examples: add zhaopin AI job scrapers + SQLite analysis

- scrape_zhaopin_ai.py: lightweight no-login list scraper
- scrape_zhaopin_full.py: SQLite storage, resumable crawl, detail-page JD
- analyze_zhaopin.py: stats by city/salary/education/experience/skills
- scrape_zhipin_ai.py: BOSS Zhipin variant (login-based, fallback)
- gitignore scraper data artifacts and browser profile

.github/workflows/release.yml

@@ -104,6 +104,24 @@ jobs:
           ref: ${{ env.SOURCE_REF }}
           fetch-depth: 1
 
+      # Record which invisible_firefox commit this build came from. The publish
+      # job turns the range previous-release..this commit into the release notes
+      # (scripts/gen_release_notes.py), and re-publishes it as a source-commit.txt
+      # asset so the NEXT release knows where to start the changelog. One leg is
+      # enough — all legs check out the same SOURCE_REF.
+      - name: Record source commit (for auto release notes)
+        if: matrix.leg == 'linux-x86_64'
+        shell: bash
+        run: git rev-parse HEAD > source-commit.txt && cat source-commit.txt
+      - name: Upload source-commit artifact
+        if: matrix.leg == 'linux-x86_64'
+        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4
+        with:
+          name: source-commit
+          path: source-commit.txt
+          if-no-files-found: error
+          retention-days: 7
+
       - name: Set up Python
         uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5
         with: { python-version: '3.11' }
@@ -344,18 +362,23 @@ jobs:
       # CLOAK + WEBGL-MASKING GUARDS — run the wrapper's e2e cloak/gamma checks
       # against THIS leg's freshly-built artifact, on its native runner. The
       # wrapper's headless=True is headed+hidden (cloak on Win/macOS, its own
-      # Xvfb on Linux), so software-GL rendering works on the GPU-less hosts.
-      # test_cloak asserts the window is hidden (Windows DWMWA_CLOAKED / macOS
-      # CGWindowAlpha) AND still renders — the macOS leg is the only place the
-      # cocoa cloak patch gets RUN. The webgl guard catches a regression of the
-      # gamma readPixels noise back to the pixelscan-maskable ±1 spike form.
+      # Xvfb on Linux). Linux (Xvfb + llvmpipe) and Windows (WARP) give a
+      # software WebGL context on the GPU-less hosts, so the WebGL-dependent
+      # assertions run there. macOS GitHub runners expose NO WebGL in the CI
+      # session at all (even vanilla Firefox; macOS has no software-GL fallback),
+      # so on the mac legs the WebGL checks self-skip and the cloak is validated
+      # via its non-blank screenshot + CGWindowAlpha == 0. test_cloak asserts the
+      # window is hidden (Windows DWMWA_CLOAKED / macOS CGWindowAlpha) AND still
+      # renders — the macOS leg is the only place the cocoa cloak patch gets RUN.
+      # The webgl guard catches a regression of the gamma readPixels noise back to
+      # the pixelscan-maskable ±1 spike form (covered on Linux + Windows).
       - name: Install pyobjc Quartz (macOS — to read the cloak window alpha)
         if: matrix.kind == 'mac'
         run: python -m pip install --quiet pyobjc-framework-Quartz
       - name: Cloak + WebGL-masking guards (headed)
         shell: bash
         run: |
-          python -m pip install --quiet -e .
+          python -m pip install --quiet ".[dev]"
           INVPW_BINARY_PATH="$FF_EXE" python -m pytest \
             tests/test_cloak.py \
             "tests/test_fingerprint_surface.py::test_webgl_readpixels_no_masking_signature" \
@@ -368,9 +391,18 @@ jobs:
     permissions:
       contents: write
     steps:
+      - name: Checkout wrapper (for scripts/gen_release_notes.py)
+        uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4
+        with: { fetch-depth: 1 }
+      - name: Set up Python
+        uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5
+        with: { python-version: '3.11' }
       - name: Download all build assets
         uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4
         with: { pattern: asset-*, path: dl, merge-multiple: true }
+      - name: Download source-commit metadata
+        uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4
+        with: { name: source-commit, path: src-meta }
       - name: Assert all 5 target archives present (no silent partial release)
         run: |
           cd dl
@@ -397,9 +429,38 @@ jobs:
           TAG="${{ github.event.inputs.release_tag }}"
           [ -z "$TAG" ] && TAG="${GITHUB_REF_NAME}"
           echo "tag=$TAG" >> "$GITHUB_OUTPUT"
-          # bare revision number for the release title: firefox-9 -> 9
-          echo "num=${TAG#firefox-}" >> "$GITHUB_OUTPUT"
+          # bare revision number for the release title: firefox-10 -> 10
+          N="${TAG#firefox-}"
+          echo "num=$N" >> "$GITHUB_OUTPUT"
+          # previous release tag, for the changelog range (firefox-10 -> firefox-9)
+          case "$N" in (*[!0-9]*|'') echo "prevtag=" >> "$GITHUB_OUTPUT";;
+                        (*) echo "prevtag=firefox-$((N-1))" >> "$GITHUB_OUTPUT";; esac
           echo "publishing DRAFT release for tag: $TAG"
+      - name: Build release notes from the source commits
+        id: notes
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          set -e
+          CUR="$(cat src-meta/source-commit.txt 2>/dev/null | tr -d '[:space:]')"
+          echo "this build's source commit: ${CUR:-<none>}"
+          # previous release's recorded source commit — gives the changelog range.
+          # Missing (first automated notes / firefox-0) -> notes omit the changelog.
+          PREV=""
+          PREVTAG="${{ steps.tag.outputs.prevtag }}"
+          if [ -n "$PREVTAG" ] && gh release download "$PREVTAG" -R "${{ github.repository }}" \
+               --pattern source-commit.txt --dir prev 2>/dev/null; then
+            PREV="$(cat prev/source-commit.txt | tr -d '[:space:]')"
+            echo "previous ($PREVTAG) source commit: $PREV"
+          else
+            echo "no previous source-commit.txt — changelog section omitted this time"
+          fi
+          python scripts/gen_release_notes.py --tag "${{ steps.tag.outputs.tag }}" \
+            --current "$CUR" --prev-sha "$PREV" --source-repo "${{ env.SOURCE_REPO }}" > body.md
+          echo "----- generated body.md -----"; cat body.md
+          # publish THIS build's source commit so the next release can diff from it
+          cp src-meta/source-commit.txt dl/source-commit.txt
       - name: Create DRAFT release with all assets
         uses: softprops/action-gh-release@3bb12739c298aeb8a4eeaf626c5b8d85266b0e65 # v2
         with:
@@ -412,13 +473,7 @@ jobs:
             dl/*.tar.gz
             dl/*.zip
             dl/checksums.txt
-          body: |
-            Patched Firefox 150.0.1 — built on GitHub Actions ($0, no mold).
-            Targets: linux-x86_64, linux-arm64, win-x86_64, macos-arm64, macos-x86_64.
-
-            DRAFT — do not publish until validate_release.py + realness gate pass on all archives.
-
-            macOS: ad-hoc signed (not notarized). After download run:
-              xattr -dr com.apple.quarantine Firefox.app
+            dl/source-commit.txt
+          body_path: body.md
         env:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}

.github/workflows/verify-cloak.yml

@@ -0,0 +1,103 @@
+# ─────────────────────────────────────────────────────────────────────────────
+# verify-cloak.yml — re-runnable CLOAK + WEBGL-MASKING GUARDS for an EXISTING
+# build run's artifacts, WITHOUT rebuilding Firefox (~3h on the mac legs).
+#
+# release.yml runs these same guards in its `gate` job against each freshly-built
+# artifact. This re-runs them against the artifacts of a PRIOR build run (input
+# `run_id`) using the CURRENT wrapper code on the default branch — so a test-only
+# fix (e.g. making the macOS leg tolerant of the runner's missing WebGL) can be
+# validated against the real binaries in ~10 min instead of paying a full rebuild.
+#
+# Same guard command as release.yml's gate. Headed-but-cloaked; zero proxy / zero
+# secrets. The macOS legs are the only place the cocoa cloak patch actually RUNS.
+# ─────────────────────────────────────────────────────────────────────────────
+name: verify-cloak
+
+on:
+  workflow_dispatch:
+    inputs:
+      run_id:
+        description: 'build run id whose asset-* artifacts to re-gate (e.g. 27346856197)'
+        required: true
+
+permissions:
+  contents: read
+  actions: read   # download-artifact needs this to read another run's artifacts
+
+jobs:
+  guard:
+    name: guard-${{ matrix.leg }}
+    runs-on: ${{ matrix.runner }}
+    timeout-minutes: 25
+    strategy:
+      fail-fast: false
+      matrix:
+        # Same legs/runners/assets as release.yml's gate matrix.
+        include:
+          - leg: linux-x86_64
+            runner: ubuntu-24.04
+            kind: linux
+            asset: firefox-150.0.1-stealth-linux-x86_64.tar.gz
+          - leg: linux-arm64
+            runner: ubuntu-24.04-arm
+            kind: linux
+            asset: firefox-150.0.1-stealth-linux-arm64.tar.gz
+          - leg: win-x86_64
+            runner: windows-latest
+            kind: win
+            asset: firefox-150.0.1-stealth-win-x86_64.zip
+          - leg: macos-arm64
+            runner: macos-15
+            kind: mac
+            asset: firefox-150.0.1-stealth-macos-arm64.tar.gz
+          - leg: macos-x86_64
+            runner: macos-15-intel
+            kind: mac
+            asset: firefox-150.0.1-stealth-macos-x86_64.tar.gz
+    steps:
+      - name: Checkout wrapper (current default branch — the FIXED tests)
+        uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4
+        with: { fetch-depth: 1 }
+      - name: Download build asset from the prior run (no rebuild)
+        uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4
+        with:
+          name: asset-${{ matrix.leg }}
+          path: art
+          run-id: ${{ github.event.inputs.run_id }}
+          github-token: ${{ secrets.GITHUB_TOKEN }}
+      - name: Set up Python
+        uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5
+        with: { python-version: '3.11' }
+      - name: Install Playwright driver (no bundled browser — we override executable_path)
+        # Single-source pin (see release.yml); the wrapper enforces juggler compat.
+        shell: bash
+        run: python -m pip install --quiet "playwright==$(cat scripts/playwright_pin.txt)"
+      - name: Linux system deps for headless firefox
+        if: matrix.kind == 'linux'
+        run: sudo "$(which python)" -m playwright install-deps firefox
+      - name: Extract + locate firefox binary
+        shell: bash
+        run: |
+          set -e
+          mkdir -p ff
+          A="art/${{ matrix.asset }}"
+          case "${{ matrix.kind }}" in
+            win)   python -c "import zipfile; zipfile.ZipFile('$A').extractall('ff')"; EXE="ff/firefox.exe";;
+            linux) tar xzf "$A" -C ff; EXE="ff/firefox";;
+            mac)   tar xzf "$A" -C ff; EXE="ff/Firefox.app/Contents/MacOS/firefox";;
+          esac
+          [ -e "$EXE" ] || { echo "ERROR: firefox binary not found at $EXE"; exit 1; }
+          chmod +x "$EXE" 2>/dev/null || true
+          echo "FF_EXE=$EXE" >> "$GITHUB_ENV"
+          echo "located: $EXE"
+      - name: Install pyobjc Quartz (macOS — to read the cloak window alpha)
+        if: matrix.kind == 'mac'
+        run: python -m pip install --quiet pyobjc-framework-Quartz
+      - name: Cloak + WebGL-masking guards (headed)
+        shell: bash
+        run: |
+          python -m pip install --quiet ".[dev]"
+          INVPW_BINARY_PATH="$FF_EXE" python -m pytest \
+            tests/test_cloak.py \
+            "tests/test_fingerprint_surface.py::test_webgl_readpixels_no_masking_signature" \
+            -m e2e -o addopts='' -q

