fix(investigation): persist large results and require graph links

Full-network pressure comparisons exceeded the worker limit before saved-result previews could run. Preserve complete calculation results and keep read-time summaries and pagination.

Incremental graph publication accepted isolated nodes after flow-authoring rules were lost. Require new steps to connect explicitly while retaining historical snapshots and status-only patches. Regression tests reproduce both failures on the previous code.

app/services/calculation_worker.py

@@ -93,8 +93,8 @@ async def execute_child(input_path: str, output_path: str):
             category = 'EXECUTION_CONTEXT_UNAVAILABLE'
         payload = {'status': 'failed', 'error': {'code': category}}
     encoded = json.dumps(payload, allow_nan=False)
-    if len(encoded.encode()) > 8 * 1024 * 1024:
-        encoded = json.dumps({'status': 'failed', 'error': {'code': 'RESULT_TOO_LARGE'}})
+    # Keep the complete calculation for persistence. MCP result reads apply
+    # their own summary/page budgets after the result has been saved.
     Path(output_path).write_text(encoded)
 
 
@@ -212,8 +212,6 @@ class CalculationWorker:
                         terminal = {'status': 'timed_out', 'error': {'code': 'EXECUTION_TIMEOUT'}}
                     elif process.returncode is not None:
                         try:
-                            if output_path.stat().st_size > 8 * 1024 * 1024:
-                                raise ValueError('Result exceeds persistence limit')
                             terminal = json.loads(output_path.read_text())
                         except (OSError, ValueError):
                             terminal = {'status': 'failed', 'error': {'code': 'CHILD_PROCESS_FAILED'}}

app/services/investigation_graph.py

@@ -39,6 +39,7 @@ def update_investigation_graph(
         ))
 
     nodes = {node.node_id: node.model_copy(deep=True) for node in previous.node_refs} if previous else {}
+    saved_ids = set(nodes)
     require(len({node.node_id for node in update.nodes}) == len(update.nodes),
             "GRAPH_NODE_DUPLICATE", "Graph update contains duplicate node ids", status=422)
     for node in update.nodes:
@@ -81,6 +82,26 @@ def update_investigation_graph(
             edges.append(edge.model_copy(deep=True))
     require(len(nodes) <= 100 and len(edges) <= 200, "GRAPH_CAPACITY_EXCEEDED",
             "Investigation graph exceeds 100 nodes or 200 edges; existing paths were retained", status=422)
+    # Validate only newly added steps. Historical snapshots and status-only
+    # patches stay valid, even if an earlier publisher omitted relationships.
+    new_ids = set(nodes) - saved_ids
+    if new_ids and len(nodes) > 1:
+        neighbours = {node_id: set() for node_id in nodes}
+        for edge in edges:
+            neighbours[edge.source].add(edge.target)
+            neighbours[edge.target].add(edge.source)
+        connected = set(saved_ids) or {next(iter(nodes))}
+        pending = list(connected)
+        while pending:
+            for node_id in neighbours[pending.pop()] - connected:
+                connected.add(node_id)
+                pending.append(node_id)
+        missing = sorted(new_ids - connected)
+        require(not missing, "GRAPH_RELATION_REQUIRED",
+                "Add explicit edges connecting these new steps to the investigation narrative: "
+                + ", ".join(missing)
+                + ". Describe branches, parallel findings or retries; edges do not schedule execution.",
+                status=422)
     lines = [source or "flowchart TD"]
     for node in nodes.values():
         if in_source(node.node_id):

docs/calculation-queue.md

@@ -43,6 +43,9 @@ API 保留 `uvicorn --workers 4`。长计算在项目 BizDB 持久化入队，
 
 调查的排队回执不写成分析证据。成功结果读取时，按原计算来源、任务及运行验证，并幂等地保存证据。未完成的计算阻止提交最终报告、`insufficient_data` 或将运行标记成功；中间进度仍可保存。
 
