管理，通过部署步骤的环境变量注入 B 上的作业，`--env-file /dev/null` 防止误读开发 `.env`。构建可并行；构建完成后，以 B 的 Docker 唯一容器名 `wq-alpha-production-deploy-lock` 互斥保护预检、迁移和服务切换。锁容器不启动、不携带凭据，正常结束或失败时删除；竞争失败的任务不会删除其他任务的锁。
 
 如果 Runner 被强制终止或 Docker 断连，可能留下锁容器。先确认没有本项目部署正在执行，再手动运行 `docker rm wq-alpha-production-deploy-lock` 后重试。不要在部署进行时删除锁。
 
 ## 3. 配置反向代理并首次运行
 
-为 `PUBLIC_ORIGIN` 对应域名配置 HTTPS，转发至 **`http://127.0.0.1:8112`**（或你的 `WEB_PORT`）。代理须运行在宿主机网络中；独立 bridge 容器中的 `127.0.0.1` 不指向宿主机。若使用容器化 1Panel/OpenResty，先确认其网络模式。AI 流式响应需要关闭代理缓冲，并允许长连接/足够长的读取超时。
+将 `PUBLIC_ORIGIN` 对应域名解析到 **B 服务器**，在 B 配置 HTTPS 反向代理，转发至 **`http://127.0.0.1:8112`**（或你的 `WEB_PORT`）。此回环地址只对 B 本机有效；A 上的代理不能通过它访问 B。代理须运行在 B 的宿主机网络中；独立 bridge 容器中的 `127.0.0.1` 不指向宿主机。若使用容器化 1Panel/OpenResty，先确认其网络模式。AI 流式响应需要关闭代理缓冲，并允许长连接/足够长的读取超时。
 
-首次部署建议在 Gitea Actions 页面手动运行工作流。需要在服务器排障运行时，须先从安全渠道将上表必填配置注入当前进程环境，再执行：
+首次部署建议在 Gitea Actions 页面选择 **`main`** 分支手动运行工作流，确认任务由 `tencent-prod-runner` 接收。需要在服务器排障运行时，在 **B 服务器的 `main` 检出目录**中先从安全渠道将上表必填配置注入当前进程环境，再执行：
 
 ```bash
 bash scripts/deploy-production.sh
@@ -62,6 +75,8 @@ bash scripts/deploy-production.sh
 
 固定 Compose 项目名为 `wq-alpha-production`，后端保持一个实例、一个 worker。脚本先验证 Compose，构建按 Git 提交标记的镜像，再检查密钥格式和数据库连通性；随后停止 web/backend、执行一次性迁移、启动服务并等待健康检查。Web 健康检查同时覆盖页面和经 Caddy 转发的数据库健康接口。迁移或启动失败时非零退出并显示容器状态，不继续标记成功。
 
+从旧服务器切换时，若旧实例仍连接同一生产数据库，须在 B 上执行迁移前停止旧实例及其定时任务。脚本和部署锁只管理 B 上的容器，不会停止 A 上的旧应用，也不能协调两台服务器对同一数据库的迁移和写入。
+
 首次部署完成后，检查公网 `/api/v1/health`，再通过真实域名登录并确认页面、写请求和 Cookie 正常。容器健康通过不能替代 HTTPS、DNS 和真实 Gitea Runner 验收。
 
 ## 4. 升级、备份和失败处理
@@ -72,7 +87,7 @@ bash scripts/deploy-production.sh
 
 如旧代码与当前 schema 兼容，可检出旧提交并指定其镜像标签启动；否则先停止应用，使用经过验证的备份恢复数据库，再用原加密密钥和对应旧版本启动。数据库恢复会丢失备份后的写入，必须人工确认后执行。本配置没有自动数据库降级，也不承诺无停机升级。
 
-以下排查命令不依赖凭据环境变量（可能含业务信息的日志请勿公开）：
+以下排查命令在 **B 服务器**运行，不依赖凭据环境变量（可能含业务信息的日志请勿公开）：
 
 ```bash
 docker ps -a --filter label=com.docker.compose.project=wq-alpha-production
@@ -84,13 +99,13 @@ docker logs --tail=100 wq-alpha-production-web-1
 
 ## 配置依据
 
-采用 compound 经验 1 中的生产/开发隔离、显式外部网络、必填凭据、固定项目名、独立迁移和部署健康检查。未照搬旧项目端口、业务 Job 或卷权限初始化：本应用后端不写 named volume，已有镜像使用 UID 10001。
+生产/开发配置隔离，使用必填凭据、固定项目名、独立迁移和部署健康检查。生产服务使用 Compose 默认网络互通，并通过该网络访问远程 PostgreSQL；无需预建数据库 Docker 网络。本应用后端不写 named volume，已有镜像使用 UID 10001。
 
-凭据存放在 Gitea Secrets，通过步骤级环境变量交给 Compose，不落地到配置文件。Docker 容器仍需持有运行时配置，因此拥有 Runner 或 Docker 管理权限的人仍可能读取它们。若此前手动创建了旧 `.env.production`，新流程不再读取它；确认配置已迁移到 Gitea 并妥善备份密钥后可自行移除旧文件。相关官方资料：[Compose 外部网络](https://docs.docker.com/reference/compose-file/networks/)、[环境变量插值](https://docs.docker.com/compose/how-tos/environment-variables/variable-interpolation/)、[Gitea Runner 标签](https://gitea.com/gitea/runner/src/branch/main/README.md)。
+凭据存放在 Gitea Secrets，通过步骤级环境变量交给 Compose，不落地到配置文件。Docker 容器仍需持有运行时配置，因此拥有 Runner 或 Docker 管理权限的人仍可能读取它们。若此前手动创建了旧 `.env.production`，新流程不再读取它；确认配置已迁移到 Gitea 并妥善备份密钥后可自行移除旧文件。相关官方资料：[Compose 网络](https://docs.docker.com/compose/how-tos/networking/)、[环境变量插值](https://docs.docker.com/compose/how-tos/environment-variables/variable-interpolation/)、[Gitea Runner 2.x 标签](https://docs.gitea.com/runner/2/labels/)、[Gitea 分支上下文](https://docs.gitea.com/usage/actions/actions-variables/)。
 
 ## 5. 1Panel 夜间全量目录同步
 
-在 1Panel 的计划任务中建立 Shell 脚本任务，执行周期由 1Panel 设置，例如每天深夜执行。每次命令明确指定一个范围：
+在 **B 服务器**的 1Panel 计划任务中建立 Shell 脚本任务，执行周期由 1Panel 设置，例如每天深夜执行。每次命令明确指定一个范围：
 
 ```bash
 docker exec wq-alpha-production-backend-1 \

docs/worldquant-knowledge-base/01-alpha-ideas/data-template-cards.md

@@ -43,3 +43,17 @@ SUE 类方向的可迁移问题是“盈利变化相对于自身历史波动是
 dateCoverage 只是一项元数据，不告诉你缺失发生在哪几年，也不完整表达横截面质量；VECTOR 非空容器也可能没有有效元素。不要把社区的 40%/60%/80% 建议阈值当官方规则。检查逐年、逐区覆盖及事件时点，比固定门槛更重要。
 
 来源：[【平台新功能】字段新特征：datecoverage的介绍、应用与勘误。](https://support.worldquantbrain.com/hc/en-us/community/posts/37581194344215)（发帖 2026-01-10，社区经验）。
+
+## 2026-09-18 补充：代理变量与样本边界
+
+[搜索量转新闻密度的注意力反转帖](https://support.worldquantbrain.com/hc/en-us/community/posts/43541109631895)（2026-09-16）用新闻事件密度代理搜索关注，作者报告 GBR 与 EUR 表现不同，EUR 变体 Fitness 均低于 1。新闻供给不等于投资者主动搜索，代理替换可能改变机制，应比较原代理、新闻量单腿及价格反转对照；这不是已验证的论文等价复现。
+
+[MAX 与彩票偏好帖](https://support.worldquantbrain.com/hc/en-us/community/posts/43508107778711)（2026-09-15）报告滚动极端涨幅除以自身波动后，Sharpe 从 0.36 到 0.75，最大回撤从 48.96% 到 5.69%；缩短为 5 日窗口虽进一步提高指标，作者也承认与短期反转重叠。可保留“极值相对自身波动”的待验证假设，但窗口更短、更强不等于彩票机制更纯。本轮未复核论文、借券成本或原始回测，也不将标题中的“完整复现”当作已验证结论。
+
+[DEU 内部人信念强度帖](https://support.worldquantbrain.com/hc/en-us/community/posts/43178926004887)（2026-09-02，本轮补录）提醒：稀疏事件填零后，大块相同排名可能代表没有事件，不能解释为经济信念中等。作者还报告有效持仓主要在 2018–2023，而较早年份为零；应以实际可投资区间评估结果。比较滚动和与均值时须核对有效观测数和缺失处理；完整固定窗口下两者只差常数，不能笼统断言换成均值必然丢失事件信息。原帖的 tail 阈值与性能改善仅是单例，不推广为稀疏数据通用模板。
+
+## 2026-09-25 补充：非线性变换与覆盖需要分开检验
+
+[EUR 财报电话会情绪概率差](https://support.worldquantbrain.com/hc/en-us/community/posts/43620895448343)（2026-09-20）报告 signed power 降相关，并描述早期稀疏截面造成权重集中。随附评论明确指出另一字段的成功不能证明“概率差 + signed_power”有效；应固定其余设置分别对照回填窗、覆盖和 truncation，拆开表达式变换与股票池变化。原帖指标属于作者自报，未复现。
+
+[ASI 借券费率方差案例](https://support.worldquantbrain.com/hc/en-us/community/posts/43624312881943)（2026-09-20）比较方差与标准差版本。即使同一有效样本下平方保持非负值的截面排序，后续线性缩放、中性化或权重处理也可能得到不同持仓，不能由“同一排序”推出整条策略等价。作者由少数 checks 推断的 robust 比率约 0.9 仅是其样本观察，不写成通用平台门槛；每次仍需读取实际 value、limit 和状态。

docs/worldquant-knowledge-base/01-alpha-ideas/relationship-and-extension-cards.md

@@ -63,3 +63,19 @@ ASI/TOP500 的 risk70 模板使用时间操作后按 exchange 中性化。来源
 ## 10. 当前优先级
 
 已有九月四张卡覆盖 SNR、机构拥挤、Wikipedia 关注和订单流。新补充方向优先选择“极性配对、同行相对位置、明确口径的实际/预测偏差”中你能取得完整数据的少量假设。优先级属于本库研究建议，未实际运行或比较这些方向的收益。
+
+## 2026-09-18 补充：差分、正交化与类别编码
+
+### 同源做差不会保证降低 PC
+
+[同源信号与噪声对冲](https://support.worldquantbrain.com/hc/en-us/community/posts/43501264459799)（2026-09-15）比较两条弱信号做差和两条强信号做差；[后续讨论](https://support.worldquantbrain.com/hc/en-us/community/posts/43536370604567)（2026-09-16）要求拆开两腿验证，但其中“做差一定会让 PC 下降”的说法不成立。数学反例：令 P 与 N 独立且方差相同，A=P+N、B=N，则 A 与 P 的相关系数为 1/√2，而 A−B 与 P 的相关系数为 1。对某些生产池成分，相关性反而会上升。
+
+保留的实验是 A、B、A−B 分别比较收益、成本、覆盖和与已有池的相关性；两腿都弱也不自动证明共享噪声被成功剥离，需要样本外和稳定性证据。[周频量价概率正交化案例](https://support.worldquantbrain.com/hc/en-us/community/posts/43500671638551)（2026-09-15）自报 Max Prod Corr 在加入运营效率控制后从 0.6806 升至 0.7068，再加国家中性化降至 0.6681。与指定控制变量正交，不等于与整个生产池低相关；作者还报告覆盖及收益损失，应把这些变化同时列入对照，不将单个检查过线称为新机制成立。
+
+### 分析师差分先对齐经济口径
+
+[analyst10 预期差异帖](https://support.worldquantbrain.com/hc/en-us/community/posts/43524498271895)（2026-09-16）提出近远期预期差、上调与下调差等候选。可取的是同单位、同口径、只改变一个经济维度的配对方式；覆盖人数与预测 surprise 不能直接相减，已表示净修订的字段也不能未经语义核对再机械差分。120 日回填是作者参数，须另外检查数据可得时点、陈旧观测及两腿实际覆盖，不当作字段的通用配置。