.gitignore

@@ -6,3 +6,9 @@ build/
 .pytest_cache/
 .venv/
 firefox-source/
+# scraper runtime artifacts (examples/) — data outputs & browser profiles
+*.db
+zhaopin_*.json
+zhaopin_*.csv
+ai_jobs.csv
+examples/.zhipin_profile/

examples/analyze_zhaopin.py

@@ -0,0 +1,169 @@
+"""智联招聘抓取数据分析 —— 读取 SQLite, 按城市/薪资/学历/经验/技能做统计。
+
+纯标准库, 无需 pandas。
+
+用法:
+  python examples/analyze_zhaopin.py                      # 全量分析
+  python examples/analyze_zhaopin.py --keyword AI         # 只看某关键词
+  python examples/analyze_zhaopin.py --city 北京          # 只看某城市
+  python examples/analyze_zhaopin.py --top 15             # 排行榜取前 15
+  python examples/analyze_zhaopin.py --db zhaopin_jobs.db
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import re
+import sqlite3
+import statistics
+from collections import Counter
+
+WORKDAYS_PER_MONTH = 21.75  # 日薪 -> 月薪折算
+
+
+def parse_salary(s: str) -> tuple[int, int] | None:
+    """把薪资字符串解析成 (月薪下限, 月薪上限) 元。无法解析返回 None。"""
+    if not s or "面议" in s:
+        return None
+    s = s.strip()
+    is_daily = "/天" in s or "元/天" in s
+    is_wan = "万" in s
+    nums = re.findall(r"\d+(?:\.\d+)?", s)
+    if not nums:
+        return None
+    vals = [float(x) for x in nums[:2]]
+    if len(vals) == 1:
+        vals = [vals[0], vals[0]]
+    lo, hi = vals[0], vals[1]
+    if is_wan:
+        lo, hi = lo * 10000, hi * 10000
+    if is_daily:
+        lo, hi = lo * WORKDAYS_PER_MONTH, hi * WORKDAYS_PER_MONTH
+    return int(lo), int(hi)
+
+
+def city_of(location: str) -> str:
+    return (location or "").split("·")[0].strip() or "未知"
+
+
+SALARY_BUCKETS = [
+    (0, 5000, "<5k"),
+    (5000, 8000, "5-8k"),
+    (8000, 12000, "8-12k"),
+    (12000, 18000, "12-18k"),
+    (18000, 25000, "18-25k"),
+    (25000, 10**9, ">=25k"),
+]
+
+
+def bucket_of(mid: float) -> str:
+    for lo, hi, label in SALARY_BUCKETS:
+        if lo <= mid < hi:
+            return label
+    return "?"
+
+
+def bar(n: int, maxn: int, width: int = 30) -> str:
+    if maxn <= 0:
+        return ""
+    return "█" * max(1, round(n / maxn * width))
+
+
+def section(title: str) -> None:
+    print("\n" + "=" * 56)
+    print(title)
+    print("=" * 56)
+
+
+def rank_table(counter: Counter, top: int, label: str) -> None:
+    if not counter:
+        print("  (无数据)")
+        return
+    maxn = counter.most_common(1)[0][1]
+    for name, n in counter.most_common(top):
+        print(f"  {name[:18]:<18} {n:>4}  {bar(n, maxn)}")
+
+
+def main() -> None:
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--db", default="zhaopin_jobs.db")
+    ap.add_argument("--keyword", default=None, help="按关键词过滤")
+    ap.add_argument("--city", default=None, help="按城市过滤")
+    ap.add_argument("--top", type=int, default=12, help="排行榜条数")
+    args = ap.parse_args()
+
+    conn = sqlite3.connect(args.db)
+    q = ("SELECT title,salary,experience,education,location,company,"
+         "job_tags,company_tags,skills FROM jobs")
+    params: list = []
+    if args.keyword:
+        q += " WHERE keyword=?"
+        params.append(args.keyword)
+    rows = conn.execute(q, params).fetchall()
+
+    if args.city:
+        rows = [r for r in rows if city_of(r[4]) == args.city]
+
+    if not rows:
+        print("没有匹配的数据。先用 scrape_zhaopin_full.py 抓一些, 或检查过滤条件。")
+        return
+
+    n_total = len(rows)
+    print(f"分析样本: {n_total} 条"
+          + (f" | 关键词={args.keyword}" if args.keyword else "")
+          + (f" | 城市={args.city}" if args.city else ""))
+
+    cities, edus, exps = Counter(), Counter(), Counter()
+    skill_freq, salary_buckets = Counter(), Counter()
+    mids: list[float] = []
+    n_salary_parsed = 0
+
+    for (title, salary, exp, edu, loc, comp, jtags, ctags, skills) in rows:
+        cities[city_of(loc)] += 1
+        edus[(edu or "未知").strip() or "未知"] += 1
+        exps[(exp or "未知").strip() or "未知"] += 1
+        # 技能: 合并 skills + job_tags
+        for src in (skills, jtags):
+            try:
+                for t in json.loads(src or "[]"):
+                    t = t.strip()
+                    if t:
+                        skill_freq[t] += 1
+            except Exception:
+                pass
+        rng = parse_salary(salary)
+        if rng:
+            n_salary_parsed += 1
+            mid = (rng[0] + rng[1]) / 2
+            mids.append(mid)
+            salary_buckets[bucket_of(mid)] += 1
+
+    section("城市分布 (Top)")
+    rank_table(cities, args.top, "城市")
+
+    section("薪资区间分布 (按月薪中位点)")
+    if mids:
+        order = {label: i for i, (_, _, label) in enumerate(SALARY_BUCKETS)}
+        maxn = max(salary_buckets.values())
+        for label in sorted(salary_buckets, key=lambda x: order.get(x, 99)):
+            n = salary_buckets[label]
+            print(f"  {label:<8} {n:>4}  {bar(n, maxn)}")
+        print(f"\n  可解析薪资: {n_salary_parsed}/{n_total} 条 (面议等已排除)")
+        print(f"  月薪中位点  平均: {statistics.mean(mids):>8,.0f} 元")
+        print(f"            中位数: {statistics.median(mids):>8,.0f} 元")
+        print(f"          最低/最高: {min(mids):,.0f} / {max(mids):,.0f} 元")
+    else:
+        print("  (无可解析薪资)")
+
+    section("学历要求")
+    rank_table(edus, args.top, "学历")
+
+    section("经验要求")
+    rank_table(exps, args.top, "经验")
+
+    section("技能/标签词频 (Top)")
+    rank_table(skill_freq, args.top, "技能")
+
+
+if __name__ == "__main__":
+    main()

examples/scrape_zhaopin_ai.py

@@ -0,0 +1,120 @@
+"""抓取智联招聘「AI 相关」岗位 —— 免登录, 渲染搜索结果页直接解析。
+
+合规提醒:
+  - 仅抓取公开展示的岗位标题/薪资/公司/标签等字段。
+  - 不抓取招聘者个人联系方式; 低频请求, 遵守目标站点服务条款, 风险自负。
+
+用法:
+  python examples/scrape_zhaopin_ai.py                         # 默认抓 "AI" 前 3 页
+  python examples/scrape_zhaopin_ai.py --keyword 大模型 --pages 5
+  python examples/scrape_zhaopin_ai.py --keyword AI --jl 530   # 530=北京
+  python examples/scrape_zhaopin_ai.py --headful               # 显示浏览器窗口
+
+城市编码(jl): 北京530 上海538 广州763 深圳765 杭州653 成都801 武汉736 南京635
+"""
+from __future__ import annotations
+
+import argparse
+import csv
+import json
+import random
+import time
+from urllib.parse import quote
+
+from invisible_playwright import InvisiblePlaywright
+
+# 在渲染后的页面里一次性提取所有卡片, 比逐个 query 更快更稳。
+# 选择器来自对真实页面的探测 (见 _probe_zhaopin.py)。
+_EXTRACT_JS = r"""
+() => {
+  const txt = (el) => el ? el.innerText.trim() : "";
+  const cards = Array.from(document.querySelectorAll(".joblist-box__item"));
+  return cards.map(card => {
+    const nameA = card.querySelector("a.jobinfo__name");
+    const info  = Array.from(card.querySelectorAll(".jobinfo__other-info-item"))
+                       .map(e => e.innerText.trim());
+    // 第一个 other-info 是地点(带图标), 取其 span; 其余按顺序是经验/学历
+    const locSpan = card.querySelector(".jobinfo__other-info-item span");
+    const companyA = card.querySelector("a.companyinfo__name");
+    return {
+      title:       txt(nameA),
+      link:        nameA ? nameA.href : "",
+      salary:      txt(card.querySelector(".jobinfo__salary")),
+      job_tags:    Array.from(card.querySelectorAll(".jobinfo__tag .joblist-box__item-tag"))
+                        .map(e => e.innerText.trim()),
+      location:    locSpan ? locSpan.innerText.trim() : (info[0] || ""),
+      experience:  info[1] || "",
+      education:   info[2] || "",
+      company:     companyA ? (companyA.getAttribute("title") || companyA.innerText).trim()
+                            : "",
+      company_url: companyA ? companyA.href : "",
+      company_tags: Array.from(card.querySelectorAll(".companyinfo__tag .joblist-box__item-tag"))
+                         .map(e => e.innerText.trim()),
+    };
+  });
+}
+"""
+
+
+def scrape(keyword: str, pages: int, jl: str | None, headful: bool) -> list[dict]:
+    results: list[dict] = []
+    seen: set[str] = set()
+    with InvisiblePlaywright(seed=42, headless=not headful) as browser:
+        page = browser.new_page()
+        for n in range(1, pages + 1):
+            url = f"https://sou.zhaopin.com/?kw={quote(keyword)}&p={n}"
+            if jl:
+                url += f"&jl={jl}"
+            try:
+                page.goto(url, wait_until="domcontentloaded", timeout=60000)
+                page.wait_for_selector(".joblist-box__item", timeout=20000)
+            except Exception:
+                print(f"第 {n} 页未加载出岗位列表, 跳过 (可能触发风控)。")
+                continue
+            page.wait_for_timeout(1500)  # 让懒加载内容补齐
+
+            rows = page.evaluate(_EXTRACT_JS)
+            new = 0
+            for r in rows:
+                key = r.get("link") or (r.get("title", "") + r.get("company", ""))
+                if key and key not in seen:
+                    seen.add(key)
+                    results.append(r)
+                    new += 1
+            print(f"第 {n} 页: 抓到 {len(rows)} 条 (新增 {new})")
+
+            time.sleep(random.uniform(3, 7))  # 低频, 降低风控
+    return results
+
+
+def save(rows: list[dict], stem: str) -> None:
+    with open(f"{stem}.json", "w", encoding="utf-8") as f:
+        json.dump(rows, f, ensure_ascii=False, indent=2)
+    if rows:
+        fields = ["title", "salary", "experience", "education", "location",
+                  "company", "job_tags", "company_tags", "link", "company_url"]
+        with open(f"{stem}.csv", "w", newline="", encoding="utf-8-sig") as f:
+            w = csv.DictWriter(f, fieldnames=fields, extrasaction="ignore")
+            w.writeheader()
+            for r in rows:
+                row = dict(r)
+                row["job_tags"] = " / ".join(row.get("job_tags") or [])
+                row["company_tags"] = " / ".join(row.get("company_tags") or [])
+                w.writerow(row)
+    print(f"\n已保存 {len(rows)} 条 → {stem}.json / {stem}.csv")
+
+
+def main() -> None:
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--keyword", default="AI", help="搜索关键词")
+    ap.add_argument("--pages", type=int, default=3, help="抓取页数")
+    ap.add_argument("--jl", default=None, help="城市编码, 不填=全国")
+    ap.add_argument("--headful", action="store_true", help="显示浏览器窗口")
+    args = ap.parse_args()
+
+    rows = scrape(args.keyword, args.pages, args.jl, args.headful)
+    save(rows, stem=f"zhaopin_{args.keyword}")
+
+
+if __name__ == "__main__":
+    main()

