@xnng/browser-relay 1.6.0

package/docs/benchmarks/browser-runtime.json

@@ -0,0 +1,254 @@
+{
+  "measuredAt": "2026-09-08T11:15:19.408Z",
+  "injectedExtensionRttMs": 0,
+  "platform": "darwin",
+  "node": "v22.23.1",
+  "chromium": "153.0.8010.12",
+  "task": "Search invoice / owner Alice / active only; await visible result",
+  "scope": "Deterministic browser/tool execution; no model inference or token accounting. Playwright over Relay is not Codex Browser Use.",
+  "summary": {
+    "legacy": {
+      "runs": 7,
+      "successes": 7,
+      "elapsedMs": 273.5,
+      "httpCalls": 10,
+      "extensionRoundTrips": 39,
+      "responseBytes": 3521,
+      "snapshotBytes": 2301
+    },
+    "actions": {
+      "runs": 7,
+      "successes": 7,
+      "elapsedMs": 278.7,
+      "httpCalls": 2,
+      "extensionRoundTrips": 2,
+      "responseBytes": 3919,
+      "snapshotBytes": 2202
+    },
+    "playwright-cdp": {
+      "runs": 7,
+      "successes": 7,
+      "elapsedMs": 334.8,
+      "httpCalls": 0,
+      "extensionRoundTrips": 73,
+      "responseBytes": null,
+      "snapshotBytes": 1425
+    }
+  },
+  "delta": {
+    "fullSnapshotBytes": 1880,
+    "changedSnapshotBytes": 160
+  },
+  "samples": [
+    {
+      "variant": "legacy",
+      "iteration": 0,
+      "success": true,
+      "elapsedMs": 292.4,
+      "httpCalls": 10,
+      "extensionRoundTrips": 39,
+      "responseBytes": 3517,
+      "snapshotBytes": 2301
+    },
+    {
+      "variant": "actions",
+      "iteration": 0,
+      "success": true,
+      "elapsedMs": 278.7,
+      "httpCalls": 2,
+      "extensionRoundTrips": 2,
+      "responseBytes": 3923,
+      "snapshotBytes": 2202
+    },
+    {
+      "variant": "playwright-cdp",
+      "iteration": 0,
+      "success": true,
+      "elapsedMs": 293,
+      "httpCalls": 0,
+      "extensionRoundTrips": 73,
+      "responseBytes": null,
+      "snapshotBytes": 1425
+    },
+    {
+      "variant": "legacy",
+      "iteration": 1,
+      "success": true,
+      "elapsedMs": 274.7,
+      "httpCalls": 10,
+      "extensionRoundTrips": 39,
+      "responseBytes": 3517,
+      "snapshotBytes": 2301
+    },
+    {
+      "variant": "actions",
+      "iteration": 1,
+      "success": true,
+      "elapsedMs": 284.1,
+      "httpCalls": 2,
+      "extensionRoundTrips": 2,
+      "responseBytes": 3926,
+      "snapshotBytes": 2202
+    },
+    {
+      "variant": "playwright-cdp",
+      "iteration": 1,
+      "success": true,
+      "elapsedMs": 329.3,
+      "httpCalls": 0,
+      "extensionRoundTrips": 73,
+      "responseBytes": null,
+      "snapshotBytes": 1425
+    },
+    {
+      "variant": "legacy",
+      "iteration": 2,
+      "success": true,
+      "elapsedMs": 291.8,
+      "httpCalls": 10,
+      "extensionRoundTrips": 39,
+      "responseBytes": 3521,
+      "snapshotBytes": 2301
+    },
+    {
+      "variant": "actions",
+      "iteration": 2,
+      "success": true,
+      "elapsedMs": 265.9,
+      "httpCalls": 2,
+      "extensionRoundTrips": 2,
+      "responseBytes": 3918,
+      "snapshotBytes": 2202
+    },
+    {
+      "variant": "playwright-cdp",
+      "iteration": 2,
+      "success": true,
+      "elapsedMs": 338.5,
+      "httpCalls": 0,
+      "extensionRoundTrips": 73,
+      "responseBytes": null,
+      "snapshotBytes": 1425
+    },
+    {
+      "variant": "legacy",
+      "iteration": 3,
+      "success": true,
+      "elapsedMs": 270.2,
+      "httpCalls": 10,
+      "extensionRoundTrips": 39,
+      "responseBytes": 3521,
+      "snapshotBytes": 2301
+    },
+    {
+      "variant": "actions",