+计算子进程的完整结果文件及 supervisor 读取不设 8MiB 整份体积上限，全网分析结果按原值保存到 `calculation_jobs.result`。
+Agent 读取仍使用 `get_calculation_result` 的摘要及 `data_path/limit/offset` 分页预算；完整结果保存与单次工具消息大小分别处理。
+
 Agent 调查执行在原生回合空闲后由宿主短轮询等待，继续维持调查心跳。新作业进入终态即触发模型读取结果，其余作业继续后台运行。状态读取遇到网络、限流或临时服务错误，最多重试 3 次，采用指数退避和随机抖动；权限拒绝、上下文失效立即停止，重试耗尽明确报错，不重新提交或取消计算。普通聊天展示任务编号，可在当前会话后续查询。异常流结束时若无法安全继续处理新结果，会明确失败而不是虚报完成；Server 中的计算结果仍保留。
 
 ## 排队期限与结果保留

docs/investigation-progress-and-presentation.md

@@ -57,6 +57,12 @@ Server 不验证依赖执行顺序或推断节点完工。
 `progress.graph_update` 包含可选 `title`、`nodes` 和 `edges`。节点沿用 `node_id/label/status/evidence_refs/action_refs/map_targets`，
 可用 `detail` 记录实际结果或失败原因；连线为 `{source,target,label?}`。仅提交新增或变化的内容，
 `update_investigation_graph` 合并节点、追加明确连线并生成 Mermaid，保留已有路径和引用。
+新建调查图使用 `flowchart TD`，按自上而下的阅读顺序展示。多步骤图必须显式连接流程关系：
+独立核查从共同问题分叉并在判断处汇总，条件分支标注条件，失败尝试连接重试或替代路线。
+新增节点须通过连线接入已有调查脉络；首次发布多个节点时须形成相连的图。
+未接入的新增节点返回 `GRAPH_RELATION_REQUIRED`（422），错误中列出需补关系的节点。
+根据真实关系修正后重试，不按节点列表顺序自动补线；关系仅解释调查，不约束执行先后。
+单个起点、仅更新已有节点、纯文字更新和已保存的历史图仍可不传连线，不重排或改写历史报告。
 “保留路径”约束图的结构，不冻结已有节点的标记。`nodes` 按 `node_id` 合并：新增节点必须提供 `label`；
 修改已保存节点时可只提交 `node_id` 与变化字段，无需再提供 `label`、全部节点或连线。
 省略字段保留原值，`label: null` 也保留名称；`status: null` 显式恢复为参考步骤。

docs/mcp-query-implementation.md

@@ -16,7 +16,7 @@ Agent 根据401 challenge 或路径/根 well-known 发现受保护资源元数
 
 普通查询默认20条；目录类最多500条，显式指定设备的时序查询最多10000条、20个标识、24小时；offset 最大10000。SCADA调查首次省略 `query_scada_readings.device_ids`，由Server读取项目完整设备目录及窗口内全部授权观测，内部自动翻页，`offset` 必须为0。其余查询按问题选定对象和窗口。SQL 有 limit/offset 和 statement_timeout。历史对比在同一条SQL中按序列、时间对数据点分页，返回 page.has_more；非零offset也标明覆盖不完整。历史曲线展示调用初始每页500点，`RESULT_TOO_LARGE` 时原位减半重试，最多读取1000条、发出16次请求，共用120秒超时（单次请求同时受该剩余总预算约束），图表最多100个有效点；这不改变模型首次查询默认值或时序接口上限。
 