examples/scrape_zhaopin_full.py

@@ -0,0 +1,277 @@
+"""智联招聘 AI 岗位爬虫（完整版）—— 免登录 + SQLite + 断点续抓 + 详情页 JD 全文。
+
+特性:
+  * 写入 SQLite (INSERT OR IGNORE 去重, 每条提交, 中断不丢数据)
+  * 断点续抓:
+      - 列表阶段: 记录每个 (keyword, jl) 已抓到的最大页码, 重跑从下一页继续
+      - 详情阶段: 只抓还没有 JD 的岗位, 重跑自动补齐
+  * 详情页: 进入每个岗位页面抓 职位描述(JD)全文 + 技能标签
+
+合规提醒: 仅抓公开岗位字段, 不抓招聘者个人联系方式; 低频自用, 遵守站点条款。
+
+用法:
+  # 抓列表(前5页) + 进详情页抓 JD
+  python examples/scrape_zhaopin_full.py --keyword AI --pages 5
+
+  # 只抓列表, 不进详情
+  python examples/scrape_zhaopin_full.py --keyword 大模型 --pages 5 --no-detail
+
+  # 中断后直接重跑同一命令 => 自动从断点继续
+  python examples/scrape_zhaopin_full.py --keyword AI --pages 5
+
+  # 导出已抓数据到 CSV
+  python examples/scrape_zhaopin_full.py --export ai_jobs.csv
+
+  # 重置某关键词的列表进度(重新从第1页抓)
+  python examples/scrape_zhaopin_full.py --keyword AI --reset
+"""
+from __future__ import annotations
+
+import argparse
+import csv
+import json
+import random
+import re
+import sqlite3
+import time
+from contextlib import closing
+from urllib.parse import quote
+
+from invisible_playwright import InvisiblePlaywright
+
+DB_DEFAULT = "zhaopin_jobs.db"
+
+_LIST_JS = r"""
+() => {
+  const txt = (el) => el ? el.innerText.trim() : "";
+  return Array.from(document.querySelectorAll(".joblist-box__item")).map(card => {
+    const nameA = card.querySelector("a.jobinfo__name");
+    const info  = Array.from(card.querySelectorAll(".jobinfo__other-info-item"))
+                       .map(e => e.innerText.trim());
+    const locSpan = card.querySelector(".jobinfo__other-info-item span");
+    const companyA = card.querySelector("a.companyinfo__name");
+    return {
+      title:   txt(nameA),
+      link:    nameA ? nameA.href : "",
+      salary:  txt(card.querySelector(".jobinfo__salary")),
+      job_tags: Array.from(card.querySelectorAll(".jobinfo__tag .joblist-box__item-tag"))
+                     .map(e => e.innerText.trim()),
+      location: locSpan ? locSpan.innerText.trim() : (info[0] || ""),
+      experience: info[1] || "",
+      education:  info[2] || "",
+      company: companyA ? (companyA.getAttribute("title") || companyA.innerText).trim() : "",
+      company_url: companyA ? companyA.href : "",
+      company_tags: Array.from(card.querySelectorAll(".companyinfo__tag .joblist-box__item-tag"))
+                         .map(e => e.innerText.trim()),
+    };
+  });
+}
+"""
+
+_DETAIL_JS = r"""
+() => {
+  const c = document.querySelector(".describtion-card__detail-content");
+  const skills = Array.from(document.querySelectorAll(".describtion-card__skills-item"))
+                      .map(e => e.innerText.trim());
+  return { jd: c ? c.innerText.trim() : "", skills };
+}
+"""
+
+
+# ── DB ──────────────────────────────────────────────────────────────────
+def init_db(path: str) -> sqlite3.Connection:
+    conn = sqlite3.connect(path)
+    conn.execute("""
+        CREATE TABLE IF NOT EXISTS jobs (
+            job_id       TEXT PRIMARY KEY,
+            title        TEXT,
+            salary       TEXT,
+            experience   TEXT,
+            education    TEXT,
+            location     TEXT,
+            company      TEXT,
+            company_url  TEXT,
+            job_tags     TEXT,
+            company_tags TEXT,
+            link         TEXT,
+            jd_text      TEXT,
+            skills       TEXT,
+            keyword      TEXT,
+            created_at   TEXT DEFAULT (datetime('now','localtime')),
+            detail_at    TEXT
+        )
+    """)
+    conn.execute("""
+        CREATE TABLE IF NOT EXISTS progress (
+            scope     TEXT PRIMARY KEY,   -- f"{keyword}|{jl}"
+            last_page INTEGER DEFAULT 0
+        )
+    """)
+    conn.commit()
+    return conn
+
+
+def job_id_from_link(link: str) -> str:
+    m = re.search(r"/jobdetail/([^.?/]+)\.htm", link)
+    return m.group(1) if m else link
+
+
+def get_last_page(conn: sqlite3.Connection, scope: str) -> int:
+    row = conn.execute("SELECT last_page FROM progress WHERE scope=?", (scope,)).fetchone()
+    return row[0] if row else 0
+
+
+def set_last_page(conn: sqlite3.Connection, scope: str, page: int) -> None:
+    conn.execute(
+        "INSERT INTO progress(scope,last_page) VALUES(?,?) "
+        "ON CONFLICT(scope) DO UPDATE SET last_page=excluded.last_page",
+        (scope, page),
+    )
+    conn.commit()
+
+
+def upsert_job(conn: sqlite3.Connection, r: dict, keyword: str) -> bool:
+    jid = job_id_from_link(r.get("link", ""))
+    if not jid:
+        return False
+    cur = conn.execute(
+        """INSERT OR IGNORE INTO jobs
+           (job_id,title,salary,experience,education,location,company,
+            company_url,job_tags,company_tags,link,keyword)
+           VALUES(?,?,?,?,?,?,?,?,?,?,?,?)""",
+        (jid, r.get("title"), r.get("salary"), r.get("experience"),
+         r.get("education"), r.get("location"), r.get("company"),
+         r.get("company_url"),
+         json.dumps(r.get("job_tags") or [], ensure_ascii=False),
+         json.dumps(r.get("company_tags") or [], ensure_ascii=False),
+         r.get("link"), keyword),
+    )
+    conn.commit()
+    return cur.rowcount > 0
+
+
+# ── 抓取 ────────────────────────────────────────────────────────────────
+def crawl_list(conn, page, keyword: str, pages: int, jl: str | None) -> None:
+    scope = f"{keyword}|{jl or ''}"
+    start = get_last_page(conn, scope) + 1
+    if start > pages:
+        print(f"[列表] '{keyword}' 已抓到第 {start-1} 页, 目标 {pages} 页, 无需续抓。")
+        return
+    print(f"[列表] '{keyword}' 从第 {start} 页抓到第 {pages} 页")
+    for n in range(start, pages + 1):
+        url = f"https://sou.zhaopin.com/?kw={quote(keyword)}&p={n}"
+        if jl:
+            url += f"&jl={jl}"
+        try:
+            page.goto(url, wait_until="domcontentloaded", timeout=60000)
+            page.wait_for_selector(".joblist-box__item", timeout=20000)
+        except Exception:
+            print(f"  第 {n} 页未加载出列表, 停止本轮 (重跑可从此页续)。")
+            break
+        page.wait_for_timeout(1200)
+        rows = page.evaluate(_LIST_JS)
+        new = sum(upsert_job(conn, r, keyword) for r in rows)
+        set_last_page(conn, scope, n)
+        print(f"  第 {n} 页: {len(rows)} 条 (新增 {new})")
+        time.sleep(random.uniform(3, 7))
+
+
+def crawl_details(conn, page, keyword: str | None, limit: int | None) -> None:
+    q = "SELECT job_id,link FROM jobs WHERE (jd_text IS NULL OR jd_text='') AND link!=''"
+    params: list = []
+    if keyword:
+        q += " AND keyword=?"
+        params.append(keyword)
+    q += " ORDER BY created_at"
+    if limit:
+        q += f" LIMIT {int(limit)}"
+    todo = conn.execute(q, params).fetchall()
+    if not todo:
+        print("[详情] 没有待补充 JD 的岗位。")
+        return
+    print(f"[详情] 待抓 JD: {len(todo)} 条")
+    for i, (jid, link) in enumerate(todo, 1):
+        try:
+            page.goto(link, wait_until="domcontentloaded", timeout=60000)
+            page.wait_for_selector(".describtion-card__detail-content", timeout=15000)
+            page.wait_for_timeout(800)
+            data = page.evaluate(_DETAIL_JS)
+        except Exception as e:
+            print(f"  [{i}/{len(todo)}] {jid} 抓取失败: {str(e)[:60]}")
+            continue
+        conn.execute(
+            "UPDATE jobs SET jd_text=?, skills=?, detail_at=datetime('now','localtime') "
+            "WHERE job_id=?",
+            (data.get("jd", ""),
+             json.dumps(data.get("skills") or [], ensure_ascii=False), jid),
+        )
+        conn.commit()  # 逐条提交 => 中断安全
+        jd_len = len(data.get("jd", ""))
+        print(f"  [{i}/{len(todo)}] {jid} JD {jd_len} 字")
+        time.sleep(random.uniform(2, 5))
+
+
+# ── 导出 ────────────────────────────────────────────────────────────────
+def export_csv(conn, path: str, keyword: str | None) -> None:
+    q = ("SELECT title,salary,experience,education,location,company,"
+         "job_tags,company_tags,skills,jd_text,link FROM jobs")
+    params: list = []
+    if keyword:
+        q += " WHERE keyword=?"
+        params.append(keyword)
+    rows = conn.execute(q, params).fetchall()
+    cols = ["title", "salary", "experience", "education", "location", "company",
+            "job_tags", "company_tags", "skills", "jd_text", "link"]
+    with open(path, "w", newline="", encoding="utf-8-sig") as f:
+        w = csv.writer(f)
+        w.writerow(cols)
+        for row in rows:
+            row = list(row)
+            for idx in (6, 7, 8):  # json 数组列 -> 用 / 连接
+                try:
+                    row[idx] = " / ".join(json.loads(row[idx] or "[]"))
+                except Exception:
+                    pass
+            w.writerow(row)
+    print(f"已导出 {len(rows)} 条 -> {path}")
+
+
+# ── main ────────────────────────────────────────────────────────────────
+def main() -> None:
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--keyword", default="AI", help="搜索关键词")
+    ap.add_argument("--pages", type=int, default=3, help="列表抓取页数")
+    ap.add_argument("--jl", default=None, help="城市编码, 不填=全国")
+    ap.add_argument("--db", default=DB_DEFAULT, help="SQLite 文件路径")
+    ap.add_argument("--no-detail", action="store_true", help="只抓列表, 不进详情页")
+    ap.add_argument("--detail-limit", type=int, default=None, help="本轮最多抓多少条 JD")
+    ap.add_argument("--headful", action="store_true", help="显示浏览器窗口")
+    ap.add_argument("--reset", action="store_true", help="重置该关键词的列表进度")
+    ap.add_argument("--export", metavar="CSV", help="导出已抓数据到 CSV 后退出")
+    args = ap.parse_args()
+
+    with closing(init_db(args.db)) as conn:
+        if args.export:
+            export_csv(conn, args.export, args.keyword if args.keyword != "AI" else None)
+            return
+        if args.reset:
+            scope = f"{args.keyword}|{args.jl or ''}"
+            conn.execute("DELETE FROM progress WHERE scope=?", (scope,))
+            conn.commit()
+            print(f"已重置进度: {scope}")
+
+        with InvisiblePlaywright(seed=42, headless=not args.headful) as browser:
+            page = browser.new_page()
+            crawl_list(conn, page, args.keyword, args.pages, args.jl)
+            if not args.no_detail:
+                crawl_details(conn, page, args.keyword, args.detail_limit)
+
+        total = conn.execute("SELECT COUNT(*) FROM jobs").fetchone()[0]
+        with_jd = conn.execute(
+            "SELECT COUNT(*) FROM jobs WHERE jd_text IS NOT NULL AND jd_text!=''"
+        ).fetchone()[0]
+        print(f"\n库内合计 {total} 条, 其中含 JD 全文 {with_jd} 条 -> {args.db}")
+
+
+if __name__ == "__main__":
+    main()