+
+### 类别编号不是有量纲的数值
+
+[JPN 另类行业分类协方差帖](https://support.worldquantbrain.com/hc/en-us/community/posts/43308165478935)（2026-09-07，本轮补录）把两套类别整数标签做时间协方差，报告低换手，同时承认解释偏弱和 OS 检查 PENDING。任意置换簇编号可能改变该协方差，而经济分类没有改变，因此不能直接将结果解释为“分类分歧强度”。至少应检验标签重编码不变性，并与基于类别成员关系的一致性度量比较；不同聚类方法也不自动意味着统计独立。本轮没有复现其回测。

docs/worldquant-knowledge-base/02-ai-mining/python-alpha.md

@@ -31,3 +31,11 @@ FastExpr 已能表达的简单机制，先用它建立基线。需要自定义
 来源：[[Python Alpha 挑战赛]Python Alpha两阶段筛选法：FastExpr快速验证 + Python深度优化](https://support.worldquantbrain.com/hc/en-us/community/posts/41135490599063)（发帖 2026-06-11，社区经验）；[从 Fast 翻译到 Python Alpha 别急着炫技：截面/时序错位是头号杀手，事件漂移和中性化层级是隐形雷](https://support.worldquantbrain.com/hc/en-us/community/posts/42047259261335)（发帖 2026-07-18，社区经验）；[【经验分享】Python Alpha 30名心得：从 FastExpr 到 Python，我用 17 个 alpha 趟出来的一条路](https://support.worldquantbrain.com/hc/en-us/community/posts/42519002511767)（发帖 2026-08-06，社区经验）；[Python Alpha 平台限制粗略小结，请大佬指正](https://support.worldquantbrain.com/hc/en-us/community/posts/40742727142679)（发帖 2026-05-26，社区经验）。
 
 本次没有安装或运行 BRAIN Labs Python SDK，所以这里只提供官方入口与迁移验证设计，不提供宣称可直接运行的 SDK 脚本。
+
+## 2026-09-25 补充：频域与 PCA 实现先验算
+
+[FFT 复现纠错帖](https://support.worldquantbrain.com/hc/en-us/community/posts/43640747372695)（2026-09-21）指出一个可直接验算的恒等式：对同一输入 x 的互补频带，slow + fast = x，因此 slow − x 与 −fast 是同一条腿，混合权重不会产生新信息。这只适用于所示互补分解及一致预处理，不能泛化为所有慢快信号必然共线。
+
+原帖的其余“修正”仍需审查。对称 Hann 窗末端为零只能推出重构后的 slow[-1] + fast[-1] = 0，不能单凭这一点断言两个分量各自为零或必然衰减固定倍数。正文提倡前值填充，但修正版实际用窗口列均值填缺失；二者不是同一种经济假设。代码 `hist[-WIN:]` 在 WIN=65 时取 65 行，与注释宣称的 66 行也不一致。14–17 倍加速、泄漏比例和回测效果均为作者报告，本库未运行代码。
+
+[PCA 特质残差反转帖](https://support.worldquantbrain.com/hc/en-us/community/posts/43692787685015)（2026-09-23）可作为待验证的自估风险模型思路，但代码 `cs[H-1:] - cs[:-H+1]` 每项相减的累计索引间隔为 H−1，不能直接标成 H 日收益和；它还将非有限收益填零，需核对缺失、上市历史与股票池影响。作者报告 IND 有效而 EUR/USA/HKG 不佳，并将 PnL 相关 0.963 称作“恒等变换”；这是近似经验，非数学恒等。不得照搬其地区结论或部分正交化系数。

docs/worldquant-knowledge-base/03-engineering/2026-operating-lessons.md

@@ -41,3 +41,19 @@ CSV 本身不保证并发原子性。需要记录任务身份、租约或领取
 可迁移的工程做法是先清点候选、小样本检查，再扩大批量；逐项记录源 Alpha、模拟引用和终态。`POST_ERROR` 不能一律视为已处理，未知受理结果须先对账，429 等临时错误与字段不支持多区域的兼容性错误应分别处理。评论补充保持原有 decay、pasteurization 等设置，避免把参数变化误归因于跨区域迁移；新的 universe 口径也要单独核对。
 
 帖子开头声称批量重跑“一定能找到”可提交候选，但正文也明确完成不代表指标或提交资格保留。本库只保留后一边界；没有可提交结果也不足以单独证明原信号质量低，仍需排查覆盖、设置和区域适用性。
+
+## 2026-09-18 补充：断点声明与代码行为要对照
+
+[OS Alpha 批量重跑 RA 两脚本](https://support.worldquantbrain.com/hc/en-us/community/posts/43524503109655)（2026-09-16）声称支持续跑，但所附代码在 `simulate_single` 返回后无条件 `mark_done(k)`，而异常或缺少 Location 的路径可能返回 `None`。这会把未确认完成的任务记为已处理；应保留已受理引用及 UNKNOWN/待对账状态，不能仅靠 done 集合跳过。进度键也没有覆盖全部 settings，改动未入键参数后可能误复用旧进度。这些是静态阅读发现，没有运行脚本或验证平台端行为。
+
+同帖还包含自动 PATCH Alpha 标签的步骤，因而不是纯读取工具。[另一篇 RA 说明](https://support.worldquantbrain.com/hc/en-us/community/posts/43410200668823)（2026-09-11）所述“不设置属性、不提交”不能自动套用到不同脚本。[AlphaSubmitter](https://support.worldquantbrain.com/hc/en-us/community/posts/43299585796247)（2026-09-07，本轮补录）宣传多轮自动重试，但没有给出足够的幂等或受理状态核对证据；POST 超时不能直接解释为服务端未创建，重发前需先对账。
+
+[意见留言墙](https://support.worldquantbrain.com/hc/en-us/community/posts/42302637812503)本轮新可见评论提出登录过期导致界面状态丢失、回测次数与 Alpha 列表数量不同等问题。它们是用户反馈，不是已确认平台故障：先保存任务引用与界面筛选状态，再分别核对请求次数、失败/取消状态和结果对象数，避免仅凭两个总数推断丢数据。本轮没有复现相关 UI 行为。
+
+## 2026-09-25 补充：缺失状态不能被默认值掩盖
+
+[回测用量帖的新增可见评论](https://support.worldquantbrain.com/hc/en-us/community/posts/43058251853847#community_comment_43553064668183)（2026-09-17，本轮补齐）指出，没有匹配到当天活动记录时直接置 used=0，可能把日期错位或接口异常解释成额度充足。适合作为审查项：缺记录保留 UNKNOWN，再核对日期与响应；评论所述日界、固定上限和具体 API 行为未由本轮核验。
+
+[ProdMemo 桥接帖评论](https://support.worldquantbrain.com/hc/en-us/community/posts/42585946296855#community_comment_43606520549399)（2026-09-19）提出旧缓存覆盖新记录的风险。合并时应保留数据类型、采集时间与来源范围，并核对部分结果是否覆盖完整结果；这是评论提供的待核审查点，本轮没有重新读取并运行桥接脚本，不能直接宣称已确认其实现缺陷。
+
+[多 Agent 研究流程](https://support.worldquantbrain.com/hc/en-us/community/posts/43623479962007)（2026-09-20）把字段证据、研究假设、实验结果与审计裁定分开，是可用的职责设计：字段存在不证明逻辑成立，回测返回结果不证明通过审计。它是作者的架构方案，不是多代理提高样本外表现的实证。

docs/worldquant-knowledge-base/04-platform/current-updates.md

@@ -28,3 +28,9 @@
 [Grand Master 1v1 Advisory 通知](https://support.worldquantbrain.com/hc/en-us/community/posts/43320365312023)（2026-09-08 发布，0 条评论）介绍约 10 分钟答疑，称试运行预约窗口截至 9 月 17 日。可准备一个具体研究瓶颈和证据；本次没有访问预约服务、核对剩余名额或预约。
 
 [关于第三方发布的论坛公告](https://support.worldquantbrain.com/hc/en-us/community/posts/43319649278359)（2026-09-08 发布，本次读正文及 2 条评论）强调平台内容的对外发布限制，涉及 Alpha、数据、文档及论坛文章。本库记录公告及来源，不作协议效力或个案法律判断；本轮仅本地整理，没有对外发布、推送仓库或改变权限。
+
+## 2026-09-18 论坛增量
+
+[ARC2026 社区积分活动通知](https://support.worldquantbrain.com/hc/en-us/community/posts/43498120902807)（2026-09-15）称，CN/HK 顾问按比赛结束时官方榜单统计，成功提交至少 2 个 ARC/RA Parent 可得 50 积分，至少 10 个可得 100 积分，取最高档、不叠加。通知明确这是社区配套活动，与比赛的美元奖池分开。本轮读取通知，未核实个人资格、积分到账或兑换比例，也未提交 Alpha。
+
+上文 Grand Master 1v1 Advisory 原公布的 9 月 17 日预约窗口已过；截至本次整理，不应继续把它列为已确认开放的预约机会，本轮未核查是否续期。新出现的招聘、免费模型与云服务推广帖保留目录，不将促销额度、模型可用性和邀请收益写成研究结论。

docs/worldquant-knowledge-base/04-platform/portfolio-osmosis.md

@@ -26,3 +26,9 @@
 先排查数据中断与不可投资表现，再按同族/相关簇留代表，然后看同时回撤和成本后的组合增量。旧 Alpha 也可重新评估；但不能只因最近几天排名好就重仓。
 
 为每次比较固定评价时段和费用设置，记录“原组合”“加入候选”“替换近似旧项”三组结果。正式配点/提交前再确认资格与生效时间。本次没有修改任何平台属性。
+
+## 2026-09-18 补充：教学倍数与支付公式分开
+
+[Theme 到 Osmosis 的说明帖](https://support.worldquantbrain.com/hc/en-us/community/posts/43529198653719)（2026-09-16）区分质量因子的加成与每日美元报酬，并明确把“主题合并后再乘 Osmosis”“不分配按 1 倍”作为文章的补充计算口径；这不等于取得完整官方支付公式。[Base Payment 说明帖](https://support.worldquantbrain.com/hc/en-us/community/posts/43561782887575)（2026-09-17）同样声明，四项得分归一化到 0–1 再等权平均只是教学假设。“交满适用配额即 Quantity 满分”也是该帖补充口径，本轮没有独立核实，不能据此计算个人收入或设置自动提交目标。
+
+[两条高倍数 Alpha 的收入分享](https://support.worldquantbrain.com/hc/en-us/community/posts/43532701388183)（2026-09-16）把两天支付差异归因于提交组合及平均方式；缺少同条件对照，不能据此推导“少交必然更赚钱”或固定的 VF、Osmosis 收入门槛。保留的研究问题是：完整 Alpha 的适用资格、组合贡献和评价窗口是否一致。本文不新增已核实的乘数、资格或支付规则，也不把作者报告视为本次复现。

docs/worldquant-knowledge-base/04-platform/spc.md

@@ -4,7 +4,7 @@ SPC 与“LLM 帮你写 FastExpr”不同：你提交投资论点 Prompt，平
 
 ## 9 月必须更新的输出格式
 
-7 月中文教程仍写 ISIN|MIC 字典；最新官方提交页（9 月 6 日更新）要求**对象数组**。结构示意如下，尖括号为类型占位符，不能原样提交：
+7 月中文教程仍写 ISIN|MIC 字典；9 月 7 日首次核查的官方提交页已要求**对象数组**。2026-09-18 再读该页（页面更新 9 月 13 日），要求仍为对象数组。结构示意如下，尖括号为类型占位符，不能原样提交：
 
 ```text
 [
@@ -20,9 +20,9 @@ SPC 与“LLM 帮你写 FastExpr”不同：你提交投资论点 Prompt，平
 ]
 ```
 
-confidence_score 范围 -1 到 1，符号对应多空；investment_horizon 为预期实现收益的天数；return_prediction 表示预测收益率。官方示例用 0.045 表示 4.5%。务必核对证券标识与交易场所，不能用臆造 ID。