+      "iteration": 3,
+      "success": true,
+      "elapsedMs": 288.1,
+      "httpCalls": 2,
+      "extensionRoundTrips": 2,
+      "responseBytes": 3919,
+      "snapshotBytes": 2202
+    },
+    {
+      "variant": "playwright-cdp",
+      "iteration": 3,
+      "success": true,
+      "elapsedMs": 323.5,
+      "httpCalls": 0,
+      "extensionRoundTrips": 73,
+      "responseBytes": null,
+      "snapshotBytes": 1425
+    },
+    {
+      "variant": "legacy",
+      "iteration": 4,
+      "success": true,
+      "elapsedMs": 273.5,
+      "httpCalls": 10,
+      "extensionRoundTrips": 39,
+      "responseBytes": 3521,
+      "snapshotBytes": 2301
+    },
+    {
+      "variant": "actions",
+      "iteration": 4,
+      "success": true,
+      "elapsedMs": 269.7,
+      "httpCalls": 2,
+      "extensionRoundTrips": 2,
+      "responseBytes": 3918,
+      "snapshotBytes": 2202
+    },
+    {
+      "variant": "playwright-cdp",
+      "iteration": 4,
+      "success": true,
+      "elapsedMs": 338.4,
+      "httpCalls": 0,
+      "extensionRoundTrips": 73,
+      "responseBytes": null,
+      "snapshotBytes": 1425
+    },
+    {
+      "variant": "legacy",
+      "iteration": 5,
+      "success": true,
+      "elapsedMs": 269.7,
+      "httpCalls": 10,
+      "extensionRoundTrips": 39,
+      "responseBytes": 3521,
+      "snapshotBytes": 2301
+    },
+    {
+      "variant": "actions",
+      "iteration": 5,
+      "success": true,
+      "elapsedMs": 278.3,
+      "httpCalls": 2,
+      "extensionRoundTrips": 2,
+      "responseBytes": 3918,
+      "snapshotBytes": 2202
+    },
+    {
+      "variant": "playwright-cdp",
+      "iteration": 5,
+      "success": true,
+      "elapsedMs": 334.8,
+      "httpCalls": 0,
+      "extensionRoundTrips": 73,
+      "responseBytes": null,
+      "snapshotBytes": 1425
+    },
+    {
+      "variant": "legacy",
+      "iteration": 6,
+      "success": true,
+      "elapsedMs": 261.4,
+      "httpCalls": 10,
+      "extensionRoundTrips": 39,
+      "responseBytes": 3521,
+      "snapshotBytes": 2301
+    },
+    {
+      "variant": "actions",
+      "iteration": 6,
+      "success": true,
+      "elapsedMs": 280.1,
+      "httpCalls": 2,
+      "extensionRoundTrips": 2,
+      "responseBytes": 3924,
+      "snapshotBytes": 2202
+    },
+    {
+      "variant": "playwright-cdp",
+      "iteration": 6,
+      "success": true,
+      "elapsedMs": 342.5,
+      "httpCalls": 0,
+      "extensionRoundTrips": 73,
+      "responseBytes": null,
+      "snapshotBytes": 1425
+    }
+  ]
+}

package/docs/benchmarks/browser-use-parity.json

@@ -0,0 +1,54 @@
+{
+  "url": "https://zh.omarchy.org/",
+  "capability": {
+    "ok": true,
+    "protocolVersion": 2,
+    "features": [
+      "observe",
+      "read",
+      "ax",
+      "refs",
+      "frames",
+      "shadow-dom",
+      "diff",
+      "actions",
+      "tasks",
+      "coordinates",
+      "drag",
+      "hover",
+      "screenshot-mapping",
+      "scoped-targets",
+      "action-readiness-wait",
+      "aria-check",
+      "tabs",
+      "observation-pages",
+      "full-links",
+      "sessions",
+      "handoff",
+      "focus"
+    ],
+    "maxActions": 100,
+    "extensionVersion": "1.5.0-dev.1",
+    "runtimeId": "0fbd0483-b48f-4ff2-98b0-8589240d99d1"
+  },
+  "elapsedMs": 3792,
+  "pages": 4,
+  "readBytes": 66862,
+  "legacyBytes": 37409,
+  "checks": {
+    "main": true,
+    "legacy": true,
+    "fullLinks": true
+  },
+  "checksRun": {
+    "unit": {
+      "passed": 105,
+      "skipped": 7
+    },
+    "browser": {
+      "passed": 28,
