cocolord · Nov 12, 2026 · Nov 12, 2026 · Nov 12, 2026 · Nov 12, 2026
@@ -720,6 +720,37 @@ the existing expansion budget. This qualifies the browser driver's target
 selection; product focus recovery and the causes of earlier partial inventories
 retain their separate evidence requirements.
 
+**Checkpoint 2026-11-12 (source-coverage execution).**
+The source lane now executes the real disposable first-Goal and runtime-readiness
+backends rather than returning packaged-only PASS receipts. Explicitly required
+backend journeys fail before PASS when their readback did not complete. Other
+scenario coverage boundaries remain valid. Existing GET/HEAD socket recovery
+includes EPIPE within the same three-attempt bound; HTTP writes stay single-attempt.
+Research-record search checks bind help dismissal to its current keyboard trigger
+and require the selected query to survive dismissal.
+
+On immutable source `eedfc5e24`, local macOS/Node26 qualification passes all
+86 scenarios across three non-overlapping 29/29/28 partitions with a common
+catalog and source tree, verified backend-readback markers and nonempty merged
+coverage. The passing partitions take 1,943.74 seconds; earlier failed runs,
+setup and repair remain separate records. This is validation cost, with no
+complete model/service/human cost comparison.
+
+The matched 200-item M2 pair retains all 15 samples and long-task windows.
+Ordinary medians are 225.77/172.99/310.77/128.45/146.52 ms, maximum 366.08 ms;
+coverage medians are 328.21/371.20/332.49/174.73/173.02 ms, maximum 421.22 ms.
+Both have zero samples above the 1-second target. Full-suite coverage M2 also
+passes: medians 325.35/270.28/326.59/164.22/166.47 ms, maximum 463.88 ms.
+The real 1,200 ms repeated Board delay is correctly rejected by ordinary
+qualification (slow median 1491.77 ms, maximum 1519.98 ms; all 15
+samples/windows retained). Existing strict-boundary and missing-sample tests
+pass without a budget change.
+
+The failed source partitions and historical frozen failures retain their original
+status. This is named local source qualification, not Linux CI, an installed-host
+upgrade, model task completion, kernel attribution or full-cost acceptance. V10
+thresholds and the remaining real-team/product milestones retain their owners.
+
 **Checkpoint 2026-10-11 (M1 proposed at `6c3c0c104` in [#6191](https://github.com/loopx-project/loopx/pull/6191), subsequently merged).**
 Verified subset: V1 reducer fixtures written from Section 5.3, mutation checks
 and the packaged page on synthetic status; V2 strip and Board sharing the work

@@ -513,6 +513,30 @@ M3、W2/W3、安装态以及 L1/G1 的真实交付和原请求路线返回仍由
 全局 Enter 到达前被重渲染替换。验证仍须在既有展开预算内展示全部 200 项。这项证据
 验证浏览器驱动的目标选择；产品焦点恢复及早期部分清单的成因仍需各自的证据。
 