examples/scrape_zhipin_ai.py

@@ -0,0 +1,112 @@
+"""抓取 BOSS 直聘「AI 相关」岗位（自用 / 学习用途）。
+
+合规提醒:
+  - 仅抓取公开展示的岗位标题/薪资/公司等字段, 不抓取招聘者个人联系方式。
+  - 低频请求, 遵守目标站点服务条款; 风险自负。
+
+首次使用:
+  python scrape_zhipin_ai.py --login      # 打开浏览器, 手动扫码登录一次
+之后:
+  python scrape_zhipin_ai.py --keyword AI --city 101010100 --pages 3
+"""
+from __future__ import annotations
+
+import argparse
+import csv
+import json
+import random
+import time
+from pathlib import Path
+from urllib.parse import quote
+
+from invisible_playwright import InvisiblePlaywright
+
+PROFILE_DIR = Path(__file__).parent / ".zhipin_profile"  # 持久化登录态
+SEED = 20240614  # 固定 seed → 跨会话指纹一致, 配合持久化 profile
+
+
+def login_flow() -> None:
+    """首次手动登录: 打开页面, 你扫码, 登录态写入 PROFILE_DIR。"""
+    with InvisiblePlaywright(seed=SEED, profile_dir=PROFILE_DIR) as ctx:
+        page = ctx.new_page()
+        page.goto("https://www.zhipin.com/web/user/?ka=header-login",
+                  wait_until="domcontentloaded")
+        print("请在打开的浏览器中扫码登录, 登录完成后回到终端按回车...")
+        input()  # 等你登录完成
+        print("登录态已保存到", PROFILE_DIR)
+
+
+def scrape(keyword: str, city: str, pages: int) -> list[dict]:
+    results: list[dict] = []
+    with InvisiblePlaywright(seed=SEED, profile_dir=PROFILE_DIR) as ctx:
+        page = ctx.new_page()
+        for n in range(1, pages + 1):
+            url = (
+                "https://www.zhipin.com/web/geek/job"
+                f"?query={quote(keyword)}&city={city}&page={n}"
+            )
+            page.goto(url, wait_until="domcontentloaded")
+
+            # 等列表渲染; 选择器需按实际页面结构核对/调整
+            try:
+                page.wait_for_selector("li.job-card-wrapper", timeout=15000)
+            except Exception:
+                print(f"第 {n} 页未出现岗位列表, 可能需要登录或触发了验证码。")
+                # 给你时间手动过验证码
+                input("处理完页面后按回车继续...")
+
+            rows = page.eval_on_selector_all(
+                "li.job-card-wrapper",
+                """els => els.map(e => ({
+                    title:   e.querySelector('.job-name')?.innerText?.trim(),
+                    salary:  e.querySelector('.salary')?.innerText?.trim(),
+                    company: e.querySelector('.company-name')?.innerText?.trim(),
+                    tags:    Array.from(e.querySelectorAll('.tag-list li'))
+                                  .map(t => t.innerText.trim()),
+                    area:    e.querySelector('.job-area')?.innerText?.trim(),
+                    link:    e.querySelector('a.job-card-left')?.href
+                             || e.querySelector('a')?.href,
+                }))""",
+            )
+            print(f"第 {n} 页抓到 {len(rows)} 条")
+            results.extend(rows)
+
+            # 低频: 随机停顿, 降低风控触发概率
+            time.sleep(random.uniform(4, 9))
+    return results
+
+
+def save(rows: list[dict], stem: str) -> None:
+    Path(f"{stem}.json").write_text(
+        json.dumps(rows, ensure_ascii=False, indent=2), encoding="utf-8"
+    )
+    if rows:
+        keys = ["title", "salary", "company", "area", "tags", "link"]
+        with open(f"{stem}.csv", "w", newline="", encoding="utf-8-sig") as f:
+            w = csv.DictWriter(f, fieldnames=keys, extrasaction="ignore")
+            w.writeheader()
+            for r in rows:
+                r = dict(r)
+                r["tags"] = " / ".join(r.get("tags") or [])
+                w.writerow(r)
+    print(f"已保存 {len(rows)} 条 → {stem}.json / {stem}.csv")
+
+
+def main() -> None:
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--login", action="store_true", help="首次手动登录")
+    ap.add_argument("--keyword", default="AI", help="搜索关键词")
+    ap.add_argument("--city", default="101010100", help="城市编码 (101010100=北京)")
+    ap.add_argument("--pages", type=int, default=3, help="抓取页数")
+    args = ap.parse_args()
+
+    if args.login:
+        login_flow()
+        return
+
+    rows = scrape(args.keyword, args.city, args.pages)
+    save(rows, stem=f"zhipin_{args.keyword}")
+
+
+if __name__ == "__main__":
+    main()

scripts/gen_release_notes.py

@@ -0,0 +1,114 @@
+#!/usr/bin/env python3
+"""Generate the GitHub release body for a firefox-N build from the actual
+invisible_firefox commits that went into it.
+
+The release tag (firefox-N) lives on the wrapper, but the binary's changes live
+on the SOURCE repo (feder-cr/invisible_firefox). We never deep-clone that history
+(it's a full Firefox fork); instead we use GitHub's compare API to list the
+commits between the PREVIOUS release's source commit and this one, and turn their
+subject lines into a short human-readable "What changed" list.
+
+  - The previous release's source commit comes from its ``source-commit.txt``
+    asset (this script's own output uploads one for the next run to read).
+  - If there's no previous source commit (first automated release) or the compare
+    fails, we fall back to a body WITHOUT the changelog section — publishing must
+    never break on note generation.
+
+This is NOT an LLM and NOT a raw ``git log`` dump: it filters out the
+non-user-facing commits (docs/chore/ci/test/style) and prints the remaining
+subjects as plain bullets. Quality rides on writing good commit subjects.
+
+Usage:
+    python scripts/gen_release_notes.py --tag firefox-10 --current <sha> \
+        [--prev-sha <sha>] [--source-repo feder-cr/invisible_firefox]
+    # reads GITHUB_TOKEN from the env for the compare API (optional for public).
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import re
+import sys
+import urllib.request
+import urllib.error
+
+# Conventional-commit prefixes that never belong in user-facing release notes.
+_SKIP = re.compile(r"^(docs|chore|ci|test|style|build)(\(|:)", re.I)
+
+
+def _api(url: str, token: str | None) -> dict:
+    headers = {"Accept": "application/vnd.github+json",
+               "User-Agent": "invisible-playwright-release-notes"}
+    if token:
+        headers["Authorization"] = f"Bearer {token}"
+    req = urllib.request.Request(url, headers=headers)
+    with urllib.request.urlopen(req, timeout=30) as r:
+        return json.load(r)
+
+
+def changelog_bullets(source_repo: str, prev_sha: str, current_sha: str,
+                      token: str | None) -> list[str]:
+    """Return the user-facing commit subjects in prev_sha..current_sha, or []."""
+    if not prev_sha or not current_sha or prev_sha == current_sha:
+        return []
+    url = f"https://api.github.com/repos/{source_repo}/compare/{prev_sha}...{current_sha}"
+    try:
+        data = _api(url, token)
+    except (urllib.error.URLError, urllib.error.HTTPError, ValueError) as e:
+        print(f"[gen_release_notes] compare API failed ({e}); no changelog section",
+              file=sys.stderr)
+        return []
+    bullets: list[str] = []
+    for c in data.get("commits", []):
+        subject = (c.get("commit", {}).get("message") or "").splitlines()[0].strip()
+        if not subject or _SKIP.match(subject):
+            continue
+        bullets.append(subject.rstrip("."))
+    return bullets
+
+
+def build_body(tag: str, current_sha: str, bullets: list[str]) -> str:
+    m = re.search(r"(\d+)", tag)
+    n = int(m.group(1)) if m else None
+    prev_label = f"firefox-{n - 1}" if n else "the previous build"
+    short = (current_sha or "")[:8]
+
+    parts = ["Patched Firefox 150.0.1, the stealth build invisible_playwright drives.", ""]
+    if bullets:
+        parts.append(f"What changed since {prev_label}:")
+        parts += [f"- {b}" for b in bullets]
+        parts.append("")
+    parts += [
+        "Builds: Linux x86_64, Linux arm64, Windows x86_64, macOS arm64, macOS x86_64.",
+        "",
+        "Most people won't grab these by hand. The wrapper fetches the right one for "
+        "your platform on first run:",
+        "",
+        "    pip install git+https://github.com/feder-cr/invisible_playwright",
+        "",
+        "If you do download manually, `checksums.txt` has the SHA256s. The macOS builds "
+        "are ad-hoc signed (not notarized), so clear the quarantine flag: "
+        "`xattr -dr com.apple.quarantine Firefox.app`",
+    ]
+    if short:
+        parts += ["", f"Built from invisible_firefox @{short}."]
+    return "\n".join(parts)
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--tag", required=True, help="release tag, e.g. firefox-10")
+    ap.add_argument("--current", required=True, help="invisible_firefox SHA this build was built from")
+    ap.add_argument("--prev-sha", default="", help="previous release's source SHA (omit for none)")
+    ap.add_argument("--source-repo", default="feder-cr/invisible_firefox")
+    args = ap.parse_args()
+
+    token = os.environ.get("GITHUB_TOKEN") or os.environ.get("GH_TOKEN")
+    bullets = changelog_bullets(args.source_repo, args.prev_sha, args.current, token)
+    sys.stdout.write(build_body(args.tag, args.current, bullets))
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

src/invisible_playwright/_fpforge/_network.py

@@ -75,10 +75,23 @@ class Network:
         self.nodes = _topsort(nodes)
         self.by_name = {n.name: n for n in self.nodes}
 
-    def sample(self, rng: random.Random) -> Dict[str, Any]:
+    def sample(
+        self,
+        rng: random.Random,
+        evidence: Optional[Dict[str, Any]] = None,
+    ) -> Dict[str, Any]:
+        """Sample the network. ``evidence`` fixes named nodes BEFORE their children
+        sample, so the children RE-CONDITION on the fixed value (not relabel after).
+        Used to pin ``gpu_class`` to the validated WebGL persona's class so the whole
+        bundle (cores/screen/fonts) stays coherent with the GPU we expose. Earlier
+        nodes still sample (RNG stream preserved → per-seed determinism)."""
+        evidence = evidence or {}
         context: Dict[str, Any] = {}
         for node in self.nodes:
-            context[node.name] = node.sample(context, rng)
+            if node.name in evidence:
+                context[node.name] = evidence[node.name]
+            else:
+                context[node.name] = node.sample(context, rng)
         return context
 
 

src/invisible_playwright/_fpforge/_sampler.py

@@ -8,7 +8,7 @@ oscpu, webdriver=false, maxTouchPoints=0) is locked by the compiled build.
 