+confidence_score 范围 -1 到 1，符号对应多空；investment_horizon 为预期实现收益的天数；return_prediction 表示预期收益变动幅度。9 月 18 日复核的官方文字要求 return_prediction 为正，多空方向由 confidence_score 表示，不能再靠负 return_prediction 表示做空。官方示例用 0.045 表示 4.5%。务必核对证券标识与交易场所，不能用臆造 ID。
 
-来源：[What do I need to submit?](https://support.worldquantbrain.com/hc/en-us/articles/41681208900759)（官方，更新 2026-09-06）；旧格式来源：[【SPC 2026】一文看懂 Systematic Prediction Challenge，写提示词就能参加！](https://support.worldquantbrain.com/hc/en-us/community/posts/41801462727191)（发帖 2026-07-08，社区经验）。
+来源：[What do I need to submit?](https://support.worldquantbrain.com/hc/en-us/articles/41681208900759)（官方，本次页面更新 2026-09-13，读取 2026-09-18）；旧格式来源：[【SPC 2026】一文看懂 Systematic Prediction Challenge，写提示词就能参加！](https://support.worldquantbrain.com/hc/en-us/community/posts/41801462727191)（发帖 2026-07-08，社区经验）。
 
 ## 已核实的规则
 
@@ -46,3 +46,9 @@ confidence_score 范围 -1 到 1，符号对应多空；investment_horizon 为
 社区“信号 5–7 天衰减”来自特定经验，不是所有 SPC 机制的固定半衰期。期限应根据来源更新节奏与前瞻验证决定。来源：[【SPC实战】信号衰减5-7天：数据源“时效性”比“全面性”更重要](https://support.worldquantbrain.com/hc/en-us/community/posts/42955680180247)（发帖 2026-08-24，社区经验）。
 
 本次未调用 LLM 生成投资组合，未提交 Prompt、调整权重或参加比赛。
+
+## 2026-09-18 补充：社区提交脚本仍需核对版本
+
+[SPC 提交 API 抓包与自动化帖](https://support.worldquantbrain.com/hc/en-us/community/posts/43322480595991)（2026-09-08，本轮补录）示例仍将 `sampleOutput` 写为 `ISIN|MIC` 字典字符串，与本次官方要求的输出内容不一致。API 包装字段是否需要字符串编码和 Prompt 实际应输出什么，是两个问题；不能因抓包样例曾被接受，就将旧字典视为现行内容规范。
+
+静态阅读还发现，该帖称已加入调权命令，但展示的命令解析只有登录检查、列表和提交；权重验证检查范围却未落实文中两位小数的声明。示例 POST 超时后没有受理状态对账，人工或自动重试可能造成重复提交。`weight=0` 在官方文字中是停止运行，不应按社区的“下线”表述推断对象被删除。这里只记录版本与恢复边界，未执行脚本、探测提交接口或调整权重。

docs/worldquant-knowledge-base/05-research-practice/validation.md

@@ -18,3 +18,9 @@
 大规模搜索时特别检查：收益是否集中在某一年；低 PC 是否只来自缺失或不同日期；高 Sharpe 是否由不可投资的极端持仓产生；同一模板换字段是否仍是同一风险来源；回填是否把旧事件变成持续仓位。
 
 本知识库中的研究方向没有经过上述实际回测终验，均是待研究候选。
+
+## 2026-09-25 补充：冻结发生在看结果之前
+
+[时间前推与冻结流程](https://support.worldquantbrain.com/hc/en-us/community/posts/43638422558103)（2026-09-21，社区方法建议）区分历史分年诊断、时间前推与前瞻观察。看过完整历史后再按年拆分，不能恢复被选择过程用掉的独立性。每个验证段开始前记录表达式、完整 settings、冻结时间和淘汰条件；改动生成新版本，保留失败段，汇总冻结后各段的日收益，而非只展示最好一段或平均各段 Sharpe。
+
+若标签或持有期跨越分割点，隔离范围应依据实际重叠，而非机械等于最长回看窗口。冻结卡和影子观察是研究建议，不是平台提交规则，也不能据此把个人留出段称为平台隐藏 OS。本轮未复核文中外部论文或运行验证实验。

docs/worldquant-knowledge-base/06-sources/README.md

@@ -1,6 +1,6 @@
 # 来源、覆盖范围与验证记录
 
-首次采集日期：2026-09-07；最近论坛同步：2026-09-11。下列初轮与扩展轮数字保留为历史记录，最新覆盖见文末周同步记录。时间边界按 Asia/Shanghai 的 2026-01-01 00:00 至整理时点；原始索引时间使用 UTC/带偏移的 ISO 8601。日历日简写通常取原来源 UTC 日期，未把它伪装成本地精确时刻。
+首次采集日期：2026-09-07；最近论坛同步：2026-09-25（有访问缺口）。下列初轮与扩展轮数字保留为历史记录，最新覆盖见文末周同步记录。时间边界按 Asia/Shanghai 的 2026-01-01 00:00 至整理时点；原始索引时间使用 UTC/带偏移的 ISO 8601。日历日简写通常取原来源 UTC 日期，未把它伪装成本地精确时刻。
 
 ## 邮箱
 
@@ -92,3 +92,33 @@ SPC 7 篇官方帮助页按 ID 直接读取；另通过支持站搜索取得 Pyt
 任务专用采集入口和状态位于本机 `.local/forum-sync/`（Git 忽略）：`collect.py` 负责分页、限流退避、互斥和逐帖原子保存；`staged/` 保留去除 HTML 和凭据类字符串的文本；`candidate-baseline.json` 为采集候选，知识整理及验证通过后才复制为 `baseline.json` 并推进 `checkpoint.json`。后续采集比较该基线的正文及评论指纹，仍须实际筛选变化内容。`run.json` 保存本次计数和失败项；部分失败不得宣称全部成功。此状态只包含任务必要内容，不保存登录响应、Cookie 或会话令牌，也未推广为全局规则或共享脚本。
 
 本轮验证：两个 CSV 各 629 个唯一帖子 ID，评论数合计 9,587；所有论坛引用均属于本次范围，122 篇帖子被知识条目引用，本地链接未发现缺失目标。新增知识内容及本地 JSON 状态的账号、密码字面值与 JWT 形式扫描通过。实际 API 采集通过；采集脚本语法检查通过，后续增量比较、失败恢复和限流退避尚未通过真实故障场景演练，不能将首次成功等同长期运行保证。
+
+## 每周论坛同步：2026-09-18
+
+本次实际采集时间为北京时间 10:11:39–10:16:36；这是执行记录，不将计划的 09:00 写成实际开始时间。认证返回 201，支持站 SSO 落地仍为 403，Community API 正常。18 页共有 1,771 篇可见帖子，排除 1,101 篇早于年初边界的帖子，范围内保留 670 篇；上次 629 个 ID 均仍可见。逐帖分页取得 9,601 条评论，670 个记录的实际取得数与列表计数均相等，零采集失败。
+
+相对 9 月 11 日内容基线，新收录 41 篇：29 篇发表于上次采集结束后，12 篇的创建时间更早。后者只能称为本轮补录，无法仅凭列表差异判断此前为何不可见或漏收。这 41 篇共取得 5 条评论。旧帖留言墙 42302637812503 新可见 9 个评论 ID，其中 7 条创建时间早于上次采集、2 条发表于 9 月 12 日；其余既有评论文本指纹未见变化，也没有缺失评论 ID。评论总数增加 14 条，其中只有 7 条的创建时间晚于上次采集，不能将净增量全称为本周发表。
+
+231 篇旧帖的更新时间变化，但只有两篇正文内容指纹变化：43320471694999 的 AI 合集新增 CNHKMCP QuantFlow 提示词入口，43030071187479 的新人指南新增第三方发布公告与 Osmosis 入口。逐一对照上周文本确认均为目录更新，没有据此扩充列表外旧帖知识。加上留言墙评论变化，共 3 篇既有帖出现内容或评论集合变化，其余仅同步元数据。
+
+41 篇新收录帖已完成标题筛查，40 篇有可提取正文并完成主旨阅读，14 篇研究机制和 6 篇工程材料另做独立核验。43287990820375 的正文仅含 12 张图片、没有可提取文字；尝试读取全部 12 个官方图片地址均返回 403，因此未逐图核验，阅读状态记为 `metadata_only`，并在检查点保留待阅读项，不把标题中的夺冠经验或 21 个 OS Alpha 视为已读证据。新收录帖的 5 条评论及留言墙新可见的 9 条评论均已阅读。宣传、充值/邀请和缺少研究证据的收入展示保留索引，不机械扩写成研究方法。
+
+知识更新包括：同源差分和正交化不保证降低生产相关性、类别编号协方差的编码依赖、分析师差分口径、代理变量及稀疏事件的有效样本边界；RA 脚本未知结果被标为完成的静态风险；支付教学假设与真实公式的区别；ARC 社区积分活动；按评价时间复盘的边界。所有社区回测数值仍为作者自报，本轮未运行或安装其代码。另通过官方 API 复核 SPC 提交说明 41681208900759（更新 2026-09-13），确认对象数组格式，并补充 return_prediction 正值与 confidence_score 多空方向的分工；没有访问提交接口或调整权重。
+
+本轮继续比较全部范围内评论的文本指纹，没有假定帖子更新时间或评论数量能完全反映评论编辑。评论指纹针对去除 HTML、规范化空白后的文字，不覆盖纯排版、链接目标或图片替换；因此“未见评论文本编辑”不能解释为所有媒体与格式均未变化。上周文本保留在本机 `.local/forum-sync/previous-2026-09-11/`，新候选仍由验证后才推进到相应成功基线，图片访问缺口另行保留；状态与采集文本不包含认证响应或会话材料。没有同步邮件、改变平台属性、预约、报名、回测、发帖或推送仓库。
+
+验证结果：两个 CSV 各有 670 个唯一帖子 ID，评论合计 9,601；采集记录及候选指纹逐项一致，142 篇帖子被知识条目引用，论坛引用均在范围内，本地链接与差异空白检查通过。新增知识内容和本地 JSON/JSONL 状态的账号密码字面值及 JWT 形式扫描通过。实际 API 采集及本周内容比对已运行；社区回测与代码未执行，自动失败恢复、限流退避仍未经过故障演练。文本及索引同步完成，图片正文访问失败保留为独立缺口。
+
+## 每周论坛同步：2026-09-25
+
+实际采集时间为北京时间 09:12:33–09:16:29，随后完成筛读和知识整理。认证返回 201，支持站 SSO 落地 403，Community API 正常。完整取得 19 页、1,859 篇可见帖子，排除 1,101 篇早于年初边界的帖子；本次范围内实际取得 758 篇、10,141 条评论，逐帖分页计数一致，无评论抓取失败。
+
+相对 9 月 18 日内容基线，新收录 89 篇，其中 72 篇发表于上次采集结束后，17 篇为历史补录。新收录帖带来 9 条评论，217 篇旧帖另有 540 个新可见评论 ID；这 549 条中，67 条发表于上次采集后、482 条创建时间更早，是本轮历史覆盖补齐。旧帖元数据更新时间变化 329 篇，但既有正文指纹无变化，可见帖既有评论没有文本编辑或缺失 ID；没有仅凭更新时间或评论数判断未变。
+
+此前收录的 41208709211671（傅里叶变换在 Alpha 开发中的应用，2026-06-14）本次列表缺失，直接官方 API 请求也返回 404，原因未定。保留原文快照、引用、索引及其历史 9 条评论，不把 404 解释为已删除。因而当前索引有 759 个 ID：758 个本次可见，1 个历史保留；阅读索引评论数合计 10,150，其中 9 条未在本次重新取得，该帖 comments_complete 记为 False。本次实际取得评论比上次增加 540，等于新可见 549 减去未能重新取得的历史 9，不能将净增量称作新发表评论。
+
+89 篇新收录帖完成标题及正文主旨筛查，专题目录只作导航，未沿目录扩充 2026 年前帖子；福利、收入展示和未答问题保留索引，不扩写为结论。新收录帖的 9 条评论及旧帖新可见的 540 条评论完成筛读，并针对实质纠错和机制建议精读。新增知识集中在冻结后验证、FFT 互补频带恒等式及修正版局限、PCA 窗口索引、聚类代表不能替代逐项 PC 终验、缺失值语义与 UNKNOWN 状态。社区数值仍为作者自述；本次没有复现回测、安装或执行社区代码，也没有将评论中的配额、评分、时区和阈值升级为官方规则。
+
+图片帖 43287990820375 仍可取得元数据，正文含 12 张图片；本次抽查其中 1 个官方图片地址仍为 403，没有把抽查写成全部重试，继续保留 metadata_only 和待阅读项。评论指纹仍仅覆盖规范化文字，不检测纯格式、链接目标或图片替换。本次保留上周 670 篇文本快照于本机 previous-2026-09-18/；可见记录验证后推进相应基线，不可访问帖保留旧指纹与重试缺口。此次为可见文本及索引完成、来源与媒体仍有缺口的部分同步，不宣称全量无缺失。
+
+验证结果：两个 CSV 均有 759 个唯一 ID；758 个本次采集记录与候选指纹、逐评论文本哈希及分页计数一致，另 1 个历史保留记录明确标为当前未完成。154 篇帖子被知识条目引用，引用均属于本次可见范围或明确保留的历史来源。本地链接、差异空白、新增内容及本地 JSON/JSONL 的账号密码字面值和 JWT 形式扫描通过。采集脚本语法检查通过；本轮只验证实际采集与静态内容比对，没有故障注入、策略回测或社区代码执行。

docs/worldquant-knowledge-base/06-sources/collection-2026.json

@@ -29,6 +29,299 @@
     "No source authentication material is included."