+**检查点 2026-11-12（源码 coverage 执行）。**
+源码分线现在执行真实的一次性首次 Goal 与运行就绪后端，不再提前返回仅适用于打包
+模式的 PASS 回执。明确需要后端的旅程未完成读回时，在 PASS 前失败；其他场景的
+合理 coverage 边界仍有效。既有 GET/HEAD socket 恢复纳入 EPIPE，仍至多三次；
+HTTP 写请求只尝试一次。
+研究记录搜索检查将提示关闭绑定当前键盘按钮，并要求关闭后保留原查询。
+
+固定源码 `eedfc5e24` 在本地 macOS/Node26 下完成全部 86 场景资格：三个互不重叠的
+29/29/28 分片共享同一完整场景目录和源码 tree，后端读回标记及非空合并 coverage 均已核对。
+本次通过的分片用时 1,943.74 秒；此前失败运行、准备和修复仍分开记录。这是验证成本，
+不构成完整模型、服务和人工成本比较。
+
+同版本 200 项 M2 对照保留全部 15 个样本和长任务窗口。普通模式五组中位数为
+225.77/172.99/310.77/128.45/146.52 ms，最大 366.08 ms；coverage 模式为
+328.21/371.20/332.49/174.73/173.02 ms，最大 421.22 ms，两者超过 1 秒目标的
+样本数均为 0。完整分片中的 coverage M2 也通过：中位数
+325.35/270.28/326.59/164.22/166.47 ms，最大 463.88 ms。真实注入 1,200 ms 的重复看板
+停顿被普通资格正确拒绝：慢组中位数 1491.77 ms、最大 1519.98 ms，全部 15 个
+样本和窗口仍保留。既有严格边界及缺失样本负例通过，预算未变。
+
+源码分片失败及历史冻结失败保留原判定。这是指定本地源码的资格，不证明 Linux CI、
+安装宿主升级、模型任务完成、内核归因或完整成本验收。V10 阈值及剩余真实团队和产品
+里程碑仍由各自既有 owner 验收。
+
 **检查点 2026-10-11（M1 在 [#6191](https://github.com/loopx-project/loopx/pull/6191) 的 `6c3c0c104` 提出，后续已合并）。**
 已验证的子集：V1 的按第 5.3 节编写的 reducer fixture、变异检查及合成状态打包页面；
 V2 的指标条与看板共用工作判定；V4 的真实隔离 Chat 动作服务与注册表 authority 上的停止、