+      "failed": 0
+    }
+  },
+  "scope": "isolated-profile public-page reading verification; not X speed comparison"
+}

package/docs/benchmarks/codex-extension-audit.json

@@ -0,0 +1,131 @@
+{
+  "auditedAt": "2026-09-09",
+  "chromeExtensionVersion": "1.26.901.11451",
+  "desktopPluginVersion": "26.901.51231",
+  "sourceMapsFound": 0,
+  "scope": "Installed published JavaScript bundles and packaged workflow docs, not the complete upstream development source. Compiled/native parts and live server-side deployment are not fully visible. No vendor code copied into project.",
+  "formatting": "Read-only Prettier 3.6.2 copies under /tmp/browser-relay-codex-audit",
+  "sourceFiles": [
+    {
+      "package": "chrome-extension",
+      "path": "manifest.json",
+      "bytes": 2739,
+      "sha256": "43c407d46765d3f85754a516b4500b7f491c067548456ba211fa82fca5bd209b"
+    },
+    {
+      "package": "chrome-extension",
+      "path": "background.js",
+      "bytes": 355753,
+      "sha256": "83b7b7708ab82ecc9aa2ccad20a2167941a706b38a9467dee0bb124d13f0033e"
+    },
+    {
+      "package": "chrome-extension",
+      "path": "content-scripts/codex.js",
+      "bytes": 35143,
+      "sha256": "d01bc3fd8c98fb5c7a42cf6f465a8e27e24d436b2d4304df2ef38c027f54cc70"
+    },
+    {
+      "package": "chrome-extension",
+      "path": "content-scripts/foreign-frame-monitor.js",
+      "bytes": 5158,
+      "sha256": "337ad4ff3ec7846134e5d39528936b1cbb68107de3c4da0e0641c712bda66129"
+    },
+    {
+      "package": "desktop-browser-plugin",
+      "path": "scripts/browser-service.mjs",
+      "bytes": 1308334,
+      "sha256": "953706b76083dd4067958dee7668c4b88beb26122d6262fa60cb79ff7e6abbb1"
+    },
+    {
+      "package": "desktop-browser-plugin",
+      "path": "scripts/browser-client.mjs",
+      "bytes": 150611,
+      "sha256": "b9b9bc2319d5ee6aa0b1e481d63bb2130d28102fc7c9080803ab5552185d9037"
+    },
+    {
+      "package": "desktop-browser-plugin",
+      "path": "docs/accessibility.md",
+      "bytes": 4685,
+      "sha256": "331ebe1105d1c1aa31e54f1e94ab9644032d6e68371b02459dd31bc44618edcb"
+    },
+    {
+      "package": "desktop-browser-plugin",
+      "path": "docs/api-use-behavior.md",
+      "bytes": 3344,
+      "sha256": "9c648a2cbd94ada50bd855c5e9cb1ff2f33f6a2eff5d450d4b310fde336ec256"
+    },
+    {
+      "package": "desktop-browser-plugin",
+      "path": "docs/tab-claiming-chrome.md",
+      "bytes": 2071,
+      "sha256": "58bdb8d36284eb72e7ed8b1099e83acc01880df4cecebbc8a5c34c7f567fcc33"
+    }
+  ],
+  "findings": [
+    {
+      "topic": "tab ownership",
+      "file": "background.js",
+      "symbols": [
+        "claimTab",
+        "getOwningSessionId",
+        "isClaimedBySession",
+        "setTabMark"
+      ],
+      "evidence": "Claims bind tabId to sessionId, turnId and extension instance. Another active session is rejected; handoff/turn marks are tracked."
+    },
+    {
+      "topic": "request lifecycle",
+      "file": "background.js",
+      "symbols": [
+        "startRequest",
+        "finishRequest",
+        "stopSession",
+        "AbortController"
+      ],
+      "evidence": "Tracks active requests, abort signals, running session state and per-tab cursors."
+    },
+    {
+      "topic": "native transport recovery",
+      "file": "background.js",
+      "symbols": [
+        "requestHost",
+        "connectNative",
+        "rejectPendingHostRequests",
+        "scheduleReconnect"
+      ],
+      "evidence": "JSON-RPC request IDs, pending deadlines and explicit rejection on native-port loss; reconnect status is exposed."