   ],
   "latest_forum_sync": {
+    "collected_at": "2026-09-25",
+    "timezone": "Asia/Shanghai",
+    "status": "partial_sync_with_source_and_media_gaps",
+    "started_at": "2026-09-25T01:12:33.054002+00:00",
+    "collection_finished_at": "2026-09-25T01:16:29.789990+00:00",
+    "previous_successful_collection_finished_at": "2026-09-18T02:16:36.188053+00:00",
+    "raw_visible_posts": 1859,
+    "raw_list_pages": 19,
+    "excluded_older_posts": 1101,
+    "retained_posts": 758,
+    "retained_comments": 10141,
+    "indexed_posts_including_unavailable": 759,
+    "historical_comments_retained_for_unavailable_posts": 9,
+    "newly_indexed_posts": 89,
+    "newly_indexed_posts_created_after_previous_collection": 72,
+    "historical_posts_newly_indexed": 17,
+    "comments_on_newly_indexed_posts": 9,
+    "newly_visible_comments_on_existing_posts": 540,
+    "historical_comments_newly_visible": 482,
+    "newly_visible_comments_created_after_previous_collection": 67,
+    "comment_count_net_change": 540,
+    "metadata_changed_existing_posts": 329,
+    "body_changed_existing_post_ids": [],
+    "comment_changed_existing_post_ids": [
+      "37427295710359",
+      "37473718017175",
+      "37759477785239",
+      "37857879534487",
+      "37858237686167",
+      "38120459173143",
+      "38151220803095",
+      "38669179039255",
+      "38767793594775",
+      "39118594803095",
+      "39183798681239",
+      "39215982466071",
+      "39225718749591",
+      "39249078858647",
+      "39304762113815",
+      "39319955780887",
+      "39367839107095",
+      "39567420541079",
+      "39729140385559",
+      "39756841978519",
+      "39845230299927",
+      "39873247817495",
+      "39922788056343",
+      "40031467208983",
+      "40089426663959",
+      "40287368788375",
+      "40373405530647",
+      "40544447671959",
+      "40734437350679",
+      "40734489248151",
+      "40737593669911",
+      "40790666727575",
+      "40847481080087",
+      "40936186021271",
+      "40962462282007",
+      "40964044763031",
+      "41020681850519",
+      "41065497021335",
+      "41138382717719",
+      "41162636720919",
+      "41236357050903",
+      "41318082888343",
+      "41320039211671",
+      "41363895396247",
+      "41391880122007",
+      "41392995550231",
+      "41396933582231",
+      "41397704497303",
+      "41417418697111",
+      "41418196103191",
+      "41420742165527",
+      "41459028441623",
+      "41510344038551",
+      "41540700959639",
+      "41550783184663",
+      "41551045362071",
+      "41551682155799",
+      "41552805539991",
+      "41553665016983",
+      "41556669366551",
+      "41567966789655",
+      "41625263438999",
+      "41651283176855",
+      "41654205318807",
+      "41655387004439",
+      "41684104419095",
+      "41692781472919",
+      "41706827651991",
+      "41719890581911",
+      "41732813691287",
+      "41801462727191",
+      "41804720051607",
+      "41825836449047",
+      "41849925214231",
+      "41854026316055",
+      "41856507417623",
+      "41877348653335",
+      "41877985926423",
+      "41878175264663",
+      "41880799695127",
+      "41881412101399",
+      "41897133795479",
+      "41898002651287",
+      "41898564566167",
+      "41903649174807",
+      "41908323673111",
+      "41909117583383",
+      "41937689411479",
+      "41939959839895",
+      "41949669579031",
+      "41974186627351",
+      "41992899715607",
+      "41993968062743",
+      "41994809192215",
+      "41995369186455",
+      "41995684431767",
+      "42007074836247",
+      "42015743397655",
+      "42018671130391",
+      "42019010387479",
+      "42019332809111",
+      "42019924663319",
+      "42020342676759",
+      "42021244122775",
+      "42021722069143",
+      "42022555985047",
+      "42029118285847",
+      "42045578950551",
+      "42047259261335",
+      "42060935441943",
+      "42061123738903",
+      "42062231758103",
+      "42064544905495",
+      "42064574062359",
+      "42065262268567",
+      "42065802192023",
+      "42078729821591",
+      "42078917602455",
+      "42160836246039",
+      "42160945573271",
+      "42161798745879",
+      "42184733027991",
+      "42185098677399",
+      "42187744015767",
+      "42187906212247",
+      "42211897526295",
+      "42215210781719",
+      "42215740602135",
+      "42216209199255",
+      "42259186135831",
+      "42274443124119",
+      "42302637812503",
+      "42359440646551",
+      "42406158561943",
+      "42407253630359",
+      "42423399281815",
+      "42423993843735",
+      "42455010729367",
+      "42460800447895",
+      "42487308197783",
+      "42509573118999",
+      "42513455206423",
+      "42513832186263",
+      "42519002511767",
+      "42549574041367",
+      "42585946296855",
+      "42587179348119",
+      "42662873714711",
+      "42686571665303",
+      "42691768951447",
+      "42714622464407",
+      "42720875852439",
+      "42721824181271",
+      "42738237663639",
+      "42740521405463",
+      "42743732814999",
+      "42756136847639",
+      "42756641286295",
+      "42757091114647",
+      "42771962021143",
+      "42772611660823",
+      "42774450536343",
+      "42775887653399",
+      "42801904467607",
+      "42945861800983",
+      "42955680180247",
+      "42972530812183",
+      "42975079714839",
+      "42981290349463",
+      "43002825233687",
+      "43006316208151",
+      "43030071187479",
+      "43031425132823",
+      "43031577079575",
+      "43058251853847",
+      "43067045830679",
+      "43085454784535",
+      "43087031637527",
+      "43147371557015",
+      "43151729306647",
+      "43178926004887",
+      "43179521660823",
+      "43179860065303",
+      "43181411147287",
+      "43208281917591",
+      "43269155035799",
+      "43287990820375",
+      "43299585796247",
+      "43309848501783",
+      "43319649278359",
+      "43319751480599",
+      "43320365312023",
+      "43320471694999",
+      "43322521365143",
+      "43322749748503",
+      "43383826751127",
+      "43410200668823",
+      "43411841446423",
+      "43438866840343",
+      "43467733543319",
+      "43477475833623",
+      "43498120902807",
+      "43499065604247",
+      "43500671638551",
+      "43501264459799",
+      "43502970330647",
+      "43508107778711",
+      "43524498271895",
+      "43524503109655",
+      "43527309427863",
+      "43527457608087",
+      "43529198653719",
+      "43529216017047",
+      "43536370604567",
+      "43541109631895",
+      "43556223197335",
+      "43561782887575"
+    ],
+    "edited_existing_comments": 0,
+    "missing_existing_comment_ids_on_visible_posts": 0,
+    "failed_comment_records": 0,
+    "missing_previously_indexed_posts": 1,
+    "unavailable_posts": [
+      {
+        "post_id": "41208709211671",
+        "direct_api_status": 404,
+        "historical_comments_retained": 9
+      }
+    ],
+    "posts_cited_in_knowledge": 154,
+    "groups": {
+      "guides": 17,
+      "ai": 160,
+      "ideas": 40,
+      "engineering": 86,
+      "general": 137,
+      "lessons": 141,
+      "optimization": 65,
+      "events": 42,
+      "portfolio": 51,
+      "data": 20
+    },
+    "new_post_texts_screened": 89,
+    "metadata_only_post_ids": [
+      "43287990820375"
+    ],
+    "new_or_newly_visible_comments_screened": 549,
+    "official_articles_checked": [],
+    "baseline_state_directory": ".local/forum-sync/",
+    "unread_media": [
+      {
+        "post_id": "43287990820375",
+        "image_count": 12,
+        "images_retried_this_run": 1,
+        "http_status": 403
+      }
+    ],
+    "limits": [
+      "758 currently visible posts and 10141 freshly retrieved comments; indexes additionally retain one unavailable post with nine historical comments.",
+      "One previously indexed post returned HTTP 404; its previous fingerprint and records are retained for retry, not deleted or marked freshly complete.",
+      "Newly visible historical posts/comments are backfill; the reason for prior visibility gaps is unknown.",
+      "One image-only post remains unread; one of its twelve official images was retried and returned HTTP 403.",
+      "Comment fingerprints cover normalized text only, not link targets, formatting or media-only changes.",
+      "Community claims and code were reviewed but not executed or backtested; no current platform rules were inferred from comments.",
+      "No email sync, simulation, external publication, submission, credential/session persistence or Git commit/push."
+    ]
+  },
+  "forum_sync_history": [
+    {
       "collected_at": "2026-09-11",
       "timezone": "Asia/Shanghai",
       "started_at": "2026-09-11T01:03:40.336597+00:00",
@@ -77,5 +370,78 @@
         "All scoped comments fetched, selective historical body reading; images and attachments not fully verified.",
         "No simulations, submissions, email sync, public publishing, or credential/session persistence."