@@ -3468,3 +3468,44 @@ qualifies the consumed replan read cost; full CLI process, durable writeback,
 model and end-to-end outcome costs are not inferred from these timings. The
 complete minimal load/context closure, R1/M3, independent Standard packages,
 workspace/meta publication and installed-host adoption remain open.
+
+
+### 2026-11-12 — Dashboard source acceptance executes its backend journeys
+
+Source coverage now executes the existing first-Goal and runtime-readiness
+journeys through the real disposable Chat service. Both source and packaged
+lanes already prepare the verified Chat bundle. Declared backend journeys
+require successful backend assertions before a PASS receipt; an early return
+cannot qualify them. Scenarios with other legitimate coverage boundaries keep
+their existing behavior. The shared browser proxy adds EPIPE to its existing
+three-attempt GET/HEAD socket recovery, recognizes typed errno or the first
+Playwright socket-error line, and never retries an HTTP write or arbitrary prose.
+Research-record search checks bind help dismissal to its current keyboard trigger
+and require the selected query to survive dismissal.
+
+On immutable source `eedfc5e24`, local macOS/Node26 qualification passes all
+86 scenarios across three non-overlapping 29/29/28 partitions with a common
+catalog and source tree, verified backend-readback markers and nonempty merged
+coverage. The passing partitions take 1,943.74 seconds; earlier failed runs,
+setup and repair remain separate records. This is validation cost, with no
+complete model/service/human cost comparison.
+
+The matched 200-item M2 pair retains all 15 samples and long-task windows.
+Ordinary medians are 225.77/172.99/310.77/128.45/146.52 ms, maximum 366.08 ms;
+coverage medians are 328.21/371.20/332.49/174.73/173.02 ms, maximum 421.22 ms.
+Both have zero samples above the 1-second target. Full-suite coverage M2 also
+passes: medians 325.35/270.28/326.59/164.22/166.47 ms, maximum 463.88 ms.
+The real 1,200 ms repeated Board delay is correctly rejected by ordinary
+qualification (slow median 1491.77 ms, maximum 1519.98 ms; all 15
+samples/windows retained). Existing strict-boundary and missing-sample tests
+pass without a budget change.
+
+Earlier failed source partitions (EPIPE, query loss and temporary-directory
+failure) remain recorded separately, as do historical frozen 1.162-second
+failures and cancelled runs. This repairs
+source-acceptance execution and its bounded transport recovery, with no render
+threshold, kernel runtime or authority change. The first-Goal fixture uses an
+exit-only Agent: registration and honest failed-Turn recovery do not prove
+natural-language task completion or model benefit. Complete minimal closure,
+R1/M3, workspace/meta publication, installed-host adoption and full-cost
+acceptance remain open.
@@ -2706,3 +2706,33 @@ dry-run 拒绝路径的中位数为 142.4→135.1 ms、934.4→911.8 ms。零条
 这些计时推断完整 CLI 进程、持久写回、模型及端到端成果成本。完整最小加载／
 上下文闭包、R1/M3、独立 Standard 包、workspace／meta 发布及安装宿主采用仍未
 完成。
+
+
+### 2026-11-12 — Dashboard 源码验收执行真实后端旅程
+
+源码 coverage 现在通过真实的一次性 Chat 服务执行既有首次 Goal 和运行就绪旅程。
+源码与打包分线均已准备经校验的 Chat bundle。声明需要后端的旅程必须完成真实后端
+断言后才能记录 PASS，提前返回不能满足该条件；其他有合理 coverage 边界的场景
+保留原有行为。共用浏览器代理在既有至多三次 GET/HEAD socket 恢复中纳入 EPIPE，
+仅识别类型化 errno 或 Playwright 首行 socket 错误，不重试 HTTP 写请求或任意诊断文字。
+研究记录搜索检查将提示关闭绑定当前键盘按钮，并要求关闭后保留原查询。
+
+固定源码 `eedfc5e24` 在本地 macOS/Node26 下完成全部 86 场景资格：三个互不重叠的
+29/29/28 分片共享同一完整场景目录和源码 tree，后端读回标记及非空合并 coverage 均已核对。
+本次通过的分片用时 1,943.74 秒；此前失败运行、准备和修复仍分开记录。这是验证成本，
+不构成完整模型、服务和人工成本比较。
+
+同版本 200 项 M2 对照保留全部 15 个样本和长任务窗口。普通模式五组中位数为
+225.77/172.99/310.77/128.45/146.52 ms，最大 366.08 ms；coverage 模式为
+328.21/371.20/332.49/174.73/173.02 ms，最大 421.22 ms，两者超过 1 秒目标的
+样本数均为 0。完整分片中的 coverage M2 也通过：中位数
+325.35/270.28/326.59/164.22/166.47 ms，最大 463.88 ms。真实注入 1,200 ms 的重复看板
+停顿被普通资格正确拒绝：慢组中位数 1491.77 ms、最大 1519.98 ms，全部 15 个
+样本和窗口仍保留。既有严格边界及缺失样本负例通过，预算未变。
+
+此前源码分片的 EPIPE、查询丢失和临时目录失败分别保留记录；历史冻结的 1.162 秒失败
+及已取消运行也保持原判定。
+本项修复源码验收执行与有界传输恢复，不改变渲染阈值、内核运行时或 authority。
+首次 Goal 夹具使用退出式 Agent：登记及如实显示失败 Turn 的恢复不证明自然语言任务
+完成或模型收益。完整最小闭包、R1/M3、workspace/meta 发布、安装宿主采用及完整成本
+验收仍保持开放。
@@ -172,6 +172,12 @@ scenarioCatalog.push(completedHistoryIsolationScenario);
 scenarioCatalog.push(agentModelSelectionScenario);
 export { scenarioCatalog };
 