-单份结果超过8MiB报错；超过12KB的调查工具消息默认返回证据资源引用，设备目录的精简路径见下段。完整快照仍保存，Server 的 `resources/read` 可以读取它。Agent 本地 MCP 只提供工具，将上游资源引用转为 `evidence_read` 中的 `mcp_tjwater_get_task_evidence` 参数，不向 OpenCode 泄露不可在本地读取的链接。`get_task_evidence` 在数据库中按 data_path、limit、offset 选页，嵌套集合和超过256字符的字符串返回路径描述；字符串按 Unicode 字符分页。页和完整资源均在数据库返回前检查8MiB上限，保留原证据编号和合成来源；普通查询要求缩小范围。其他结果仍保留执行后的整体8MiB检查，不能把它理解为所有查询均有数据库端字节限制。原始/清洗实测、模型基线、分析结果、比较和 synthetic 均明确标记。非有限数值转为 null 并记录数据缺口。
+同步查询单份结果超过8MiB报错；后台计算作业的完整结果保存不设此上限，读取时仍执行工具消息摘要和分页预算；超过12KB的调查工具消息默认返回证据资源引用，设备目录的精简路径见下段。完整快照仍保存，Server 的 `resources/read` 可以读取它。Agent 本地 MCP 只提供工具，将上游资源引用转为 `evidence_read` 中的 `mcp_tjwater_get_task_evidence` 参数，不向 OpenCode 泄露不可在本地读取的链接。`get_task_evidence` 在数据库中按 data_path、limit、offset 选页，嵌套集合和超过256字符的字符串返回路径描述；字符串按 Unicode 字符分页。页和完整资源均在数据库返回前检查8MiB上限，保留原证据编号和合成来源；普通查询要求缩小范围。其他结果仍保留执行后的整体8MiB检查，不能把它理解为所有查询均有数据库端字节限制。原始/清洗实测、模型基线、分析结果、比较和 synthetic 均明确标记。非有限数值转为 null 并记录数据缺口。
 
 `list_scada_devices` 默认每页100个、最多500个，其他集合仍沿用各自的分页默认值。设备目录超过12,000字节时，先内联本页全部设备的 `device_id`、`device_type` 和 `measurement_unit`，省略其他属性并明确标注；模型映射、坐标等通过 `get_scada_device` 核验，任务证据和幂等回执保留完整属性。此时 `truncated` 表示属性省略，设备是否还有后续页以 `page.has_more` 为准，按本页实际设备数递增 `offset`。精简目录仍超预算时沿用证据引用或缩页路径。
 

tests/unit/test_calculation_worker.py

@@ -1,5 +1,6 @@
 """Worker boundaries use real managed subprocesses, without external databases."""
 import asyncio
+import json
 from functools import wraps
 from types import SimpleNamespace
 from contextlib import asynccontextmanager
@@ -27,6 +28,66 @@ def run_async(function):
     return wrapper
 
 
+@run_async
+async def test_child_keeps_complete_result_above_eight_mib(monkeypatch, tmp_path):
+    from app.mcp import calculation_jobs
+    from app.services import calculation_worker
+
+    data = {"comparisons": "x" * (8 * 1024 * 1024), "map_payload": {"features": ["coverage"]}}
+
+    @asynccontextmanager
+    async def scope(row):
+        yield object()
+
+    async def calculate(scope, row):
+        return SimpleNamespace(model_dump=lambda **kwargs: data)
+
+    monkeypatch.setattr(calculation_worker, "execution_scope", scope)
+    monkeypatch.setattr(calculation_jobs, "execute_calculation_job", calculate)
+    input_path, output_path = tmp_path / "input.json", tmp_path / "result.json"
+    input_path.write_text("{}")
+    await calculation_worker.execute_child(str(input_path), str(output_path))
+    assert output_path.stat().st_size > 8 * 1024 * 1024
+    assert json.loads(output_path.read_text()) == {"status": "succeeded", "result": data}
+
+
+@run_async
+async def test_supervisor_persists_complete_result_above_eight_mib(monkeypatch, tmp_path):
+    monkeypatch.setenv("CALCULATION_WORKER_STATE_DIR", str(tmp_path))
+    original = asyncio.create_subprocess_exec
+
+    async def spawn(*args, **kwargs):
+        # A real child writes the same file protocol as the calculation process.
+        code = ('import json,pathlib,sys; pathlib.Path(sys.argv[1]).write_text('
+                'json.dumps({"status":"succeeded","result":{"comparisons":"x"*(8*1024*1024),'
+                '"map_payload":{"features":["coverage"]}}}))')
+        return await original(sys.executable, "-c", code, args[-1], start_new_session=True)
+
+    monkeypatch.setattr(asyncio, "create_subprocess_exec", spawn)
+    results = []
+
+    class Repository:
+        async def heartbeat(self, *args):
+            return "running"
+
+        async def finish(self, row, owner, **terminal):
+            results.append(terminal)
+            return True
+
+    @asynccontextmanager
+    async def repository(project):
+        yield Repository()
+
+    worker = CalculationWorker(WorkerSettings(heartbeat_seconds=.02), repository=repository)
+    await asyncio.wait_for(worker.run_job({"project_id": "project", "job_id": "large-result",
+                                          "timeout_seconds": 10}), 15)
+    assert len(results) == 1
+    assert results[0]["status"] == "succeeded"
+    assert results[0]["result"] == {
+        "comparisons": "x" * (8 * 1024 * 1024), "map_payload": {"features": ["coverage"]},
+    }
+
+
 def test_worker_settings_require_valid_concurrency(monkeypatch):
     settings = WorkerSettings.from_environment()
     assert settings.max_concurrency == 2