       ]
+    },
+    {
+      "collected_at": "2026-09-18",
+      "timezone": "Asia/Shanghai",
+      "started_at": "2026-09-18T02:11:39.967012+00:00",
+      "collection_finished_at": "2026-09-18T02:16:36.188053+00:00",
+      "previous_successful_collection_finished_at": "2026-09-11T01:06:33.369594+00:00",
+      "raw_visible_posts": 1771,
+      "raw_list_pages": 18,
+      "excluded_older_posts": 1101,
+      "retained_posts": 670,
+      "retained_comments": 9601,
+      "newly_indexed_posts": 41,
+      "newly_indexed_posts_created_after_previous_collection": 29,
+      "historical_posts_newly_indexed": 12,
+      "comments_on_newly_indexed_posts": 5,
+      "newly_visible_comments_on_existing_posts": 9,
+      "historical_comments_newly_visible": 7,
+      "newly_visible_comments_created_after_previous_collection": 7,
+      "comment_count_net_change": 14,
+      "metadata_changed_existing_posts": 231,
+      "body_changed_existing_post_ids": [
+        "43320471694999",
+        "43030071187479"
+      ],
+      "comment_changed_existing_post_ids": [
+        "42302637812503"
+      ],
+      "edited_existing_comments": 0,
+      "missing_existing_comment_ids": 0,
+      "failed_comment_records": 0,
+      "missing_previously_indexed_posts": 0,
+      "posts_cited_in_knowledge": 142,
+      "groups": {
+        "portfolio": 50,
+        "ai": 154,
+        "ideas": 30,
+        "lessons": 134,
+        "optimization": 62,
+        "data": 20,
+        "events": 41,
+        "engineering": 75,
+        "guides": 8,
+        "general": 96
+      },
+      "new_post_texts_screened": 40,
+      "metadata_only_post_ids": [
+        "43287990820375"
+      ],
+      "new_or_newly_visible_comments_read": 14,
+      "official_articles_checked": [
+        {
+          "id": "41681208900759",
+          "updated_at": "2026-09-13T15:38:21Z",
+          "read_at": "2026-09-18"
         }
+      ],
+      "baseline_state_directory": ".local/forum-sync/",
+      "limits": [
+        "Newly visible records can predate the previous run; cause of historical visibility gaps is unknown.",
+        "One new post is image-only and its body has not been visually reviewed; other images and attachments not fully verified.",
+        "Community backtest claims not reproduced; code findings are static, not live platform tests.",
+        "No simulations, submissions, email sync, public publishing, or credential/session persistence.",
+        "Comment fingerprints cover normalized text only, not link destinations, formatting, or embedded media changes."
+      ],
+      "unread_media": [
+        {
+          "post_id": "43287990820375",
+          "image_count": 12,
+          "http_status": 403
+        }
+      ]
+    }
+  ]
 }

docs/worldquant-knowledge-base/07-professional-knowledge/data-and-signal-semantics.md

@@ -45,3 +45,11 @@
 增加同类字段可能只是叠加相关噪声和缺失。帖子记录了同域克隆失败、跨类别组合更好的样本，但没有证明多类别组合必然有效。来源：[字段堆越多alpha越牛？我替你们试过了，真不是这么回事](https://support.worldquantbrain.com/hc/en-us/community/posts/41908323673111)（2026-07-13）。
 
 保持主字段、方向与覆盖语义后，才讨论删除冗余算子；降低 OpCount 不能靠偷偷改变信号。来源：[【Quant101】OpCount 拯救计划：因子化简](https://support.worldquantbrain.com/hc/en-us/community/posts/41105918774807)（2026-06-10）。详见 [优化流程](../08-alpha-optimization/diagnostic-workflow.md)。
+
+## 2026-09-25 补充：缺失设置是实验变量
+
+[按数据类型选择回填](https://support.worldquantbrain.com/hc/en-us/community/posts/43273417454487)（2026-09-05，本轮补录）和 [NanHandling 讨论](https://support.worldquantbrain.com/hc/en-us/community/posts/43562519639063)（2026-09-17，本轮补录）提醒，事件未发生、状态尚未更新与采集缺失不能用同一填法。NanHandling ON/OFF、算子及字段组合可能改变有效股票数和信号含义；文中观测不是平台所有算子的统一定义，尤其不能推广“其他情况无所谓”。
+
+[新闻密度帖的新增可见评论](https://support.worldquantbrain.com/hc/en-us/community/posts/43541109631895#community_comment_43582240044695)（评论发表于 2026-09-18）提出：若无新闻日为缺失且均值忽略缺失，得到的是有新闻日的条件均值，并非每个交易日的新闻密度。应先核对实际返回值，再比较零填充、原缺失和有限回填，同时观察覆盖与集中度；不能预先把缺失都当无事件。[预处理讨论评论](https://support.worldquantbrain.com/hc/en-us/community/posts/43527457608087#community_comment_43552980010775)（2026-09-17，本轮补齐）进一步提醒，回填后的 Sharpe 提升可能只是股票构成改变，需并看信号年龄、多空数量、换手及分组暴露。这些是评论者的机制建议，不是已确认因果结果。
+
+历史频域来源 41208709211671 在本次列表缺失，2026-09-25 直接 API 请求返回 404；保留此前归纳和来源，不据此判断作者删除或内容失效。

docs/worldquant-knowledge-base/08-alpha-optimization/portfolio-and-correlation.md

@@ -33,3 +33,9 @@
 Osmosis 社区帖称有多个 Region×Delay、每赛道 Alpha 数量与点数要求；具体资格和锁定时区未由本轮完整核实，不将该帖直接升级为当前官方规则。来源：[🌟 玩转 Osmosis：让你的 Alphas 成为最高 2 倍收益放大器！(机制详解与实操指南)](https://support.worldquantbrain.com/hc/en-us/community/posts/43002825233687)（2026-08-26）。
 
 关于低 PC 与高 Fitness 的取舍，社区有单变量实验支持按质量预算思考，不能只追最低 PC。来源：[关于降Prod Correlation的实践（一）](https://support.worldquantbrain.com/hc/en-us/community/posts/39118594803095)（2026-03-18）。具体权重和阈值需要依据你的池验证，本任务未修改组合、提交 Alpha 或分配点数。
+
+## 2026-09-25 补充：聚类代表不能替代每个候选的终验
+
+[本地 PnL 聚类预筛工具](https://support.worldquantbrain.com/hc/en-us/community/posts/43300341649687)（2026-09-07，本轮补录）按候选间相关性合并簇，只查代表项并把 PC 结果传播给成员。减少重复检查可作为调度思路，但 PASS/FAIL 传播没有充分保证。以同一日期样本上中心化、标准化的 A、B、P 为例，柯西不等式给出 |corr(A,P)−corr(B,P)| ≤ √(2(1−corr(A,B)))；即使 corr(A,B)=0.9，上界仍约 0.447，远不能保证跨不过提交门槛。这是本库数学边界，不是平台 PC 公式；链式连通聚类还不保证任意两成员都达到聚类阈值。
+
+[双桶去相关案例](https://support.worldquantbrain.com/hc/en-us/community/posts/43624302417559)（2026-09-20）报告 PC 下降同时 margin 下降，也给出分桶轴与信号重叠导致信号被抹去的反例。因此比较应同时记录成本、覆盖、子区域与生产相关性，不能只凭某项过线判定改善。[Value Factor 讨论](https://support.worldquantbrain.com/hc/en-us/community/posts/43589077012759)（2026-09-18）区分单体质量、自有池增量和平台全池评价；其教学示例不是实测收益或支付公式。上述数值和工具效果均未复现。

docs/worldquant-knowledge-base/09-experience-lessons/failure-casebook.md

@@ -32,3 +32,9 @@
 触发条件；完整基线；失败原值；已经试过的修复；对照；新的失败项；适用边界；是否复现；下次何时允许重试。把网络故障、语义错误、检查失败和经济假设被否定分成不同类别。
 
 本库是当前任务交付物，没有将这些社区经验自动写入全局 Skill、Hook 或硬性提交规则。
+
+## 2026-09-18 补充：复盘要对齐评价时间
+
+[OS Comb 六月至七月复盘](https://support.worldquantbrain.com/hc/en-us/community/posts/43438866840343)（2026-09-12）只讨论已更新的六月 3.3+、七月 3.5+，明确八月未刷新，不把后续研究工作解释为此前指标上升的原因。作者把变化归于跨区、多品类和入池质量；这是个人归因，未隔离市场环境、旧池变化和平台统计窗口。可保留的方法是给研究动作、提交入池日和指标评价期分别标记时间，避免倒置因果。
+
+[负 Combine 到 GM 的复盘上篇](https://support.worldquantbrain.com/hc/en-us/community/posts/43540397996311)及[下篇](https://support.worldquantbrain.com/hc/en-us/community/posts/43540418549527)（均为 2026-09-16）再次强调过提交检查不等于成本后质量、复杂度与多样性需要分开看。晋级与收入结果均为作者自报，不能将成功者经历称为保证可复制的方法，也不能把地区或数据类别数量直接当作有效分散度。

docs/worldquant-knowledge-base/README.md

@@ -1,8 +1,8 @@
 # WorldQuant 研究重启知识库
 
-首次整理：2026-09-07；论坛最近同步：2026-09-11，Asia/Shanghai。论坛严格按 **2026-01-01 起的发帖时间**筛选；旧帖即使今年更新、被邮件或合集引用，也不纳入论坛研究来源。面向个人研究，不包含已运行或保证有效的 Alpha。
+首次整理：2026-09-07；论坛最近同步：2026-09-25（有访问缺口），Asia/Shanghai。论坛严格按 **2026-01-01 起的发帖时间**筛选；旧帖即使今年更新、被邮件或合集引用，也不纳入论坛研究来源。面向个人研究，不包含已运行或保证有效的 Alpha。
 
-9 月 11 日增量：论坛索引现有 **629 篇帖子、9,587 条评论**，新增收录 6 篇（其中 1 篇为 9 月 5 日漏收来源），并补齐原有三篇帖子的 6 条历史评论。新增 ARC2026 官方 FAQ 复核、跨区域重跑经验及论坛公告，见 [平台更新](04-platform/current-updates.md) 和 [工程经验](03-engineering/2026-operating-lessons.md)。首次完整比较基线已建立；旧帖更新时间变化不能直接等同正文编辑，详见 [来源与覆盖](06-sources/README.md)。
+9 月 25 日增量：本次实际取得 **758 篇可见帖子、10,141 条评论**。新收录 89 篇，其中 72 篇发表于上次采集后、17 篇为历史补录；217 篇旧帖出现新可见评论，未见既有正文或评论文本编辑。更新[时间前推验证](05-research-practice/validation.md)、[FFT/PCA 实现边界](02-ai-mining/python-alpha.md)、[相关性预筛](08-alpha-optimization/portfolio-and-correlation.md)、[数据语义](07-professional-knowledge/data-and-signal-semantics.md)、[工程状态](03-engineering/2026-operating-lessons.md)和[事件信号对照](01-alpha-ideas/data-template-cards.md)。另有 1 篇旧帖返回 404，历史记录仍保留，索引共 759 篇；1 篇图片正文继续待读。详见[来源与覆盖](06-sources/README.md)。
 
 **先更新研究方向，再恢复回测规模。** 当前最值得投入的是有经济解释的低相关信号、能保留失败证据的 AI 实验闭环，以及组合层的增量价值。不要把“提交数量”“Token 消耗”当作研究进展。
 

frontend/src/research/CandidateTable.tsx

@@ -0,0 +1,93 @@
+import { useState } from "react";
+import { Pagination } from "@douyinfe/semi-ui-19";
+import { WorkspaceTable } from "../components/WorkspaceTable";
+import type { ResearchCandidate } from "./workspaceTypes";
+
+/** Selection spans pages; eligibility is supplied by the producing workflow. */
+export function CandidateTable({
+  candidates,
+  selected,
+  onSelectionChange,
+  disabled = false,
+  requireValidation = false,
+}: {
+  candidates: ResearchCandidate[];
+  selected: string[];
+  onSelectionChange: (ids: string[]) => void;
+  disabled?: boolean;
+  requireValidation?: boolean;
+}) {
+  const [page, setPage] = useState(1);
+  const [pageSize, setPageSize] = useState(25);
+  return (
+    <>
+      <WorkspaceTable<ResearchCandidate>
+        dataSource={candidates.slice((page - 1) * pageSize, page * pageSize)}
+        className="research-candidate-table"
+        rowKey="client_item_id"
+        scroll={{ x: 1120, y: 400 }}
+        empty="暂无候选表达式。"
+        rowSelection={{
+          width: 48,
+          disabled,
+          selectedRowKeys: selected,
+          onChange: (keys) => onSelectionChange((keys ?? []).map(String)),
+          getCheckboxProps: (candidate) => ({
+            disabled:
+              disabled ||
+              (requireValidation && candidate.validation?.status !== "valid"),
+            "aria-label": `选择候选 ${candidate.client_item_id}`,
+            title: requireValidation
+              ? [
+                  ...(candidate.validation?.syntax ?? []),
+                  ...(candidate.validation?.types ?? []),
+                  ...(candidate.validation?.availability ?? []),
+                ].join("；")
+              : undefined,
+          }),
+        }}
+        columns={[
+          {
+            title: "表达式",
+            dataIndex: "expression",
+            width: 420,
+            render: (value) => <code title={value}>{value}</code>,
+          },
+          ...(
+            [
+              ["neutralization", "Neutralization", 170],
+              ["decay", "Decay", 90],
+              ["truncation", "Truncation", 120],
+              ["maxTrade", "Max Trade", 120],
+              ["nanHandling", "NaN Handling", 150],
+            ] as const
+          ).map(([key, title, width]) => ({
+            key,
+            title,
+            width,
+            render: (_: unknown, candidate: ResearchCandidate) =>
+              candidate.settings[key],
+          })),
+        ]}
+      />
+      <div className="table-pagination workspace-table-footer">
+        <span>
+          共 {candidates.length} 个候选 · 已选 {selected.length} 个
+        </span>
+        <Pagination
+          currentPage={page}
+          pageSize={pageSize}
+          total={candidates.length}
+          showSizeChanger
+          pageSizeOpts={[25, 50, 100]}
+          preventPageChangeOnPageSizeChange
+          onPageChange={setPage}
+          onPageSizeChange={(size) => {
+            setPageSize(size);
+            setPage(1);
+          }}
+        />
+      </div>
+    </>
+  );
+}

frontend/src/research/ExperimentView.tsx

@@ -1,19 +1,12 @@
 import { useState } from "react";
-import {
+import { Banner, Button, Toast } from "@douyinfe/semi-ui-19";
-  Banner,
-  Button,
-  Checkbox,
-  Table,
-  Tag,
-  Toast,
-} from "@douyinfe/semi-ui-19";
 import { post, formatTime } from "../api";
 import type { UIAction } from "../ai/types";
 import type { Experiment } from "./workspaceTypes";
 import { EvaluationPanel } from "./EvaluationPanel";
 import { LineagePanel } from "./LineagePanel";
 import { ResearchSelect } from "./ResearchSelect";
-import { validationLabel } from "./workspaceTypes";
+import { CandidateTable } from "./CandidateTable";
 import { DeleteResearchButton } from "./DeleteResearchButton";
 
 export function ExperimentView({
@@ -33,7 +26,7 @@ export function ExperimentView({
   const [runId, setRunId] = useState(experiment.backtest_run_ids[0]);
   const [busy, setBusy] = useState(false);
   const valid = experiment.candidates.filter(
-    (c) => c.validation.status === "valid",
+    (c) => c.validation?.status === "valid",
   );
   async function preview() {
     setBusy(true);
@@ -72,7 +65,10 @@ export function ExperimentView({
           生成回测确认预览{selected.length ? `（${selected.length}）` : ""}
         </Button>
       </div>
+      <details>
+        <summary>研究描述</summary>
         <p>{experiment.hypothesis}</p>
+      </details>
       {experiment.archived ? (
         <Banner
           type="info"
@@ -92,175 +88,12 @@ export function ExperimentView({
         type="info"
         description="候选已保存为不可变研究记录。本地校验不保证平台可执行；在回测预览中确认后才开始模拟。"
       />
-      <div className="research-table-scroll">