+export function requireBackendReadback(scenario, result) {
+  if (scenario.requiresBackendReadback && result.backendReadbackVerified !== true) {
+    throw new Error(`Required backend readback did not execute: ${scenario.id}`);
+  }
+}
+
 async function main() {
   const requestedScenario = process.env.LOOPX_PERSONAL_WORKSPACE_SCENARIO;
   const scenarios = requestedScenario
@@ -207,10 +213,12 @@ async function main() {
           ? browser
           : { newPage: (options = {}) => browser.newPage({ locale: "zh-CN", ...options }) };
         const result = await scenario.run({ browser: scenarioBrowser, collectCoverage, url });
+        requireBackendReadback(scenario, result);
         coverageEntries.push(...result.coverageEntries);
         results[scenario.id] = {
           status: "PASS",
           note: result.note,
+          ...(scenario.requiresBackendReadback ? {backend_readback_verified: true} : {}),
           duration_ms: Date.now() - startedAt,
         };
       } catch (error) {

@@ -6,7 +6,7 @@ import { tmpdir } from "node:os";
 import { join, resolve } from "node:path";
 import { createInterface } from "node:readline";
 import { resolveTestPython } from "../../scripts/test-python.mjs";
-import { outputDir, packaged, repoRoot } from "./fixture.mjs";
+import { outputDir, repoRoot } from "./fixture.mjs";
 import { fetchWorkspaceBackend, openWorkspacePage } from "./scenario-context.mjs";
 
 // End to end against the real service: a disposable `loopx chat` on a fresh
@@ -23,9 +23,10 @@ import { fetchWorkspaceBackend, openWorkspacePage } from "./scenario-context.mjs
 
 export const firstGoalHomeLiveScenario = {
   id: "first-goal-home-live",
+  requiresBackendReadback: true,
   async run({ browser, collectCoverage, url }) {
-    // The real service serves the packaged bundle; the development lane has none to serve.
-    if (!packaged) return { coverageEntries: [], note: "live first-Goal home runs against the packaged bundle only" };
+    // Both lanes prepare the verified Chat bundle for the real backend. The
+    // source lane keeps Vite's source maps while proxying this same owner.
     const root = await mkdtemp(join(tmpdir(), "loopx-first-goal-live-"));
     const python = resolveTestPython({ repoRoot });
     const child = spawn(python, ["-u", resolve(repoRoot, "examples/personal-workspace-browser/fresh-install-backend.py"), root], { cwd: repoRoot, env: { ...process.env, PYTHONPATH: repoRoot }, stdio: ["ignore", "pipe", "pipe"] });
@@ -130,7 +131,7 @@ export const firstGoalHomeLiveScenario = {
       await page.locator(".personal-goal-row", { hasText: title }).first().waitFor({ state: "visible", timeout: 20_000 });
       await firstGoal.waitFor({ state: "hidden", timeout: 20_000 });
       await page.screenshot({ path: resolve(outputDir, "first-goal-home-live-home.png"), animations: "disabled" });
-      return { coverageEntries: await session.close(), note: "real loopx chat on a fresh runtime: zero-Goal discovery with its own warning; first Goal registered from the page through the real action owner with the stub Agent's failed first Turn reported honestly; registry written, directory lists it" };
+      return { coverageEntries: await session.close(), backendReadbackVerified: true, note: "real loopx chat on a fresh runtime: zero-Goal discovery with its own warning; first Goal registered from the page through the real action owner with the stub Agent's failed first Turn reported honestly; registry written, directory lists it" };
     } catch (error) {
       if (diagnostics) process.stderr.write(diagnostics);
       await session?.close().catch(() => {});

@@ -8,9 +8,19 @@ import {outputDir, repoRoot} from "./fixture.mjs";
 import {startHistoryOwner} from "./research-history.mjs";
 
 async function readEmptyScope(page, results, message) {
-  await results.getByRole('button',{name:'No matches in loaded records',exact:true}).focus();
+  const search = results.getByRole('searchbox');
+  const query = await search.inputValue();
+  const trigger = results.getByRole('button',{name:'No matches in loaded records',exact:true});
+  await trigger.focus();
+  // Enter opens the help even when the trigger retained focus after an earlier
+  // dismissal; focusing an already focused element does not fire onFocus.
+  await trigger.press('Enter');
   await page.getByRole('tooltip').getByText(message,{exact:true}).waitFor();
-  await page.keyboard.press('Escape');
+  // Bind dismissal to the current trigger after rendering. A global Escape
+  // can reach the search input and clear the filter instead of only the help.
+  await trigger.press('Escape');
+  await page.getByRole('tooltip').waitFor({state:'hidden'});
+  assert.equal(await search.inputValue(), query, 'Dismissing scope help preserves the selected query');
 }
 