tests/unit/test_contaminant_simulation_mcp.py

@@ -42,27 +42,29 @@ def context(monkeypatch):
     return scope, binding
 
 
-def test_replay_prediction_can_extend_past_cutoff_without_reading_future_observations(monkeypatch):
+@pytest.mark.parametrize("include_impact", [True, False])
+def test_replay_prediction_can_extend_past_cutoff_without_reading_future_observations(monkeypatch, include_impact):
     scope, binding = context(monkeypatch)
     worker = AsyncMock(return_value={"impact": {"affected_node_count": 2}})
     monkeypatch.setattr(analysis_service, "calculate_replay_contaminant_simulation", worker)
-    result = asyncio.run(query_service.execute_query(scope, binding, TOOLS["run_contaminant_simulation"], args()))
+    parameters = args(include_impact_coverage=include_impact)
+    result = asyncio.run(query_service.execute_query(scope, binding, TOOLS["run_contaminant_simulation"], parameters))
     assert result.origin == "synthetic" and result.data["result"]["impact"]["affected_node_count"] == 2
     assert "预测" in result.data_gaps[0]
     scope.connection.commit.assert_awaited_once()
     worker.assert_awaited_once()
+    assert worker.await_args.args[0].include_impact_coverage is include_impact
     assert worker.await_args.kwargs == {"network": "project", "username": "actor"}
     assert should_queue(TOOLS["run_contaminant_simulation"], args())
 
 
-@pytest.mark.parametrize("problem", ["future_start", "no_impact", "reader", "permission"])
+@pytest.mark.parametrize("problem", ["future_start", "reader", "permission"])
 def test_replay_admission_keeps_time_permission_and_reader_boundaries(monkeypatch, problem):
     scope, binding = context(monkeypatch)
     parameters = args()
     if problem == "future_start": parameters = args(start_time="2026-07-16T08:15:00+08:00")
-    if problem == "no_impact": parameters = args(include_impact_coverage=False)
     if problem == "reader": binding.purpose = "task_reader"
     if problem == "permission": binding.permission_cap.remove("scada.view")
     with pytest.raises(HTTPException) as caught:
         asyncio.run(query_service.authorize_query(scope, binding, TOOLS["run_contaminant_simulation"], parameters))
-    assert caught.value.status_code == (422 if problem in {"future_start", "no_impact"} else 403)
+    assert caught.value.status_code == (422 if problem == "future_start" else 403)

tests/unit/test_investigation_graph.py

@@ -239,6 +239,85 @@ def test_graph_renderer_escapes_labels_and_keeps_explicit_parallel_and_failed_pa
     assert graph.node_refs[2].status == "reference"
 