+      <CandidateTable
-        <Table<Experiment["candidates"][number]>
+        candidates={experiment.candidates}
-          className="research-table"
+        selected={selected}
-          dataSource={experiment.candidates}
+        onSelectionChange={setSelected}
-          rowKey="client_item_id"
+        requireValidation
-          pagination={false}
-          size="small"
-          style={{ minWidth: 840 }}
-          empty="暂无候选表达式。"
-          columns={[
-            {
-              title: "选择",
-              key: "selection",
-              width: 64,
-              render: (_, candidate) => (
-                <Checkbox
-                  aria-label={`选择候选 ${candidate.client_item_id}`}
-                  disabled={candidate.validation.status !== "valid"}
-                  checked={selected.includes(candidate.client_item_id)}
-                  onChange={(event) =>
-                    setSelected((old) =>
-                      event.target.checked
-                        ? [...old, candidate.client_item_id]
-                        : old.filter((id) => id !== candidate.client_item_id),
-                    )
-                  }
       />
-              ),
-            },
-            {
-              title: "候选表达式",
-              key: "expression",
-              render: (_, candidate) => (
-                <>
-                  <code>{candidate.expression}</code>
-                  <details>
-                    <summary>绑定与改动</summary>
-                    <pre>
-                      {JSON.stringify(
-                        {
-                          bindings: candidate.bindings,
-                          changes: candidate.changes,
-                        },
-                        null,
-                        2,
-                      )}
-                    </pre>
-                  </details>
-                </>
-              ),
-            },
-            {
-              title: "市场与设置",
-              key: "settings",
-              width: 240,
-              render: (_, candidate) => (
-                <>
-                  {candidate.settings.region} / {candidate.settings.universe} /
-                  D{candidate.settings.delay}
-                  <br />
-                  {candidate.settings.neutralization} · decay{" "}
-                  {candidate.settings.decay}
-                </>
-              ),
-            },
-            {
-              title: "校验",
-              key: "validation",
-              width: 180,
-              render: (_, candidate) => (
-                <>
-                  <Tag
-                    color={
-                      candidate.validation.status === "valid"
-                        ? "green"
-                        : "orange"
-                    }
-                  >
-                    {validationLabel[candidate.validation.status]}
-                  </Tag>
-                  {(["syntax", "types", "availability"] as const).map((key) =>
-                    candidate.validation[key].map((issue, index) => (
-                      <p key={`${key}-${index}`}>
-                        {
-                          {
-                            syntax: "语法",
-                            types: "类型",
-                            availability: "可用性",
-                          }[key]
-                        }
-                        ：{issue}
-                      </p>
-                    )),
-                  )}
-                </>
-              ),
-            },
-          ]}
-        />
-      </div>
-      <div className="research-lineage">
-        <strong>研究来源</strong>
-        <span>实验 {experiment.id}</span>
-        {experiment.evidence.template?.id && (
-          <Button
-            theme="borderless"
-            onClick={() =>
-              onAction({
-                type: "open_template",
-                asset_id: experiment.evidence.template!.id,
-                version: experiment.evidence.template!.version,
-                nonce: Date.now(),
-              })
-            }
-          >
-            模板 v{experiment.evidence.template.version}
-          </Button>
-        )}
-        {experiment.inputs.map((input) => (
-          <Button
-            key={input.id}
-            theme="borderless"
-            onClick={() =>
-              onAction({
-                type: "open_research_input",
-                input_id: input.id,
-                nonce: Date.now(),
-              })
-            }
-          >
-            {input.name} · {input.scope.region}/{input.scope.universe}/D
-            {input.scope.delay}
-          </Button>
-        ))}
-        {experiment.parents.map((parent) => (
-          <Button
-            key={`${parent.kind}:${parent.id}`}
-            theme="borderless"
-            onClick={() =>
-              onAction(
-                parent.kind === "alpha"
-                  ? {
-                      type: "open_alpha",
-                      alpha_id: parent.id,
-                      nonce: Date.now(),
-                    }
-                  : {
-                      type: "open_experiment",
-                      experiment_id: parent.id,
-                      nonce: Date.now(),
-                    },
-              )
-            }
-          >
-            父来源：{parent.id}
-          </Button>
-        ))}
-        {experiment.backtest_run_ids.map((id) => (
-          <Button
-            key={id}
-            theme="borderless"
-            onClick={() =>
-              onAction({ type: "open_backtest", run_id: id, nonce: Date.now() })
-            }
-          >
-            查看关联回测
-          </Button>
-        ))}
-      </div>
       <div className="inline-actions">
         <Button onClick={() => setAssessment((v) => !v)}>评估研究结果</Button>
         <Button onClick={() => setRelations((v) => !v)}>查看变体关系</Button>

frontend/src/research/ResearchWorkspace.tsx

@@ -33,17 +33,11 @@ import {
 } from "./workspaceTypes";
 import { TemplateEditor } from "./TemplateEditor";
 import { ExperimentView } from "./ExperimentView";
+import { TemplateCandidateSet } from "./TemplateCandidateSet";
 import { DeleteResearchButton } from "./DeleteResearchButton";
 import { ComparisonPanel } from "./ComparisonPanel";
 import "./workspace.css";
 
-type ImportResult = {
-  templates: Template[];
-  digest: string;
-  conflicts: unknown[];
-  errors: unknown[];
-  differences?: unknown[];
-};
 export function ResearchWorkspace({
   page,
   active,
@@ -90,11 +84,27 @@ export function ResearchWorkspace({
   const [assetTotal, setAssetTotal] = useState(0);
   const [historyTotal, setHistoryTotal] = useState(0);
   const refreshSequence = useRef(0);
-  const [importText, setImportText] = useState("");
+  const CandidateView =
-  const [importResult, setImportResult] = useState<ImportResult | null>(null);
+    experiment?.kind === "template" ? TemplateCandidateSet : ExperimentView;
   const selectedInputs = inputs.filter((input) => inputIds.includes(input.id));
   const dirty =
     !asset || JSON.stringify(template) !== JSON.stringify(asset.content);
+  const generationKey = JSON.stringify([
+    asset?.id,
+    asset?.version,
+    template,
+    inputIds,
+    inputs,
+    settings,
+    mode,
+    limit,
+    seed,
+  ]);
+  const currentGeneration = useRef(generationKey);
+  currentGeneration.current = generationKey;
+  useEffect(() => {
+    if (page === "templates" && detailMode === "editor") setExperiment(null);
+  }, [page, template, inputIds, inputs, settings, mode, limit, seed]);
   async function task(label: string, action: () => Promise<void>) {
     setBusy(label);
     setError("");
@@ -108,13 +118,12 @@ export function ResearchWorkspace({
   }
   async function refresh() {
     const sequence = ++refreshSequence.current;
-    const [nextAssets, nextInputs, nextHistory] = await Promise.all([
+    const [nextAssets, nextHistory] = await Promise.all([
       api<{ items: Asset[]; total: number }>(
         `/research/assets?kind=template&limit=25&offset=${assetPage * 25}&q=${encodeURIComponent(search)}`,
       ),
-      Promise.resolve({ items: inputs }),
       api<{ items: typeof history; total: number }>(
-        `/research/experiments?limit=25&offset=${historyPage * 25}`,
+        `/research/experiments?kind=${page === "templates" ? "template" : "variant"}&limit=25&offset=${historyPage * 25}`,
       ),
     ]);
     if (sequence !== refreshSequence.current) return;
@@ -127,7 +136,6 @@ export function ResearchWorkspace({
       Math.min(current, Math.max(0, Math.ceil(nextHistory.total / 25) - 1)),
     );
     setAssets(nextAssets.items);
-    setInputs(nextInputs.items);
     setHistory(nextHistory.items);
   }
   useEffect(() => {
@@ -170,32 +178,34 @@ export function ResearchWorkspace({
         const next = await api<Asset>(
           `/research/assets/${action.asset_id}${action.version ? `?version=${action.version}` : ""}`,
         );
-        setAsset(next);
+        openTemplate(next);
-        setTemplate(next.content);
-        const feature = next.provenance?.feature;
-        if (feature) {
-          setInputIds(feature.content.input_ids);
-          setHypothesis(feature.content.hypothesis);
-          const fixed = feature.provenance?.inputs || [];
-          setInputs((old) => [
-            ...old,
-            ...fixed.filter((i) => !old.some((o) => o.id === i.id)),
-          ]);
-          const first = fixed[0];
-          if (first)
-            setSettings((old) => ({
-              ...old,
-              region: first.scope.region,
-              universe: first.scope.universe,
-              delay: first.scope.delay,
-            }));
-        }
       });
     if (action.type === "open_variant") {
       setParent(action.alpha_id);
       setMethod("structure");
     }
   }, [active, action]);
+  function openTemplate(next: Asset) {
+    setAsset(next);
+    setTemplate(next.content);
+    setExperiment(null);
+    const feature = next.provenance?.feature;
+    const fixed = feature?.provenance?.inputs || [];
+    setInputIds(feature?.content.input_ids || []);
+    setInputs(fixed);
+    setHypothesis(feature?.content.hypothesis || "");
+    const first = fixed[0];
+    setSettings({
+      ...initialSettings,
+      ...(first
+        ? {
+            region: first.scope.region,
+            universe: first.scope.universe,
+            delay: first.scope.delay,
+          }
+        : {}),
+    });
+  }
   async function save() {
     const next = asset
       ? await api<Asset>(`/research/assets/${asset.id}`, {
@@ -258,13 +268,19 @@ export function ResearchWorkspace({
       asset_id: asset!.id,
       version: asset!.version,
       ...researchSelection(inputIds, inputs),
-      hypothesis,
+      hypothesis:
+        page === "templates"
+          ? template.description.trim() || `使用模板：${template.name}`
+          : hypothesis,
       settings,
       mode,
       limit,
       seed,
       parent_alpha_ids: parent ? parent.split(/[,，\s]+/).filter(Boolean) : [],
     });
+    if (page === "templates" && currentGeneration.current !== generationKey) {
+      throw new Error("回测准备已改变，请按当前配置重新生成候选集合");
+    }
     setExperiment(next);
     setDetailTab("prepare");
     await refresh();
@@ -314,6 +330,10 @@ export function ResearchWorkspace({
                 setAsset(null);
                 setTemplate(blankTemplate());
                 setExperiment(null);
+                setInputIds([]);
+                setInputs([]);
+                setHypothesis("");
+                setSettings(initialSettings);
                 setDetailTab("editor");
                 setDetailMode("editor");
                 setDetailOpen(true);
@@ -368,7 +388,7 @@ export function ResearchWorkspace({
               dataSource={assets}
               rowKey="id"
               scroll={{ x: 800, y: "100%" }}
-              empty="还没有模板。新建或导入一个模板开始研究。"
+              empty="还没有模板。新建一个模板开始研究。"
               columns={[
                 {
                   title: "模板名称",
@@ -383,9 +403,7 @@ export function ResearchWorkspace({
                           const next = await api<Asset>(
                             `/research/assets/${item.id}`,