 Graph:
 
-    gpu (root, 444 real Windows ANGLE renderers)
+    gpu (root, 474 real Windows ANGLE renderers)
      │
      └─> gpu_class (deterministic classifier, 6 classes)
           ├─> hw_concurrency       (CPT per class)
@@ -28,7 +28,7 @@ Sampling is deterministic per stealth_seed via a private random.Random.
 import json
 import os
 import re
-from typing import Any, Dict
+from typing import Any, Dict, Optional
 
 from ._network import Network, Node
 
@@ -110,6 +110,16 @@ def classify_gpu(gpu_value: Dict[str, str]) -> str:
 
     if re.search(r"Intel.*HD Graphics (3000|4000|2500)", r):
         return "integrated_old"
+    # Discrete Intel Arc DESKTOP/dGPU cards (A-series / B-series, e.g. A750,
+    # A770, B580) are discrete GPUs (~RTX 3060 tier for A7xx), NOT the
+    # integrated "Arc 130T/140T/Graphics" iGPUs in Core Ultra chips. Route the
+    # discrete SKUs to a coherent discrete-GPU class so the conditioned bundle
+    # (cores, screen, storage) matches a real discrete-GPU machine; A3xx are
+    # entry discrete -> low_end, A5xx/A7xx/Bxxx -> mid_range. Bare "Arc 1x0(T/V)"
+    # integrated names do NOT match and fall through to integrated_modern below.
+    m = re.search(r"Intel.*\bArc(?:\(TM\))?\s+([AB])(\d)\d\d\b", r)
+    if m:
+        return "low_end" if m.group(2) == "3" else "mid_range"
     if re.search(
         r"Intel.*(HD Graphics (4[56]|5\d\d|6\d\d)|UHD Graphics|Graphics Family|Iris|Arc)",
         r,
@@ -328,8 +338,15 @@ class Forge:
         self.seed = int(seed)
         self._rng = random.Random(self.seed)
 
-    def sample(self) -> Dict[str, Any]:
-        bundle = _NETWORK.sample(self._rng)
+    def sample(self, fixed_gpu_class: Optional[str] = None) -> Dict[str, Any]:
+        # fixed_gpu_class pins gpu_class so the WHOLE bundle (cores/screen/fonts) is
+        # drawn coherently for the WebGL persona's class we expose on Windows/mac.
+        # The default (no fix) path calls _NETWORK.sample(rng) with one arg so existing
+        # monkeypatches/tests keep working.
+        if fixed_gpu_class:
+            bundle = _NETWORK.sample(self._rng, evidence={"gpu_class": fixed_gpu_class})
+        else:
+            bundle = _NETWORK.sample(self._rng)
         gpu = bundle["gpu"]
         screen = bundle["screen"]
         audio = bundle["audio"]
@@ -339,7 +356,7 @@ class Forge:
             "stealth_seed": self.seed,
             # Locked identity
             **_LOCKED,
-            # GPU (coherent pair from 444 pool)
+            # GPU (coherent pair from 474 pool)
             "webgl_renderer": gpu["renderer"],
             "webgl_vendor": gpu["vendor"],
             "gpu_class": bundle["gpu_class"],
@@ -392,6 +409,6 @@ class Forge:
         }
 
 
-def sample(seed: int) -> Dict[str, Any]:
-    """Convenience: `Forge(seed).sample()`."""
-    return Forge(seed).sample()
+def sample(seed: int, fixed_gpu_class: Optional[str] = None) -> Dict[str, Any]:
+    """Convenience: `Forge(seed).sample(fixed_gpu_class)`."""
+    return Forge(seed).sample(fixed_gpu_class)

src/invisible_playwright/_fpforge/data/cpt_audio_given_class.json

@@ -1,5 +1,5 @@
 {
-  "_meta": "audio (rate/latency/channels) given gpu_class",
+  "_meta": "audio (rate/latency/channels) given gpu_class. NOTE 2026-06-14: maxChannelCount reflects the OS DEFAULT OUTPUT DEVICE (stereo for the vast majority of users), NOT the GPU — so channels=2 dominates every class (~78-92%) with only a small 6/8 surround tail. The previous tables emitted 45-100% surround on mid/high/workstation, which is unrealistic and lifted FP Pro tampering_ml (surround on a typical consumer profile reads as a coherence anomaly). Rate/latency tuples are unchanged.",
   "table": {
     "integrated_old": [
       {
@@ -26,7 +26,7 @@
           "latency": 30,
           "channels": 2
         },
-        "prob": 0.6
+        "prob": 0.62
       },
       {
         "value": {
@@ -34,7 +34,7 @@
           "latency": 40,
           "channels": 2
         },
-        "prob": 0.25
+        "prob": 0.3
       },
       {
         "value": {
@@ -42,7 +42,7 @@
           "latency": 25,
           "channels": 6
         },
-        "prob": 0.15
+        "prob": 0.08
       }
     ],
     "low_end": [
@@ -52,7 +52,7 @@
           "latency": 40,
           "channels": 2
         },
-        "prob": 0.55
+        "prob": 0.6
       },
       {
         "value": {
@@ -60,7 +60,7 @@
           "latency": 50,
           "channels": 2
         },
-        "prob": 0.3
+        "prob": 0.32
       },
       {
         "value": {
@@ -68,7 +68,7 @@
           "latency": 30,
           "channels": 6
         },
-        "prob": 0.15
+        "prob": 0.08
       }
     ],
     "mid_range": [
@@ -78,31 +78,39 @@
           "latency": 25,
           "channels": 2
         },
-        "prob": 0.45
+        "prob": 0.5
       },
       {
         "value": {
           "rate": 48000,
           "latency": 20,
-          "channels": 6
+          "channels": 2
         },
         "prob": 0.3
       },
-      {
-        "value": {
-          "rate": 48000,
-          "latency": 20,
-          "channels": 8
-        },
-        "prob": 0.15
-      },
       {
         "value": {
           "rate": 44100,
           "latency": 30,
           "channels": 2
         },
-        "prob": 0.1
+        "prob": 0.12
+      },
+      {
+        "value": {
+          "rate": 48000,
+          "latency": 20,
+          "channels": 6
+        },
+        "prob": 0.06
+      },
+      {
+        "value": {
+          "rate": 48000,
+          "latency": 20,
+          "channels": 8
+        },
+        "prob": 0.02
       }
     ],
     "high_end": [
@@ -110,51 +118,75 @@
         "value": {
           "rate": 48000,
           "latency": 15,
-          "channels": 6
+          "channels": 2
         },
-        "prob": 0.3
+        "prob": 0.6
       },
       {
         "value": {
-          "rate": 48000,
-          "latency": 15,
-          "channels": 8
-        },
-        "prob": 0.3
-      },
-      {
-        "value": {
-          "rate": 48000,
+          "rate": 96000,
           "latency": 15,
           "channels": 2
         },
-        "prob": 0.2
+        "prob": 0.18
       },
       {
         "value": {
-          "rate": 96000,
+          "rate": 48000,
           "latency": 15,
           "channels": 6
         },
         "prob": 0.1
       },
+      {
+        "value": {
+          "rate": 48000,
+          "latency": 15,
+          "channels": 8
+        },
+        "prob": 0.05
+      },
+      {
+        "value": {
+          "rate": 96000,
+          "latency": 15,
+          "channels": 6
+        },
+        "prob": 0.05
+      },
       {
         "value": {
           "rate": 96000,
           "latency": 15,
           "channels": 8
         },
-        "prob": 0.1
+        "prob": 0.02
       }
     ],
     "workstation": [
+      {
+        "value": {
+          "rate": 48000,
+          "latency": 10,
+          "channels": 2
+        },
+        "prob": 0.45
+      },
+      {
+        "value": {
+          "rate": 96000,
+          "latency": 10,
+          "channels": 2
+        },
+        "prob": 0.2
+      },
       {
         "value": {
           "rate": 48000,
           "latency": 10,
           "channels": 8
         },
-        "prob": 0.25
+        "prob": 0.12
       },
       {
         "value": {
@@ -162,7 +194,7 @@
           "latency": 10,
           "channels": 8
         },
-        "prob": 0.3
+        "prob": 0.1
       },
       {
         "value": {
@@ -170,7 +202,7 @@
           "latency": 10,
           "channels": 6
         },
-        "prob": 0.2
+        "prob": 0.08
       },
       {
         "value": {
@@ -178,16 +210,8 @@
           "latency": 10,
           "channels": 8
         },
-        "prob": 0.15
-      },
-      {
-        "value": {
-          "rate": 48000,
-          "latency": 15,
-          "channels": 6
-        },
-        "prob": 0.1
+        "prob": 0.05
       }
     ]
   }
-}
+}

src/invisible_playwright/_fpforge/data/cpt_hwc_given_class_tier.json

@@ -36,29 +36,21 @@
       },
       {
         "value": 8,
-        "prob": 0.3
-      },
-      {
-        "value": 12,
-        "prob": 0.05
+        "prob": 0.35
       }
     ],
     "[\"integrated_modern\", \"budget\"]": [
-      {
-        "value": 4,
-        "prob": 0.55
-      },
       {
         "value": 6,
-        "prob": 0.2
+        "prob": 0.45
       },
       {
         "value": 8,
-        "prob": 0.2
+        "prob": 0.4
       },
       {
         "value": 12,
-        "prob": 0.05
+        "prob": 0.15
       }
     ],
     "[\"integrated_modern\", \"standard\"]": [
@@ -178,11 +170,7 @@
       },
       {
         "value": 12,
-        "prob": 0.1
-      },
-      {
-        "value": 16,
-        "prob": 0.05
+        "prob": 0.15
       }
     ],
     "[\"mid_range\", \"standard\"]": [

src/invisible_playwright/_fpforge/data/cpt_screen_given_class_tier.json

@@ -108,16 +108,6 @@
       }
     ],
     "[\"integrated_modern\", \"budget\"]": [
-      {
-        "value": {
-          "w": 1366,
-          "h": 768,
-          "aw": 1366,
-          "ah": 728,
-          "dpr": 1.0
-        },
-        "prob": 0.3
-      },
       {
         "value": {
           "w": 1920,
@@ -126,14 +116,24 @@
           "ah": 1040,
           "dpr": 1.0
         },
-        "prob": 0.65
+        "prob": 0.8
       },
       {
         "value": {
-          "w": 1600,
-          "h": 900,
-          "aw": 1600,
-          "ah": 860,
+          "w": 2560,
+          "h": 1440,
+          "aw": 2560,
+          "ah": 1400,
+          "dpr": 1.0
+        },
+        "prob": 0.15
+      },
+      {
+        "value": {
+          "w": 1920,
+          "h": 1200,
+          "aw": 1920,
+          "ah": 1160,
           "dpr": 1.0
         },
         "prob": 0.05
@@ -758,4 +758,4 @@
       }
     ]
   }
-}
+}

src/invisible_playwright/_fpforge/data/cpt_storage_given_class_tier.json

@@ -48,29 +48,21 @@
       },
       {
         "value": 500000,
-        "prob": 0.3
-      },
-      {
-        "value": 1000000,
-        "prob": 0.05
+        "prob": 0.35
       }
     ],
     "[\"integrated_modern\", \"budget\"]": [
-      {
-        "value": 64000,
-        "prob": 0.2
-      },
-      {
-        "value": 128000,
-        "prob": 0.3
-      },
       {
         "value": 256000,
         "prob": 0.3
       },
       {
         "value": 500000,
-        "prob": 0.2
+        "prob": 0.45
+      },
+      {
+        "value": 1000000,
+        "prob": 0.25
       }
     ],
     "[\"integrated_modern\", \"standard\"]": [
@@ -302,4 +294,4 @@
       }
     ]
   }
-}
+}

src/invisible_playwright/_fpforge/profile.py

@@ -178,7 +178,11 @@ def _apply_pins_to_raw(raw: Dict[str, Any], pin: Dict[str, Any]) -> Dict[str, An
     return out
 
 
-def generate_profile(seed: int, pin: Optional[Dict[str, Any]] = None) -> Profile:
+def generate_profile(
+    seed: int,
+    pin: Optional[Dict[str, Any]] = None,
+    fixed_gpu_class: Optional[str] = None,
+) -> Profile:
     """Return a deterministic Profile for the given integer seed.
 
     pin: optional dict of dotted-path keys (e.g. "screen.width", "gpu.renderer")
@@ -215,7 +219,11 @@ def generate_profile(seed: int, pin: Optional[Dict[str, Any]] = None) -> Profile
         for key in pin:
             _validate_pin_key(key)
 