+    },
+    {
+      "topic": "heartbeat cleanup",
+      "file": "background.js",
+      "symbols": [
+        "client-heartbeat-alarm",
+        "stopActiveSessions"
+      ],
+      "evidence": "Alarm periodically checks clients and stops sessions/ detaches debuggers if none remain connected."
+    },
+    {
+      "topic": "context and prompts",
+      "file": "docs/accessibility.md / docs/api-use-behavior.md",
+      "symbols": [
+        "AX state",
+        "diff",
+        "current indices",
+        "authoritative signal"
+      ],
+      "evidence": "Accessibility-first workflow, short known action groups, fresh indices, cheapest relevant observation and one meaningful success signal."
+    },
+    {
+      "topic": "native internal branches differ",
+      "file": "scripts/browser-service.mjs",
+      "symbols": [
+        "clickLocator",
+        "performPointerAction",
+        "resolvePointerActionTarget",
+        "focusLocator",
+        "readCheckedState"
+      ],
+      "evidence": "Published service contains separate locator and accessibility paths. Do not assume those routes have identical control/event/occlusion semantics; main results use live CUA AX traces."
+    }
+  ]
+}

package/docs/benchmarks/codex-native-protocol.md

@@ -0,0 +1,69 @@
+# Codex 原生同题测试协议
+
+日期：2026-09-09。已完成原生 AX 单步、AX 批量、原生 Playwright 各 5 次，同题全部成功。
+原始数据见 `codex-native.json`，连接状态与历史诊断见 `codex-native-status.json`。
+本文件保留协议与后续复跑约束；实际执行的范围及偏差见下文。
+
+## 本次执行范围
+
+先进行交互式流程探索，再做确定性回放。正式回放中的“单步观察”表示每个动作后读取状态，
+不表示每步都有一次新的模型决策。AX 两种路径成对交替，各 5 次；原生 Playwright 后续单独跑 5 次。
+开始每题都重载页面并重新取得 AX 索引。表单任务的业务等待由 AX 自带的状态稳定等待或
+Playwright 的结果可见等待满足，没有插入额外固定 sleep。
+
+原生视口 1466 × 925；Relay 已补充同视口的可见 Chromium 对照。原生 Chrome 版本读取被
+URL 策略阻止，记为 unknown，没有绕过。两个浏览器配置与后端不完全相同。
+探索轮的经过时间与回放轮的内层工具时间分开保存；计时闭包问题修正后才开始正式回放。
+15 次所有状态以去重字符串字典保存，每条操作引用对应 stateId，保留逐操作毫秒计时。
+
+[CUA 回放代码](codex-native-replay.js)保存了本次执行方法，整理为统一入口。先启动
+`npm run bench:serve`，在 CUA 用 `cua.createBrowserTab('chrome', fixtureUrl, {sessionName:'🧪 原生同题对比'})`
+得到 `nativeBench`，读取初始状态，再将回放代码放入持久 CUA REPL。它不是普通 Node 脚本，
+也没有导入替代的浏览器驱动。每一对 AX 流程在一个外层工具调用中执行，原始观察在 `samples` 中保留。
+
+## 连接与安装
+
+1. 通过 CUA 的 Chrome 原生应用窗口访问官方扩展商店，安装 ChatGPT 扩展。
+2. 按[官方说明](https://learn.chatgpt.com/docs/chrome-extension)，桌面应用“设置 → Computer use”中启用对应浏览器，直到显示 Manage。
+3. 用 `cua.getState()` 确认浏览器可用，再用返回的浏览器 ID 新建专用测试标签。
+4. 如果 CUA 拒绝操作 Codex 自身设置，由用户完成设置部分。不能通过脚本替代被禁止的操作。
+5. 原生 `getApp('com.google.Chrome')` 与扩展支持的 Browser Tab 是两条不同通路。即使前者成功，也不能将其结果标为后者的性能。
+
+## 同题与验证
+
+启动 `npm run bench:serve`，使用其打印的本机 URL。服务不进行外部业务操作。
+
+统一任务文本：
+
+> 在当前测试页把 Search 填为 invoice，Owner 选为 Alice，勾选 Active only，点击 Search records。等待页面显示 Found: invoice / alice / active，然后停止。使用当前页面状态定位元素。
+
+本次工具比较可使用已知任务，但不能包装成未知网页的模型盲测。若评测模型能力，应另开干净会话，不向模型提供源码、预先写好的定位器或成功轨迹。
+
+## 两种原生流程
+
+- 单步观察：初始 AX；每个会影响下一步选择的动作后读取 AX；看到成功状态后结束。
+- 短批处理：初始 AX；在已知目标和状态允许时把动作放在一个 `js` 调用；遇到下拉菜单等新 UI 时先观察并重新取得索引；最后读取 AX 核验结果。
+
+禁止跨 reload 复用 AX 索引。下拉框不能假定支持 `setValue`；应按实际暴露的能力操作。不加没有依据的固定 sleep，也不以 DOM 写入直接伪造业务结果。首轮发现的限制和恢复步骤必须保留，不能只统计成功重试。
+
+## 计时口径
+
+| 指标 | 定义 |
+| --- | --- |
+| 准备时间 | 创建浏览器、安装/连接、打开测试页、页面重置；单独记录 |