                           );
-                          setAsset(next);
+                          openTemplate(next);
-                          setTemplate(next.content);
-                          setExperiment(null);
                           setDetailTab("editor");
                           setDetailMode("editor");
                           setDetailOpen(true);
@@ -529,7 +547,7 @@ export function ResearchWorkspace({
         {error && <Banner type="danger" description={error} />}
         {detailMode === "experiment" ? (
           experiment && (
-            <ExperimentView
+            <CandidateView
               key={experiment.id}
               experiment={experiment}
               onAction={onAction}
@@ -544,7 +562,7 @@ export function ResearchWorkspace({
             {asset?.archived && (
               <Banner
                 type="info"
-                description="此模板已删除，当前显示保留的历史版本；可另存新模板继续编辑。"
+                description="此模板已删除，当前仅显示保留的历史版本。"
               />
             )}
             {page === "templates" && (
@@ -554,12 +572,13 @@ export function ResearchWorkspace({
                 onChange={setDetailTab}
                 tabPaneMotion={false}
                 tabList={[
-                  { itemKey: "editor", tab: "模板信息编辑" },
+                  { itemKey: "editor", tab: "模板详情" },
                   { itemKey: "prepare", tab: "回测准备" },
                 ]}
               />
             )}
-            <div hidden={page === "templates" && detailTab !== "editor"}>
+            {page === "variants" && (
+              <div>
                 <section className="research-card">
                   <h3>研究输入与假设</h3>
                   {page === "variants" && (
@@ -678,6 +697,7 @@ export function ResearchWorkspace({
                   )}
                 </section>
               </div>
+            )}
             {page !== "variants" || method === "structure" ? (
               <>
                 <div hidden={page === "templates" && detailTab !== "editor"}>
@@ -692,7 +712,26 @@ export function ResearchWorkspace({
                       value={template}
                       onChange={setTemplate}
                       inputs={selectedInputs}
+                      definitionsOnly={page === "templates"}
                     />
+                    {page === "templates" ? (
+                      <div className="research-toolbar">
+                        <Button
+                          theme="solid"
+                          disabled={!!busy || asset?.archived || !dirty}
+                          onClick={() => void task("保存模板", save)}
+                        >
+                          保存
+                        </Button>
+                        <Button
+                          disabled={!!busy || !asset || asset.archived}
+                          onClick={() => void task("新增版本", save)}
+                        >
+                          新增版本
+                        </Button>
+                      </div>
+                    ) : (
+                      <>
                         {asset && (
                           <div className="research-toolbar">
                             <span>版本记录</span>
@@ -705,7 +744,10 @@ export function ResearchWorkspace({
                                     assets.find((item) => item.id === asset.id)
                                       ?.version || asset.version,
                                 },
-                            (_, i) => ({ value: i + 1, label: `v${i + 1}` }),
+                                (_, i) => ({
+                                  value: i + 1,
+                                  label: `v${i + 1}`,
+                                }),
                               )}
                               onChange={(version) =>
                                 void task("读取历史版本", async () => {
@@ -749,28 +791,21 @@ export function ResearchWorkspace({
                           )}
                           <span>组合规模：{combination}</span>
                         </div>
-                    {page === "templates" && (
+                      </>
-                      <Button
-                        theme="solid"
-                        disabled={!!busy || dirty}
-                        onClick={() => setDetailTab("prepare")}
-                      >
-                        下一步：回测准备
-                      </Button>
                     )}
                   </section>
                 </div>
                 <div hidden={page === "templates" && detailTab !== "prepare"}>
                   <section className="research-card">
                     <h3>回测准备</h3>
-                    <p className="research-hint">
+                    <div aria-label="回测表达式">
-                      配置展开方式与模拟设置 → 生成候选研究集合 →
+                      <h4>表达式</h4>
-                      生成回测确认预览 → 确认后开始回测。
+                      <pre>{template.expression}</pre>
-                    </p>
+                    </div>
                     {page === "templates" && (
                       <>
                         <label>
-                          固定研究输入
+                          数据准备
                           <ResearchDataInput
                             label="选择回测数据准备"
                             ids={inputIds}
@@ -789,24 +824,10 @@ export function ResearchWorkspace({
                             }}
                           />
                         </label>
-                        <label>
-                          本次研究假设
-                          <TextArea
-                            aria-label="回测研究假设"
-                            value={hypothesis}
-                            onChange={setHypothesis}
-                            autosize={{ minRows: 2, maxRows: 5 }}
-                          />
-                        </label>
-                        <p>
-                          模板：{template.name || "未命名"} ·{" "}
-                          {asset ? `v${asset.version}` : "未保存"} · 组合规模：
-                          {combination}
-                        </p>
                         {dirty && (
                           <Banner
                             type="warning"
-                            description="请返回模板信息编辑，保存模板版本后再生成候选。"
+                            description="请返回模板详情，保存模板版本后再生成候选。"
                           />
                         )}
                       </>
@@ -844,6 +865,9 @@ export function ResearchWorkspace({
                       </label>
                     </div>
                     <SimulationSettingsEditor
+                      validationMode={
+                        page === "templates" ? "combination" : "platform"
+                      }
                       value={settings}
                       onChange={setSettings}
                       {...options}
@@ -857,20 +881,22 @@ export function ResearchWorkspace({
                       disabled={
                         dirty ||
                         !inputIds.length ||
-                        !hypothesis ||
+                        (page === "variants" && !hypothesis) ||
                         !!busy ||
                         !settingsValid
                       }
                       onClick={() => void task("生成候选", expand)}
                     >
-                      生成候选研究集合
+                      生成候选集合
                     </Button>
-                    {(!inputIds.length || !hypothesis || !settingsValid) && (
+                    {(!inputIds.length ||
+                      (page === "variants" && !hypothesis) ||
+                      !settingsValid) && (
                       <p className="research-hint">
                         待完成：
                         {[
                           !inputIds.length && "选择固定研究输入",
-                          !hypothesis && "填写研究假设",
+                          page === "variants" && !hypothesis && "填写研究假设",
                           !settingsValid && "检查模拟设置",
                         ]
                           .filter(Boolean)
@@ -908,7 +934,7 @@ export function ResearchWorkspace({
             )}
             <div hidden={page === "templates" && detailTab !== "prepare"}>
               {experiment && (
-                <ExperimentView
+                <CandidateView
                   key={experiment.id}
                   experiment={experiment}
                   onAction={onAction}
@@ -919,65 +945,6 @@ export function ResearchWorkspace({
                 />
               )}
             </div>
-            {page === "templates" && detailTab === "editor" && (
-              <details className="research-card">
-                <summary>导入旧模板</summary>
-                <p>
-                  粘贴 cnhk 模板 JSON
-                  或模板数组。先预览转换结果，同名模板不会自动覆盖。
-                </p>
-                <TextArea
-                  aria-label="导入模板 JSON"
-                  value={importText}
-                  onChange={(text) => {
-                    setImportText(text);
-                    setImportResult(null);
-                  }}
-                  autosize={{ minRows: 4, maxRows: 12 }}
-                />
-                <Button
-                  disabled={!!busy || !importText}
-                  onClick={() =>
-                    void task("预览导入", async () => {
-                      const data = JSON.parse(importText);
-                      setImportResult(
-                        await post("/research/templates/import-preview", {
-                          templates: Array.isArray(data) ? data : [data],
-                        }),
-                      );
-                    })
-                  }
-                >
-                  预览导入差异
-                </Button>
-                {importResult && (
-                  <>
-                    <pre>{JSON.stringify(importResult, null, 2)}</pre>
-                    <Button
-                      disabled={
-                        !!busy ||
-                        !!importResult.errors?.length ||
-                        !!importResult.conflicts?.length
-                      }
-                      onClick={() =>
-                        void task("导入模板", async () => {
-                          await post("/research/templates/import", {
-                            templates: importResult.templates,
-                            digest: importResult.digest,
-                          });
-                          setImportResult(null);
-                          setImportText("");
-                          await refresh();
-                          Toast.success("导入完成");
-                        })
-                      }
-                    >
-                      确认导入
-                    </Button>
-                  </>
-                )}
-              </details>
-            )}
             {page === "variants" && (
               <ComparisonPanel key={parent} baseline={parent} />
             )}

frontend/src/research/TemplateCandidateSet.tsx

@@ -0,0 +1,106 @@
+import { useRef, useState } from "react";
+import { Banner, Button, Toast } from "@douyinfe/semi-ui-19";
+import { post } from "../api";
+import type { UIAction } from "../ai/types";
+import { CandidateTable } from "./CandidateTable";
+import { DeleteResearchButton } from "./DeleteResearchButton";
+import type { Experiment } from "./workspaceTypes";
+
+/** The generated collection is the confirmation surface for a template backtest. */
+export function TemplateCandidateSet({
+  experiment,
+  onAction,
+  onDelete,
+  deleting,
+}: {
+  experiment: Experiment;
+  onAction: (action: UIAction) => void;
+  onDelete?: () => Promise<void>;
+  deleting?: boolean;
+}) {
+  const [selected, setSelected] = useState(() =>
+    experiment.candidates.map((c) => c.client_item_id),
+  );
+  const [busy, setBusy] = useState(false);
+  const [runs, setRuns] = useState(experiment.backtest_run_ids);
+  const attempt = useRef<{ selection: string; key: string } | null>(null);
+  const pending = useRef(false);
+  async function backtest() {
+    if (pending.current || !selected.length) return;
+    pending.current = true;
+    setBusy(true);
+    // Retain the same request key after network failure: retry must not create a second run.