-    raw = _sample_raw(int(seed))
+    # fixed_gpu_class re-conditions the whole bundle on a chosen class (used so the
+    # bundle stays coherent with the validated WebGL persona we expose on Windows/mac).
+    # An explicit gpu.class_tier pin still wins.
+    eff_class = (pin or {}).get("gpu.class_tier") or fixed_gpu_class
+    raw = _sample_raw(int(seed), fixed_gpu_class=eff_class)
     if pin:
         raw = _apply_pins_to_raw(raw, pin)
 

src/invisible_playwright/_webgl_personas.py

@@ -0,0 +1,163 @@
+"""Empirically-calibrated WebGL GPU personas for Windows ANGLE D3D11.
+
+We expose a FALSE GPU (this is a multi-user tool — never leak each host's real GPU),
+chosen deterministically per seed from a small set of renderer-string "buckets" that
+Firefox's SanitizeRenderer emits and that FP Pro's tampering_ml scores as CLEAN.
+
+## What actually gates a persona (calibrated 2026-06-14, supersedes the old theory)
+
+The blocker is NOT anti_detect and NOT a "render-vs-renderer" check. It is FP Pro's
+**tampering_ml** (gate <=0.5), a holistic ML coherence score. We reverse-engineered its
+GPU sensitivity with single-variable A/Bs on demo.fingerprint.com (deterministic per
+(seed, renderer, IP); tools in tests/_gpu_isolate.py / _gpu_landscape.py / _gpu_sweep.py /
+_gpu_sweep2.py / _gpu_persona_pure.py). Findings:
+
+  1. tampering_ml = f(renderer STRING, seed baseline = canvas/audio). The renderer string
+     carries a STABLE per-bucket penalty; the seed sets the floor it adds to.
+  2. gpu_class is IRRELEVANT to tampering_ml (nv_980 scored identically on mid_range /
+     high_end / premium / workstation). So pairing a fake GPU with a "matching" hardware
+     tier does NOT help the score (we still set a coherent class — see gpu_class below —
+     for OTHER detectors that cross-check cores/screen, just not for this).
+  3. It is NOT render-consistency: a cross-vendor AMD string is CLEAN on our Intel-Arc
+     host. So the real silicon's pixels are not the dominant signal; falsifying to a
+     different vendor works — IF the string is one FP Pro scores low.
+
+Sweep over all 10 Windows SanitizeRenderer buckets x 10 seeds (clean = tml<=0.5 AND not
+anti_detect), on our Intel Arc A750 host:
+  - amd_r9 (Radeon R9 200 Series) ...... 10/10 clean, max tml 0.346   <- SHIP
+  - intel_arc (Arc A750) ............... 10/10 clean, max tml 0.377   <- SHIP
+  - amd_hd5850 ......................... 9/10 (fails the hardest seed)
+  - amd_hd3200 / intel_hd .............. 6/10 (seed-dependent, risky)
+  - intel_hd400 ........................ 3/10
+  - ALL NVIDIA (8800/480/980) .......... 0/10 (penalized everywhere, ~0.7-0.99)
+  - intel_945 (ancient Intel) .......... 0/10
+So only TWO buckets are robustly clean across profiles. We ship exactly those, weighted
+to real-world prevalence ("Radeon R9 200 Series" is the bucket for ALL modern AMD = a big
+real slice; "Arc A750" covers Intel discrete = rarer). Cross-vendor, so the fleet is not a
+single-GPU cluster. More names require lowering the seed floor first (see CAVEAT 2).
+
+## ⚠️ CAVEATS
+ 1. HOST-INDEPENDENCE NOT PROVEN. Everything above was measured on ONE host (Intel Arc
+    A750). The host's real render is embedded in the seed baseline, so the clean-bucket set
+    *might* be host-dependent (on a real NVIDIA host, maybe nv_980 is clean and amd_r9 is
+    not). This MUST be validated on a non-Arc machine before trusting it fleet-wide; if it
+    turns out host-dependent, add a pre-launch host-GPU-class probe and pick a bucket per
+    detected class. Until then: safe for Arc hosts (incl. the dev's), unvalidated elsewhere.
+ 2. DIVERSITY CEILING = 2 names because "hard" seeds (high canvas/audio floor, e.g. seed 4
+    ~0.35) only stay clean on the 2 best buckets. Lowering that floor (an fpforge CPT fix —
+    candidate: 8-channel audio + 1TB storage emitted on a mid_range profile) would unlock
+    amd_hd5850 / intel_hd for more seeds => up to ~5 names. Follow-up, not done yet.
+
+## Load-bearing format requirements (unchanged, still true)
+ - renderer MUST end ", D3D11)" (full ANGLE wire format) or SanitizeRenderer returns
+   "Generic Renderer" (a tell). The C++ passes our string through SanitizeRenderer, which
+   buckets "AMD Radeon R9 200 Series" -> "Radeon R9 200 Series" and "Arc A750" -> itself.
+ - the forced extension list MUST be the EXACT NATIVE ORDER getSupportedExtensions returns.
+   The set+order is fixed by Firefox+ANGLE on D3D11 FL11_0 (VENDOR-INDEPENDENT — verified
+   via 20-agent source study), so ONE list is correct for both personas. A reorder is caught
+   (tampering_ml 0.34 -> 0.84). The lists below are the verbatim native-order Arc capture.
+
+Calibration data + sweep tooling live in the local workbench (not shipped).
+"""
+from __future__ import annotations
+
+import sys
+from typing import Dict, List, Optional
+
+# Vendor-independent ext lists (native order, Arc host capture). Identical for every persona
+# because the set+order is fixed by Firefox+ANGLE on D3D11 FL11_0, not by the GPU vendor.
+_EXT1 = (
+    "ANGLE_instanced_arrays,EXT_blend_minmax,EXT_color_buffer_half_float,EXT_float_blend,"
+    "EXT_frag_depth,EXT_shader_texture_lod,EXT_sRGB,EXT_texture_compression_bptc,"
+    "EXT_texture_compression_rgtc,EXT_texture_filter_anisotropic,OES_element_index_uint,"
+    "OES_fbo_render_mipmap,OES_standard_derivatives,OES_texture_float,OES_texture_float_linear,"
+    "OES_texture_half_float,OES_texture_half_float_linear,OES_vertex_array_object,"
+    "WEBGL_color_buffer_float,WEBGL_compressed_texture_s3tc,WEBGL_compressed_texture_s3tc_srgb,"
+    "WEBGL_debug_renderer_info,WEBGL_debug_shaders,WEBGL_depth_texture,WEBGL_draw_buffers,"
+    "WEBGL_lose_context,WEBGL_provoking_vertex"
+)
+_EXT2 = (
+    "EXT_color_buffer_float,EXT_float_blend,EXT_texture_compression_bptc,"
+    "EXT_texture_compression_rgtc,EXT_texture_filter_anisotropic,OES_draw_buffers_indexed,"
+    "OES_texture_float_linear,OVR_multiview2,WEBGL_compressed_texture_s3tc,"
+    "WEBGL_compressed_texture_s3tc_srgb,WEBGL_debug_renderer_info,WEBGL_debug_shaders,"
+    "WEBGL_lose_context,WEBGL_provoking_vertex"
+)
+
+
+def _p(key, renderer, vendor, gpu_class, weight):
+    return {"key": key, "renderer": renderer, "vendor": vendor,
+            "gpu_class": gpu_class, "weight": weight, "ext1": _EXT1, "ext2": _EXT2}
+
+
+# Only the two robustly-clean Windows buckets (calibration sweep 2026-06-14). Both discrete,
+# so gpu_class=mid_range keeps cores/screen coherent with the declared GPU for OTHER detectors
+# (gpu_class does NOT affect tampering_ml). Weights ~ real-world prevalence of the BUCKET:
+# "Radeon R9 200 Series" represents ALL modern AMD (large real slice); "Arc A750" = Intel
+# discrete (rarer). Cross-vendor => the fleet is not a single-GPU cluster.
+_PERSONAS: List[Dict] = [
+    _p("amd_radeon_r9", "ANGLE (AMD, AMD Radeon R9 200 Series Direct3D11 vs_5_0 ps_5_0, D3D11)",
+       "Google Inc. (AMD)", "mid_range", 70),    # -> bucket "Radeon R9 200 Series"; tml 0.03-0.35
+    _p("intel_arc_a750", "ANGLE (Intel, Intel(R) Arc(TM) A750 Graphics Direct3D11 vs_5_0 ps_5_0, D3D11)",
+       "Google Inc. (Intel)", "mid_range", 30),  # -> bucket "Intel(R) Arc(TM) A750 Graphics"; tml 0.02-0.38
+]
+
+_TOTAL_W = sum(p["weight"] for p in _PERSONAS)
+
+# ENABLED: we falsify the GPU on Windows/mac. Validated clean on an Intel Arc host (see the
+# HOST-INDEPENDENCE caveat in the module docstring — unvalidated on non-Arc hosts). On Linux
+# select_persona returns None: there prefs.py spoofs profile.gpu.renderer directly.
+_ENABLED = True
+
+
+def select_persona(seed: int) -> Optional[Dict]:
+    """Deterministic, prevalence-weighted persona for this seed (None on Linux).
+
+    Same seed -> same persona (fppro_consistency: identity stable per seed). Different seeds
+    spread across the persona mix by weight. None on Linux (the sampled profile.gpu.renderer
+    is spoofed directly there).
+    """
+    if not _ENABLED or sys.platform.startswith("linux") or not _PERSONAS:
+        return None
+    h = (int(seed) * 2654435761) % _TOTAL_W
+    cum = 0
+    for p in _PERSONAS:
+        cum += p["weight"]
+        if h < cum:
+            return p
+    return _PERSONAS[-1]
+
+
+def forced_gpu_class(seed: int) -> Optional[str]:
+    """The gpu_class the forge conditions the WHOLE bundle on (== the selected persona's class),
+    so cores/screen/fonts stay coherent with the GPU we expose. Does NOT affect FP Pro
+    tampering_ml (proven) but matters for detectors that cross-check hardware tier. None on Linux."""
+    p = select_persona(seed)
+    return p["gpu_class"] if p else None
+
+
+# ── Render-noise seed pool (canvas/WebGL gamma) ──────────────────────────────
+# zoom.stealth.fpp.hw_seed drives the per-seed canvas2D + WebGL readPixels gamma
+# LUT in C++. The render-image HASH it produces is the DOMINANT FP Pro tampering_ml
+# driver (proven 2026-06-14: holding a fixed profile and varying ONLY hw_seed moved
+# tml 0.25->0.75). The monotonic gamma preserves the GPU's render structure, so some
+# hw_seeds yield a "suspicious" render hash. We therefore DECOUPLE the render-noise
+# seed from the identity seed and pick from a calibrated pool of hw_seeds that score
+# CLEAN even on the hardest attribute profile (sweep 1..30 vs the worst seed: these
+# 14 all gave tml<=0.285). Diversity is preserved (14 distinct render hashes spread
+# across the population — real GPUs cluster to few canvas hashes anyway); identity
+# stays per-seed (the rest of the fingerprint differs). Same seed -> same render seed
+# (fppro_consistency holds).
+# CAVEAT: the render hash = f(host GPU render, gamma), so this pool is calibrated on
+# the Intel-Arc host. On other GPUs the clean set may differ (host-independence open,
+# same as the personas) — Option B (substitution = GPU-independent render hash) would
+# remove that dependency. Validate per-host or move to B before trusting fleet-wide.
+CLEAN_RENDER_SEEDS = [19, 10, 28, 24, 23, 16, 11, 30, 17, 22, 3, 9, 12, 26]
+
+
+def render_noise_seed(seed: int) -> int:
+    """Deterministic clean render-noise seed for hw_seed (decoupled from identity).
+
+    Maps the identity seed into CLEAN_RENDER_SEEDS so every session gets a calibrated
+    clean canvas/WebGL render hash while keeping per-user diversity. Stable per seed."""
+    return CLEAN_RENDER_SEEDS[(int(seed) * 2654435761) % len(CLEAN_RENDER_SEEDS)]

src/invisible_playwright/async_api.py

@@ -9,8 +9,9 @@ from typing import Any, Dict, Optional, Union
 from playwright.async_api import Browser, BrowserContext, Playwright, async_playwright
 