+| 工具执行时间 | 每次 CUA `js` 内用 `Date.now()` 记录 API 调用开始、结束，累加；含该工具内部状态稳定等待 |
+| 会话经过时间 | 初始观察开始到成功状态返回，包含模型决策及调用间调度；不能称为纯工具延迟 |
+| Agent 工具调用数 | 外层 `js` 调用数量 |
+| 浏览器 API 次数 | click、setValue、pressKey、getAXState 等实际数量；与外层调用数分列 |
+| 观察量 | 每次 AX 状态 UTF-8 字节数之和；截图单独记录编码、尺寸及实际可获取的字节数 |
+| 结果 | 精确成功文本、是否纠错、失败阶段、用户接管、超时；失败样本保留 |
+
+至少交替运行两种流程各 5 次。浏览器版本、模式（headed/headless）、视口、设备缩放、模型和推理档位、网络条件必须记录；无法从环境确认的值写 unknown，不猜测。
+
+Relay 已有结果来自 headless Chromium；实际 Chrome 原生通常是 headed，两者环境并不完全一致。公平的性能结论需要补跑相同可见浏览器条件下的 Relay 对照。原生接口没有暴露的 CDP 命令数量记 null，不能写 0，也不能拿 Relay 逻辑命令数与原生外层工具次数比较。
+
+不能给整个 Codex 原生链路套用“Relay→扩展 50 ms”数字：这是 Relay 专属注入点。原生网络链路没有等效注入时，只报告原生默认环境数据。
+
+## 报告要求
+
+保留每次原始观察与计时记录，只归档该测试页内容；不保存用户其他标签或账号信息。分别报告工具时间、会话经过时间、调用数、字节数、成功/失败情况。用中位数和范围呈现小样本，不外推综合模型成功率、token 成本或生产网站胜率。

package/docs/benchmarks/codex-native-replay.js

@@ -0,0 +1,116 @@
+// Paste into mcp__cua_repl after creating the fixture tab as `nativeBench`.
+// This is a CUA REPL snippet, not a standalone Node/Playwright runner.
+// State is kept in one object: do not rebind a log captured by an older cell.
+var nativeReplay = { samples: [], current: null };
+
+nativeReplay.measure = async function (operation, fn) {
+  const start = Date.now();
+  try {
+    const result = await fn();
+    nativeReplay.current.steps.push({
+      operation,
+      elapsedMs: Date.now() - start,
+      success: true,
+      ...(typeof result === "string"
+        ? { state: result, stateBytes: new TextEncoder().encode(result).length }
+        : {}),
+    });
+    return result;
+  } catch (error) {
+    nativeReplay.current.steps.push({
+      operation, elapsedMs: Date.now() - start, success: false, error: String(error),
+    });
+    throw error;
+  }
+};
+
+nativeReplay.observe = async function (full = false) {
+  const state = await nativeReplay.measure("getAXState", () =>
+    nativeBench.getAXState({ emit: false, ...(full ? { disableDiffing: true } : {}) }),
+  );
+  const current = nativeReplay.current;
+  if (!state.includes("The following is a diff")) current.nodes = new Map();
+  for (const removed of state.matchAll(/Removed element IDs:\s*([^\n]+)/g)) {
+    for (const part of removed[1].split(",")) {
+      const [from, to = from] = part.trim().split(/[-–]/).map(Number);
+      for (const id of current.nodes.keys())
+        if (id >= from && id <= to) current.nodes.delete(id);
+    }
+  }
+  for (const line of state.split("\n")) {
+    const match = line.replace(/^[~+]\s*/, "").match(/^\s*(\d+)\s+(.+)$/);
+    if (match) current.nodes.set(Number(match[1]), match[2]);
+  }
+  return state;
+};
+
+nativeReplay.ref = function (pattern) {
+  const matches = [...nativeReplay.current.nodes].filter(([, text]) => pattern.test(text));
+  if (matches.length !== 1) throw new Error(`Expected one current AX target: ${pattern}`);
+  return matches[0][0];
+};
+
+nativeReplay.run = async function (variant, iteration, invocation) {
+  if (!["ax-single", "ax-batch", "native-playwright"].includes(variant))
+    throw new Error("Unknown variant");
+  await nativeBench.reload(); // Intentionally outside measured workflow.