 
+@pytest.mark.parametrize("edges", [[], [{"source": "check", "target": "check"}]])
+def test_new_multistep_graph_requires_relationships_instead_of_isolated_nodes(edges):
+    with pytest.raises(DomainError) as error:
+        update_investigation_graph(None, InvestigationGraphUpdate.model_validate({
+            "nodes": [{"node_id": "read", "label": "观测核验"},
+                      {"node_id": "check", "label": "异常判断"}],
+            "edges": edges,
+        }))
+    assert error.value.code == "GRAPH_RELATION_REQUIRED"
+    assert "check" in str(error.value)
+
+
+def test_new_branch_must_connect_to_the_saved_narrative_without_inventing_edges():
+    original = update_investigation_graph(None, InvestigationGraphUpdate.model_validate({
+        "nodes": [{"node_id": "root", "label": "调查起点"}],
+    }))
+    with pytest.raises(DomainError) as error:
+        update_investigation_graph(original, InvestigationGraphUpdate.model_validate({
+            "nodes": [{"node_id": "left", "label": "独立核查"},
+                      {"node_id": "right", "label": "核查发现"}],
+            "edges": [{"source": "left", "target": "right"}],
+        }))
+    assert error.value.code == "GRAPH_RELATION_REQUIRED"
+    assert original.edges == []
+    assert len(original.node_refs) == 1
+
+
+def test_connected_parallel_branches_and_join_render_top_down():
+    graph = update_investigation_graph(None, InvestigationGraphUpdate.model_validate({
+        "nodes": [{"node_id": key, "label": label} for key, label in (
+            ("start", "异常核验"), ("isolation", "隔离分析"),
+            ("impact", "爆管影响"), ("summary", "汇总判断"))],
+        "edges": [{"source": source, "target": target, "label": label}
+                  for source, target, label in (
+                      ("start", "isolation", "并行核查"), ("start", "impact", "并行核查"),
+                      ("isolation", "summary", "汇总依据"), ("impact", "summary", "汇总依据"))],
+    }))
+    assert graph.mermaid.startswith("flowchart TD\n")
+    assert len(graph.edges) == 4
+    assert 'start -->|"并行核查"| impact' in graph.mermaid
+    patched = update_investigation_graph(graph, InvestigationGraphUpdate.model_validate({
+        "nodes": [{"node_id": "impact", "status": "recorded"}],
+    }))
+    assert patched.edges == graph.edges
+
+
+def test_existing_disconnected_snapshot_remains_readable_and_patchable():
+    original = InvestigationDiagram.model_validate({
+        "title": "已保存的旧图", "mermaid": 'flowchart TD\n  read["观测"]\n  check["判断"]',
+        "edges": [], "node_refs": [{"node_id": "read", "label": "观测"},
+                                   {"node_id": "check", "label": "判断"}],
+    })
+    patched = update_investigation_graph(original, InvestigationGraphUpdate.model_validate({
+        "nodes": [{"node_id": "check", "status": "recorded"}],
+    }))
+    assert patched.mermaid == original.mermaid
+    assert patched.edges == original.edges == []
+    assert original.node_refs[1].status is None
+
+
+@pytest.mark.anyio
+async def test_publication_rejects_missing_links_and_accepts_a_corrected_graph():
+    task, repo, service = setup_service()
+    first = await publish(service, repo, task["task_id"], nodes=[
+        {"node_id": "read", "label": "观测核验"},
+    ])
+    original = first.task.latest_progress.progress.diagram.model_dump(mode="json")
+    with pytest.raises(DomainError) as error:
+        await publish(service, repo, task["task_id"], nodes=[
+            {"node_id": "check", "label": "异常判断"},
+        ])
+    assert error.value.code == "GRAPH_RELATION_REQUIRED"
+    assert (await service.detail(task["task_id"])).latest_progress.progress.diagram.model_dump(mode="json") == original
+    corrected = await publish(service, repo, task["task_id"], nodes=[
+        {"node_id": "check", "label": "异常判断"},
+    ], edges=[{"source": "read", "target": "check", "label": "核验依据"}])
+    assert 'read -->|"核验依据"| check' in corrected.task.latest_progress.progress.diagram.mermaid
+
+
 def test_first_structured_update_preserves_an_existing_mermaid_without_reinterpreting_it():
     original = InvestigationDiagram.model_validate({
         "title": "原调查路线", "mermaid": 'flowchart LR\nA{核查条件} -->|具备| B[原路径]\nA --> C[另一分支]',