+    const selection = JSON.stringify([...selected].sort());
+    if (attempt.current?.selection !== selection) {
+      attempt.current = { selection, key: crypto.randomUUID() };
+    }
+    try {
+      const result = await post<{ backtest_run_id: string }>(
+        `/research/experiments/${experiment.id}/backtest`,
+        {
+          candidate_ids: selected,
+          idempotency_key: attempt.current.key,
+        },
+      );
+      setRuns((old) => [...new Set([...old, result.backtest_run_id])]);
+      onAction({
+        type: "open_backtest",
+        run_id: result.backtest_run_id,
+        nonce: Date.now(),
+      });
+    } catch (error) {
+      Toast.error((error as Error).message);
+    } finally {
+      pending.current = false;
+      setBusy(false);
+    }
+  }
+  return (
+    <section className="research-card" aria-label="回测候选集合">
+      <h3>回测候选集合</h3>
+      <details>
+        <summary>研究描述</summary>
+        <p>{experiment.hypothesis}</p>
+      </details>
+      {experiment.archived && (
+        <Banner type="info" description="此候选集合已删除，保留历史关联。" />
+      )}
+      <p>确认表达式和参数后，点击“回测”将所选候选交给回测研究批量执行。</p>
+      <CandidateTable
+        candidates={experiment.candidates}
+        selected={selected}
+        onSelectionChange={setSelected}
+        disabled={busy || deleting || experiment.archived}
+      />
+      <div className="research-toolbar">
+        <Button
+          theme="solid"
+          disabled={!selected.length || deleting || experiment.archived}
+          loading={busy}
+          onClick={() => void backtest()}
+        >
+          回测（{selected.length}）
+        </Button>
+        {onDelete && !experiment.archived && (
+          <DeleteResearchButton
+            name={experiment.name}
+            label="候选集合"
+            disabled={busy || deleting}
+            onConfirm={onDelete}
+          />
+        )}
+        {runs.map((id, index) => (
+          <Button
+            key={id}
+            theme="borderless"
+            onClick={() =>
+              onAction({ type: "open_backtest", run_id: id, nonce: Date.now() })
+            }
+          >
+            关联回测 {index + 1}
+          </Button>
+        ))}
+      </div>
+    </section>
+  );
+}

frontend/src/research/TemplateEditor.tsx

@@ -14,10 +14,12 @@ export function TemplateEditor({
   value,
   onChange,
   inputs,
+  definitionsOnly = false,
 }: {
   value: Template;
   onChange: (value: Template) => void;
   inputs: InputSnapshot[];
+  definitionsOnly?: boolean;
 }) {
   function expression(text: string) {
     const names = [
@@ -33,7 +35,9 @@ export function TemplateEditor({
       variables: Object.fromEntries(
         names.map((name) => [
           name,
-          value.variables[name] || {
+          Object.hasOwn(value.variables, name)
+            ? value.variables[name]
+            : {
                 kind: "field",
                 field_type: "MATRIX",
                 values: [],
@@ -95,7 +99,56 @@ export function TemplateEditor({
         使用 {"{name}"}{" "}
         声明变量；重复占位符共享同一个取值。组合片段只复用表达式，不修改平台算子定义。
       </p>
-      {Object.entries(value.variables).map(([name, item]) => (
+      {definitionsOnly && Object.keys(value.variables).length > 0 && (
+        <h4>字段配置</h4>
+      )}
+      {Object.entries(value.variables).map(([name, item]) =>
+        definitionsOnly ? (
+          <div className="research-variable-definition" key={name}>
+            <strong>{`{${name}}`}</strong>
+            <label>
+              字段类型
+              <ResearchSelect
+                label={`${name} 字段类型`}
+                value={item.kind === "field" ? item.field_type : item.kind}
+                optionList={[
+                  ...["MATRIX", "VECTOR", "GROUP"].map((value) => ({
+                    value,
+                    label: value,
+                  })),
+                  ...Object.entries(kinds)
+                    .filter(([kind]) => kind !== "field")
+                    .map(([value, label]) => ({ value, label })),
+                ]}
+                onChange={(type) => {
+                  const isField = ["MATRIX", "VECTOR", "GROUP"].includes(
+                    String(type),
+                  );
+                  variable(name, {
+                    kind: isField ? "field" : (type as Variable["kind"]),
+                    ...(isField
+                      ? { field_type: type as Variable["field_type"] }
+                      : {}),
+                    description: item.description || "",
+                    values: [],
+                  });
+                }}
+              />
+            </label>
+            <label>
+              字段描述
+              <Input
+                aria-label={`${name} 字段描述`}
+                value={item.description || ""}
+                maxLength={3000}
+                placeholder="说明字段含义与用途"
+                onChange={(description) =>
+                  variable(name, { ...item, description })
+                }
+              />
+            </label>
+          </div>
+        ) : (
           <div className="research-variable" key={name}>
             <strong>{`{${name}}`}</strong>
             <ResearchSelect
@@ -145,7 +198,8 @@ export function TemplateEditor({
                   values: text
                     .split("\n")
                     .map((v) =>
-                    ["integer", "number"].includes(item.kind) && v.trim() !== ""
+                      ["integer", "number"].includes(item.kind) &&
+                      v.trim() !== ""
                         ? Number(v)
                         : v,
                     ),
@@ -174,7 +228,8 @@ export function TemplateEditor({
               </Button>
             )}
           </div>
-      ))}
+        ),
+      )}
     </div>
   );
 }

frontend/src/research/workspace.css

@@ -374,3 +374,32 @@
   flex-shrink: 0;
   padding: 8px 0 12px;
 }
+
+.research-variable-definition {
+  display: grid;
+  grid-template-columns: minmax(90px, 0.5fr) minmax(140px, 1fr) minmax(
+      200px,
+      2fr
+    );
+  gap: 12px;
+  align-items: center;
+  border-top: 1px solid var(--semi-color-border);
+}
+.research-variable-definition > strong {
+  overflow-wrap: anywhere;
+}
+@media (max-width: 640px) {
+  .research-variable-definition {
+    grid-template-columns: minmax(0, 1fr);
+    gap: 0;
+    padding-top: 12px;
+  }
+}
+
+.research-card .research-candidate-table .semi-checkbox {
+  margin: 0;
+  flex-direction: row;
+}
+.research-candidate-table .workspace-cell code {
+  white-space: nowrap;
+}

frontend/src/research/workspaceTypes.ts

@@ -10,6 +10,7 @@ export type Variable = {
     | "string"
     | "fragment";
   values: (string | number)[];
+  description?: string;
   field_type?: "MATRIX" | "VECTOR" | "GROUP" | null;
 };
 export type Template = {
@@ -18,7 +19,6 @@ export type Template = {
   expression: string;
   variables: Record<string, Variable>;
   category: "template" | "fragment";
-  scope?: InputSnapshot["scope"] | null;
 };
 export type Asset = {
   id: string;
@@ -55,7 +55,7 @@ export type InputSnapshot = {
   created_at: string;
 };
 export type ResearchCandidate = Candidate & {
-  validation: {
+  validation?: {
     status: string;
     syntax: string[];
     types: string[];

frontend/src/settings/SimulationSettingsEditor.tsx

@@ -28,7 +28,9 @@ export function SimulationSettingsEditor({
   fixedScopes,
   disabled = false,
   onValidityChange,
+  validationMode = "platform",
 }: {
+  validationMode?: "platform" | "combination";
   alphaType?: "REGULAR" | "SUPER";
   value: SimulationSettings;
   onChange: (value: SimulationSettings) => void;
@@ -43,10 +45,16 @@ export function SimulationSettingsEditor({
 }) {
   const scope = settingsScope(value);
   const row = rows.find((r) => sameScope(r, scope));
-  const errors = settingsErrors(value, rows, fixedScopes);
+  const combinationOnly = validationMode === "combination";
+  const errors = combinationOnly
+    ? !fixedScopes?.length || fixedScopes.some((s) => !sameScope(s, scope))
+      ? ["所选数据准备必须具有相同组合，并与回测组合一致"]
+      : []
+    : settingsErrors(value, rows, fixedScopes);
   const [scopeReady, setScopeReady] = useState(false);
-  const valid =
+  const valid = combinationOnly
-    !loading &&
+    ? !errors.length
+    : !loading &&
       !error &&
       !errors.length &&
       (fixedScopes !== undefined || scopeReady);
@@ -54,7 +62,19 @@ export function SimulationSettingsEditor({
     onValidityChange?.(valid);
   }, [valid, onValidityChange]);
   function field(key: string) {
-    const options = fieldOptions(row, key);
+    const options = combinationOnly
+      ? key === "neutralization"
+        ? {
+            choices: [
+              ...new Set([
+                value.neutralization,
+                ...(row?.neutralizations ??
+                  rows.flatMap((r) => r.neutralizations)),
+              ]),
+            ],
+          }
+        : fieldOptions(undefined, key)
+      : fieldOptions(row, key);
     const current =
       value[key as keyof SimulationSettings] ??
       (key === "maxPosition" ? "OFF" : undefined);
@@ -71,7 +91,7 @@ export function SimulationSettingsEditor({
             value={typeof current === "boolean" ? String(current) : current}
             disabled={
               disabled ||
-              loading ||
+              (loading && !combinationOnly) ||
               !options.choices.length ||
               (options.choices.length === 1 && !message)
             }
@@ -97,7 +117,7 @@ export function SimulationSettingsEditor({
                 ? current
                 : undefined
             }
-            disabled={disabled || loading}
+            disabled={disabled || (loading && !combinationOnly)}
             min={options.minimum}
             max={options.maximum}
             precision={
@@ -129,16 +149,30 @@ export function SimulationSettingsEditor({
         选项更新时间：{formatTime(fetchedAt ?? null)}
       </small>
       {error && (
-        <Banner type="warning" description={`${error}；请重试同步合法设置`} />
+        <Banner
+          type="warning"
+          description={
+            combinationOnly
+              ? "参数选项暂不可用，可使用当前参数继续生成候选。"
+              : `${error}；请重试同步合法设置`
+          }
+        />
       )}
       {!loading && !rows.length && !error && (
-        <Banner type="info" description="尚无参数选项，请先同步合法设置。" />
+        <Banner
+          type="info"
+          description={
+            combinationOnly
+              ? "可同步参数选项辅助选择，当前参数仍可用于生成候选。"
+              : "尚无参数选项，请先同步合法设置。"
+          }
+        />
       )}
       <ScopePicker
         value={scope}
         options={asScopeOptions(rows)}
         readOnly={fixedScopes !== undefined}
-        disabled={disabled || loading}
+        disabled={disabled || (loading && !combinationOnly)}
         onValidityChange={setScopeReady}
         onChange={(next) =>
           onChange({

scripts/deploy-production.sh

@@ -6,7 +6,7 @@ umask 077
 
 cd "$(dirname "${BASH_SOURCE[0]}")/.."
 # Fail before touching the host; never read a checkout's local .env file.
-for key in WQ_EMAIL WQ_PASSWORD DATABASE_URL ADMIN_PASSWORD ENCRYPTION_KEY DATABASE_NETWORK PUBLIC_ORIGIN; do
+for key in WQ_EMAIL WQ_PASSWORD DATABASE_URL ADMIN_PASSWORD ENCRYPTION_KEY PUBLIC_ORIGIN; do
   if [[ -z "${!key:-}" ]]; then
     echo "Missing required Gitea configuration: $key" >&2
     exit 1