+  nativeReplay.current = { steps: [], nodes: new Map() };
+  const start = Date.now();
+  let success = false, error;
+  try {
+    let finalState;
+    if (variant === "native-playwright") {
+      const pw = nativeBench.playwright;
+      await nativeReplay.measure("domSnapshot", () => pw.domSnapshot());
+      await nativeReplay.measure("fill(Search)", () =>
+        pw.getByRole("textbox", { name: "Search", exact: true }).fill("invoice"));
+      await nativeReplay.measure("selectOption(Owner)", () =>
+        pw.getByRole("combobox", { name: "Owner" }).selectOption({ label: "Alice" }));
+      await nativeReplay.measure("check(Active)", () =>
+        pw.getByRole("checkbox", { name: "Active only" }).check());
+      await nativeReplay.measure("click(Search records)", () =>
+        pw.getByRole("button", { name: "Search records" }).click());
+      await nativeReplay.measure("waitFor(result)", () =>
+        pw.getByText("Found: invoice / alice / active", { exact: true })
+          .waitFor({ state: "visible", timeoutMs: 5000 }));
+      finalState = await nativeReplay.measure("domSnapshot", () => pw.domSnapshot());
+    } else {
+      await nativeReplay.observe(true);
+      await nativeReplay.measure("setValue(Search)", () =>
+        nativeBench.setValue(nativeReplay.ref(/^text field .*Search/), "invoice"));
+      if (variant === "ax-single") await nativeReplay.observe();
+      await nativeReplay.measure("setValue(Owner)", () =>
+        nativeBench.setValue(nativeReplay.ref(/^pop up button .*Owner/), "Alice"));
+      if (variant === "ax-single") await nativeReplay.observe();
+      await nativeReplay.measure("setValue(Active)", () =>
+        nativeBench.setValue(nativeReplay.ref(/^checkbox .*Active only/), "1"));
+      if (variant === "ax-single") await nativeReplay.observe();
+      await nativeReplay.measure("click(Search records)", () =>
+        nativeBench.click(nativeReplay.ref(/^button Search records$/)));
+      finalState = await nativeReplay.observe();
+    }
+    success = finalState.includes("Found: invoice / alice / active");
+    if (!success) throw new Error("Expected success text not visible");
+  } catch (failure) {
+    error = String(failure);
+  }
+  const steps = nativeReplay.current.steps;
+  const sample = {
+    variant, iteration, invocation, success, error,
+    scriptElapsedMs: Date.now() - start,
+    toolApiMs: steps.reduce((sum, step) => sum + step.elapsedMs, 0),
+    stateBytes: steps.reduce((sum, step) => sum + (step.stateBytes || 0), 0),
+    apiCalls: steps.length,
+    observations: steps.filter((step) => step.state != null).length,
+    steps,
+  };
+  nativeReplay.samples.push(sample);
+  return { ...sample, steps: undefined };
+};
+
+// Run each AX pair in its own CUA call, changing i from 0 through 4:
+// const order = i % 2 ? ["ax-batch", "ax-single"] : ["ax-single", "ax-batch"];
+// for (const variant of order)
+//   nodeRepl.write(await nativeReplay.run(variant, i, `pair-${i}`));
+// Then run native-playwright iterations 0–4, retaining invocation IDs.
+// Read raw observations from nativeReplay.samples; export with nodeRepl.write.