 from ._fpforge import Profile, generate_profile
+from ._webgl_personas import forced_gpu_class
 from ._geo import prepare_session_geo
-from ._headless import make_virtual_display
+from ._headless import cloak_prefs, make_virtual_display
 from ._proxy import configure_proxy as _configure_proxy_shared
 from .download import ensure_binary
 from .launcher import _CHROME_H, _CHROME_W, _TASKBAR_H, _tz_env
@@ -68,7 +69,9 @@ class InvisiblePlaywright:
         self._profile_dir: Optional[Path] = Path(profile_dir) if profile_dir else None
         # reCAPTCHA pre-seed gated server-side; respect persistent profile.
         self._prep_recaptcha = bool(prep_recaptcha) and self._profile_dir is None
-        self._profile: Profile = generate_profile(self.seed, pin=self._pin)
+        self._profile: Profile = generate_profile(
+            self.seed, pin=self._pin, fixed_gpu_class=forced_gpu_class(self.seed)
+        )
         self._pw: Optional[Playwright] = None
         self._browser: Optional[Browser] = None
         self._persistent_context: Optional[BrowserContext] = None
@@ -95,10 +98,19 @@ class InvisiblePlaywright:
             extra_prefs=self._extra_prefs,
             virtual_display=bool(self._headless and _sys.platform == "win32"),
         )
-        prefs["invisible_playwright.humanize"] = bool(self._humanize)
+        # Windows & macOS hide the headless window via the binary's own cloak
+        # (DWMWA_CLOAK / NSWindow alpha) — inject the pref so the patched build
+        # cloaks its chrome windows. setdefault: an explicit user override wins.
+        # (Mirrors launcher._build_prefs; the sync path always did this, async
+        # didn't — so async headless=True never cloaked AND crashed below.)
+        if self._headless and _sys.platform in ("win32", "darwin"):
+            for _k, _v in cloak_prefs().items():
+                prefs.setdefault(_k, _v)
+        # stealthfox.* is the namespace the binary's Juggler reads (see launcher.py note).
+        prefs["stealthfox.humanize"] = bool(self._humanize)
         if self._humanize:
             cap = 1.5 if self._humanize is True else float(self._humanize)
-            prefs["invisible_playwright.humanize.maxTime"] = str(cap)
+            prefs["stealthfox.humanize.maxTime"] = str(cap)
         playwright_proxy = _configure_proxy_shared(self._proxy, prefs)
         pw_headless = self._resolve_headless()
         env = self._build_env()
@@ -223,8 +235,13 @@ class InvisiblePlaywright:
         if not self._headless:
             return False
         vd = make_virtual_display()
-        vd.start()
-        self._virtual_display = vd
+        # Linux: Xvfb to start. Windows/macOS: make_virtual_display() returns
+        # None (the binary self-cloaks via cloak_prefs injected in __aenter__),
+        # so there is nothing to start — guarding the None was the missing piece
+        # that made async headless=True crash with AttributeError on Windows.
+        if vd is not None:
+            vd.start()
+            self._virtual_display = vd
         return False
 
 

src/invisible_playwright/config.py

@@ -38,6 +38,7 @@ import secrets
 from typing import Any, Dict, List, Optional, Union
 
 from ._fpforge import generate_profile
+from ._webgl_personas import forced_gpu_class
 from .prefs import translate_profile_to_prefs
 
 
@@ -83,7 +84,7 @@ def get_default_stealth_prefs(
         ``playwright.firefox.launch()`` or ``launch_persistent_context()``.
     """
     resolved_seed = int(seed) if seed is not None else secrets.randbits(31)
-    profile = generate_profile(resolved_seed, pin=pin)
+    profile = generate_profile(resolved_seed, pin=pin, fixed_gpu_class=forced_gpu_class(resolved_seed))
     prefs = translate_profile_to_prefs(
         profile,
         locale=locale,
@@ -91,10 +92,11 @@ def get_default_stealth_prefs(
         extra_prefs=extra_prefs,
         virtual_display=virtual_display,
     )
-    prefs["invisible_playwright.humanize"] = bool(humanize)
+    # stealthfox.* is the namespace the binary's Juggler reads (see launcher.py note).
+    prefs["stealthfox.humanize"] = bool(humanize)
     if humanize:
         max_seconds = float(humanize) if not isinstance(humanize, bool) else 1.5
-        prefs["invisible_playwright.humanize.maxTime"] = str(max_seconds)
+        prefs["stealthfox.humanize.maxTime"] = str(max_seconds)
     return prefs
 
 

src/invisible_playwright/constants.py

@@ -7,7 +7,7 @@ bugfixes don't force a multi-hour Firefox rebuild.
 from __future__ import annotations
 
 # Bump this when a new patched Firefox build is released on GitHub.
-BINARY_VERSION: str = "firefox-9"
+BINARY_VERSION: str = "firefox-10"
 
 # Releases known to be broken — ensure_binary() refuses them with a clear error
 # instead of handing the user an unusable binary. firefox-8 was packaged without

src/invisible_playwright/launcher.py

@@ -8,6 +8,7 @@ from typing import Any, Dict, Optional, Union
 from playwright.sync_api import Browser, BrowserContext, Playwright, sync_playwright
 
 from ._fpforge import Profile, generate_profile
+from ._webgl_personas import forced_gpu_class
 from ._geo import prepare_session_geo
 from ._headless import cloak_prefs, make_virtual_display
 from ._proxy import configure_proxy as _configure_proxy_shared
@@ -178,7 +179,9 @@ class InvisiblePlaywright:
         # persistent profile_dir is in use, respect its existing cookies
         # and DON'T enable pre-seed (the profile owns its own state).
         self._prep_recaptcha = bool(prep_recaptcha) and self._profile_dir is None
-        self._profile: Profile = generate_profile(self.seed, pin=self._pin)
+        self._profile: Profile = generate_profile(
+            self.seed, pin=self._pin, fixed_gpu_class=forced_gpu_class(self.seed)
+        )
         self._pw: Optional[Playwright] = None
         self._browser: Optional[Browser] = None
         self._persistent_context: Optional[BrowserContext] = None
@@ -346,9 +349,13 @@ class InvisiblePlaywright:
         if self._headless and _sys.platform in ("win32", "darwin"):
             for _k, _v in cloak_prefs().items():
                 prefs.setdefault(_k, _v)
-        prefs["invisible_playwright.humanize"] = bool(self._humanize)
+        # Pref namespace MUST be stealthfox.* — that's what the binary's Juggler
+        # reads (PageHandler.js gates the Bezier mouse path on `stealthfox.humanize`).
+        # The old `invisible_playwright.*` name was a dead no-op (nothing read it), so
+        # humanize silently never fired and every click teleported the cursor.
+        prefs["stealthfox.humanize"] = bool(self._humanize)
         if self._humanize:
-            prefs["invisible_playwright.humanize.maxTime"] = str(self._humanize_max_seconds())
+            prefs["stealthfox.humanize.maxTime"] = str(self._humanize_max_seconds())
         return prefs
 
     def _build_env(self) -> Dict[str, str]:

src/invisible_playwright/prefs.py

@@ -21,6 +21,7 @@ import sys
 from typing import Any, Dict, Optional
 
 from ._fpforge import Profile
+from ._webgl_personas import render_noise_seed, select_persona
 
 
 # ──────────────────────────────────────────────────────────────────────
@@ -231,6 +232,17 @@ _BASELINE: Dict[str, Any] = {
     "network.proxy.socks_remote_dns":                     True,
     "network.proxy.failover_direct":                      False,
 
+    # TLS ClientHello fingerprint — match stock Firefox byte-for-byte.
+    # The Playwright/Juggler Firefox build this binary derives from re-enables
+    # cipher 0xC009 (TLS_ECDHE_ECDSA_WITH_AES_128_CBC_SHA), which retail Firefox
+    # 150 does NOT offer. That extra (17th) cipher shifts our JA3/JA4 away from
+    # any real Firefox (ja4 t13d1717h2 vs stock t13d1617h2). A ClientHello that
+    # matches no real browser is itself a consistency tell. Disabling it makes
+    # JA3/JA4/peetprint byte-identical to retail FF150 (verified on tls.peet.ws).
+    # Stock Firefox ships without 0xC009 and works on the whole web, so this only
+    # improves fingerprint consistency — it cannot break connectivity.
+    "security.ssl3.ecdhe_ecdsa_aes_128_sha":              False,
+
     # Safebrowsing — chatty and fingerprintable.
     "browser.safebrowsing.malware.enabled":               False,
     "browser.safebrowsing.phishing.enabled":              False,
@@ -437,22 +449,41 @@ def _accept_language(locale: str) -> str:
 def _font_metrics_for_platform(profile_metrics: str) -> str:
     """Return ``zoom.stealth.font.metrics`` value.
 
-    Windows: empty string. The C++ width-scale hook is a no-op and
-    Firefox renders Arial/Segoe/Calibri/etc. at their native canonical
-    widths. Applying the Bayesian-sampled per-font factors on a Windows
-    build would *distort* real metrics and surface as a font_preferences
-    width anomaly to FP Pro / reCAPTCHA.
+    The C++ whitelist hook (``gfxPlatformFontList::FindAndAddFamiliesLocked``)
+    backs EVERY whitelisted *named* family with the list-head family on every
+    platform. Without per-font width factors, that means each named font
+    (Arial, Times New Roman, Courier New, …) renders with identical glyphs and
+    collapses to a SINGLE canvas ``measureText`` width — a non-physical
+    1-distinct-width result that strict JS-sensor anti-bots flag via their
+    font probe. The per-font factors in ``profile_metrics``
+    (``arial|0.978,arial black|1.168,…``) spread the fabricated families back
+    to distinct, realistic, deterministic-per-seed widths, so we apply them on
+    EVERY platform (previously suppressed on Windows/mac, which left the
+    collapse in place — only the CSS-generic vector, which FP Pro probes, was
+    ever correct there).
 
-    Linux: prepend generic-family compensation factors so DejaVu /
-    Liberation render at the widths Windows JS expects, then append the
-    per-font factors that make each fabricated family detectable by
-    width-diff probes.
+    These factors only key *named* families. CSS generics
+    (serif/sans-serif/monospace/system-ui) bypass the whitelist entirely and
+    render at the host's native widths, so they are never present in
+    ``profile_metrics`` and stay unfactored — FP Pro's ``font_preferences``
+    probe (which measures the generics) is unaffected. That is also why
+    applying named-font factors here does NOT distort the canonical generic
+    widths.
+
+    Linux ADDITIONALLY needs generic-family compensation
+    (``_LINUX_GENERIC_FONT_FACTORS``) because DejaVu/Liberation generics render
+    wider/narrower than the Windows widths the spoofed profile claims; on
+    Windows/mac the generics already render native, so no generic compensation
+    is applied — only the named-font factors.
     """
     if not profile_metrics:
         return ""
     if sys.platform.startswith("linux"):
         return _LINUX_GENERIC_FONT_FACTORS + profile_metrics
-    return ""  # Windows: NEVER apply width-scale factors.
+    # Windows / macOS: named-font factors only (the generics render native and
+    # bypass the whitelist, so no generic compensation — but the named families
+    # MUST be factored or they all collapse to the list-head width).
+    return profile_metrics
 
 
 def translate_profile_to_prefs(
@@ -479,21 +510,32 @@ def translate_profile_to_prefs(
     # GPU / WebGL renderer/vendor.
     # On Linux we spoof to a Windows ANGLE renderer string (profile.gpu.renderer)
     # so cross-platform sessions report a consistent Windows GPU identity.
-    # On Windows, spoofing a different GPU creates a renderer/parameters hash
-    # mismatch: FP Pro hashes all 81 CN-set getParameter() values including
-    # enum 7937 (RENDERER). Setting GTX 980 while ANGLE returns Intel Arc A750
-    # parameters produces an OOD (hash 23d0a74b vs vanilla 66544db) that FP Pro
-    # ML scores at ~0.70 (confirmed: direct SF146 vs vanilla on same machine).
-    # Fix: leave renderer/vendor empty on Windows → ANGLE reports native hardware
-    # (SanitizeRenderer path at ClientWebGLContext.cpp:2592-2595) → consistent.
+    # On Windows/mac, spoofing a renderer string ALONE is unsafe — the ~81
+    # getParameter values stay real, so a name↔params hash mismatch FP Pro flags
+    # (setting GTX 980 over real Arc A750 params scored ~0.70). Instead we apply a
+    # VALIDATED PERSONA (see _webgl_personas): a {renderer, vendor} whose params are
+    # the shared ANGLE D3D11 caps (vendor-independent — identical on any host, per the
+    # ANGLE source) and whose extension list is FORCED below. That is a coherent fake
+    # GPU that passes FP Pro host-independently (the host's real GPU never leaks). If no
+    # validated persona exists for the sampled gpu_class yet, fall back to the host-real
+    # renderer (empty → native ANGLE; SanitizeRenderer at ClientWebGLContext.cpp:2592).
+    _persona = None
     if sys.platform.startswith("linux"):
         prefs["zoom.stealth.webgl.renderer"] = profile.gpu.renderer
         prefs["zoom.stealth.webgl.vendor"]   = profile.gpu.vendor
         _renderer_lo = (profile.gpu.renderer or "").lower()
     else:
-        prefs["zoom.stealth.webgl.renderer"] = ""
-        prefs["zoom.stealth.webgl.vendor"]   = ""
-        _renderer_lo = "intel"  # test hardware is Intel Arc A750
+        _persona = select_persona(profile.seed)
+        if _persona:
+            prefs["zoom.stealth.webgl.renderer"] = _persona["renderer"]
+            prefs["zoom.stealth.webgl.vendor"]   = _persona["vendor"]
+        else:
+            prefs["zoom.stealth.webgl.renderer"] = ""
+            prefs["zoom.stealth.webgl.vendor"]   = ""
+        # Canvas-noise mask is calibrated to the REAL host GPU's rendering variance — the canvas is
+        # drawn by real hardware, NOT the persona's claimed GPU, so it must NOT follow the persona
+        # (a non-Intel persona on an Intel host would over-noise). Deployment host is Intel.
+        _renderer_lo = "intel"
 