 export const researchRecordSearchScenario = {

@@ -7,7 +7,7 @@ import { join, resolve } from "node:path";
 import { createInterface } from "node:readline";
 import { waitForHttp } from "../dashboard-browser-smoke-support.mjs";
 import { resolveTestPython } from "../../scripts/test-python.mjs";
-import { outputDir, packaged, repoRoot } from "./fixture.mjs";
+import { outputDir, repoRoot } from "./fixture.mjs";
 import { fetchWorkspaceBackend, openWorkspacePage } from "./scenario-context.mjs";
 
 // End to end against the real service: a disposable `loopx chat` whose Codex,
@@ -19,8 +19,8 @@ import { fetchWorkspaceBackend, openWorkspacePage } from "./scenario-context.mjs
 
 export const runtimeReadinessLiveScenario = {
   id: "runtime-readiness-live",
+  requiresBackendReadback: true,
   async run({ browser, collectCoverage, url }) {
-    if (!packaged) return { coverageEntries: [], note: "live runtime readiness runs against the packaged bundle only" };
     const root = await mkdtemp(join(tmpdir(), "loopx-readiness-live-"));
     const python = resolveTestPython({ repoRoot });
     const child = spawn(python, ["-u", resolve(repoRoot, "examples/personal-workspace-browser/runtime-readiness-backend.py"), root, "--with-goal"], { cwd: repoRoot, env: { ...process.env, PYTHONPATH: repoRoot }, stdio: ["ignore", "pipe", "pipe"] });
@@ -152,7 +152,7 @@ export const runtimeReadinessLiveScenario = {
       const codex = catalog.adapters.find((row) => row.agent_id === "codex");
       assert.equal(codex.available, true, "The real catalog now reports Codex available");
       assert.equal(codex.unavailable_reason, undefined);
-      return { coverageEntries: await session.close(), note: "real loopx chat service, live PATH probe, real re-check" };
+      return { coverageEntries: await session.close(), backendReadbackVerified: true, note: "real loopx chat service, live PATH probe, real re-check" };
     } catch (error) {
       await session?.page.screenshot({ path: resolve(outputDir, "runtime-readiness-live-failure.png"), animations: "disabled" }).catch(() => {});
       if (session?.errors.length) process.stderr.write(`${JSON.stringify(session.errors)}\n`);

@@ -14,8 +14,13 @@ export async function fetchWorkspaceBackend(route, backendUrl, { pathname } = {}
     try {
       return await route.fetch({ url: target.href, headers });
     } catch (error) {
+      // Playwright preserves the native socket code in its first error line.
+      // Only that transport signature (or a typed errno), never response prose,
+      // permits the existing bounded read retry. Writes remain single-attempt.
+      const socketCode = error?.code ?? /^route\.fetch: (?:connect|read|write) (ECONNRESET|ECONNREFUSED|EPIPE)(?:\s|$)/u
+        .exec(String(error?.message ?? error).split("\n", 1)[0])?.[1];
       const retryable = ["GET", "HEAD"].includes(request.method())
-        && /\b(?:ECONNRESET|ECONNREFUSED)\b/u.test(String(error?.message ?? error))
+        && ["ECONNRESET", "ECONNREFUSED", "EPIPE"].includes(socketCode)
         && attempt < 2;
       if (!retryable) throw error;
       await new Promise(resolve => setTimeout(resolve, 50 * (attempt + 1)));