     # MSAA: on Windows, pin to 4 (Firefox default for ANGLE) so gl.SAMPLES is
     # constant across all sessions. Different MSAA values cause different CN-set
@@ -522,7 +564,8 @@ def translate_profile_to_prefs(
     prefs["zoom.stealth.screen.dpr"]          = profile.screen.dpr
     prefs["layout.css.devPixelsPerPx"]        = str(profile.screen.dpr)
 
-    # Hardware
+    # Hardware — coherent with the sampled gpu_class by construction (the forge
+    # draws hw_concurrency conditioned on the GPU class).
     prefs["zoom.stealth.hw_concurrency"]      = profile.hardware.concurrency
     prefs["zoom.stealth.storage.quota_mb"]    = profile.hardware.storage_quota_mb
 
@@ -566,8 +609,12 @@ def translate_profile_to_prefs(
     # Cross-process seed (canvas noise + DWrite gamma share this). Only
     # zoom.stealth.fpp.hw_seed is read by the C++; the old zoom.stealth.seed
     # alias was never declared in the yaml and read by nothing — dropped
-    # 2026-06-10.
-    prefs["zoom.stealth.fpp.hw_seed"] = profile.seed
+    # 2026-06-10. The render-noise seed is DECOUPLED from the identity seed and
+    # drawn from a calibrated CLEAN pool: the canvas/WebGL render HASH it drives
+    # is the dominant FP Pro tampering_ml signal, and some hw_seeds yield a
+    # "suspicious" render hash. render_noise_seed() maps to the clean pool while
+    # keeping per-seed determinism + diversity. See _webgl_personas.
+    prefs["zoom.stealth.fpp.hw_seed"] = render_noise_seed(profile.seed)
 
     # Synthetic host ICE candidate — injected by C++ when addr_ct==0 (SOCKS5
     # proxy suppresses all local addresses so Firefox can't gather host cands).
@@ -577,13 +624,22 @@ def translate_profile_to_prefs(
     _lan_ip = f"192.168.{(_s >> 8) % 254 + 1}.{_s % 254 + 1}"
     prefs["zoom.stealth.webrtc.host_ip"] = _lan_ip
 
-    # On Windows, native ANGLE extension list already matches real Windows users.
-    # The baseline hard-codes a curated _WEBGL1/2_EXTENSIONS list designed for
-    # Linux Mesa → clear it so Windows sessions report the native extension set
-    # (hash matches real Intel Arc A750 vanilla captures).
+    # Windows/mac extension list:
+    #  - persona active → FORCE the validated extension list. A non-Intel host's native
+    #    extensions would mismatch the persona's renderer (renderer says AMD/Intel-Arc but
+    #    extensions are the host's), so the persona must carry its own list to stay
+    #    host-independent.
+    #  - no persona → clear so the host-real renderer reports its native extension set
+    #    (matches real vanilla captures for that host's GPU).
     if not sys.platform.startswith("linux"):
-        prefs["zoom.stealth.webgl.extensions"]  = ""
-        prefs["zoom.stealth.webgl2.extensions"] = ""
+        if _persona:
+            # The persona carries its OWN extension lists in EXACT NATIVE ORDER — a
+            # reordered/foreign list is flagged by FP Pro (verified 2026-06-13).
+            prefs["zoom.stealth.webgl.extensions"]  = _persona["ext1"]
+            prefs["zoom.stealth.webgl2.extensions"] = _persona["ext2"]
+        else:
+            prefs["zoom.stealth.webgl.extensions"]  = ""
+            prefs["zoom.stealth.webgl2.extensions"] = ""
 
     # Linux Xvfb workarounds (no-op on Windows).
     if sys.platform.startswith("linux"):

tests/test_cloak.py

@@ -91,9 +91,18 @@ def test_cloak_hides_window_but_keeps_rendering(firefox_binary):
         shot = page.screenshot()
         assert len(shot) > 3000, "cloaked window produced a blank screenshot (rendering paused)"
 
-        # 2) real WebGL present (native headless has none) -> headed pipeline intact.
+        # 2) headed pipeline intact: a real WebGL context (Playwright's native
+        #    headless has none). Linux (Xvfb + llvmpipe) and Windows (WARP) give a
+        #    software context on the GPU-less runners, so a missing context there
+        #    is a real regression -> hard fail. macOS GitHub runners expose NO
+        #    WebGL in the CI session at all (even vanilla Firefox), and macOS has
+        #    no software-GL fallback; the cloak's "still rendering" property is
+        #    already proven by the non-blank screenshot above, so we don't also
+        #    require a live WebGL context there.
         renderer = page.evaluate(_WEBGL_RENDERER)
-        assert renderer and renderer != "NO-WEBGL", f"no real WebGL under cloak: {renderer!r}"
+        webgl_ok = bool(renderer) and renderer != "NO-WEBGL"
+        if not (sys.platform == "darwin" and not webgl_ok):
+            assert webgl_ok, f"no real WebGL under cloak: {renderer!r}"
 
         # 3) the window is actually hidden (per-platform).
         if sys.platform == "win32":

tests/test_detectors_e2e.py

@@ -6,21 +6,32 @@ libraries and uses their FULL API surface:
   * BotD (@fingerprintjs/botd, MIT) — the client-side bot detector that
     FingerprintJS Pro itself uses. We assert the aggregate verdict
     (``detect().bot == False``) AND every one of its ~18 individual detectors
-    (``getDetections()``) returns ``bot == False``. The per-detector view is
-    why we could delete our hand-rolled ``test_botd_*`` mirrors — the real
-    library now covers each detector, with the same granularity.
+    (``getDetections()``) returns ``bot == False``.
   * FingerprintJS open-source (MIT) — ``get()`` must return a ``visitorId``
-    that is STABLE across two fresh launches with the same seed (an
-    over-randomized spoof drifts), and a RICH component set (the fingerprint
-    surface is real, not a stub).
+    that is STABLE across two fresh launches with the same seed, and a RICH
+    component set (the fingerprint surface is real, not a stub).
+  * fpscanner (antoinevastel/fpscanner 1.0.6, MIT) — ``collectFingerprint()``
+    runs ~21 bot-detection rules in the browser. We assert the **engine-agnostic**
+    subset (webdriver / selenium / bot-UA / platform / timezone / language) is
+    clean. We deliberately do NOT assert the Chrome/GPU-only rules (hasCDP,
+    hasPlaywright, hasSwiftshaderRenderer, hasMissingChromeObject, …): they're
+    trivially clean on Firefox, and the GPU ones can legitimately fire on a
+    software-WebGL CI host (Xvfb/llvmpipe) — asserting them would false-red.
+  * CreepJS (abrahamjuliot/creepjs, MIT, pinned) — the gold-standard Firefox-aware
+    headless/stealth/lie detector. It exposes its result on ``window.Fingerprint``.
+    We assert ``headlessRating == 0`` (webdriver + headless-UA tells) and the
+    JS-proxy stealth tells are absent. ``stealthRating`` / ``totalLies`` /
+    ``likeHeadlessRating`` are LOGGED, not hard-asserted, because some of their
+    sub-signals (hasBadWebGL, prefers-light-color) are GPU/theme-sensitive and
+    differ on a GPU-less CI host.
 
 Everything is hermetic: the libraries are vendored (tests/vendor/) and served
-from a localhost HTTP server — no external CDN call (Firefox tracking-protection
-blocks the CDN anyway) and no IP/network dependency. Runs identically on a dev
-box and on a GitHub runner.
+from a localhost HTTP server — no external CDN call. For CreepJS, every non-local
+request is aborted, so its optional crowd-comparison POST never runs and the
+verdict is computed purely locally. Runs identically on a dev box and a GH runner.
 
-NOT covered: FingerprintJS *Pro* (commercial, server-side, IP/residential
-analysis) — can't be self-hosted, stays the local realness gate.
+NOT covered: FingerprintJS *Pro* (commercial, server-side) — stays the local
+realness gate.
 """
 from __future__ import annotations
 
@@ -36,19 +47,33 @@ from invisible_playwright import InvisiblePlaywright
 _VENDOR = Path(__file__).parent / "vendor"
 _BOTD = "botd-2.0.0.esm.js"
 _FPJS = "fingerprintjs-5.2.0.umd.min.js"
+_FPSCANNER = "fpscanner-1.0.6.es.js"
+_CREEPJS = "creepjs-10aa672.js"  # pinned abrahamjuliot/creepjs@10aa6724
+
+# fpscanner rules that are MEANINGFUL on Firefox and GPU-independent — these must
+# stay clean. The omitted rules are Chrome-only (hasCDP/hasPlaywright/
+# hasMissingChromeObject/hasHighCPUCount/hasImpossibleDeviceMemory/
+# headlessChromeScreenResolution) or GPU-sensitive on a software-WebGL CI host
+# (hasSwiftshaderRenderer/hasGPUMismatch/hasMismatchWebGLInWorker).
+_FPSCANNER_AGNOSTIC = [
+    "hasWebdriver", "hasWebdriverIframe", "hasWebdriverWorker", "hasWebdriverWritable",
+    "hasSeleniumProperty", "hasBotUserAgent", "hasPlatformMismatch",
+    "hasMismatchLanguages", "hasUTCTimezone", "hasMismatchPlatformIframe",
+    "hasMismatchPlatformWorker", "hasInconsistentEtsl",
+]
 
 _PAGE = f"""<!doctype html><html><head><meta charset="utf-8">
 <title>detectors</title>
 <script src="/{_FPJS}"></script>
 </head><body><h1 id="state">loading</h1>
 <script type="module">
-window.__botd = null; window.__fp = null; window.__err = "";
+window.__botd = null; window.__fp = null; window.__fps = null; window.__err = "";
 (async () => {{
   try {{
     const Botd = await import("/{_BOTD}");
-    const botd = await Botd.load();          // load() collects internally
-    const verdict = botd.detect();           // {{bot:false}} | {{bot:true,botKind}}
-    const raw = botd.getDetections() || {{}}; // per-detector verdicts
+    const botd = await Botd.load();
+    const verdict = botd.detect();
+    const raw = botd.getDetections() || {{}};
     const detections = {{}};
     for (const k in raw) detections[k] = {{ bot: raw[k].bot, botKind: raw[k].botKind || null }};
     window.__botd = {{ bot: verdict.bot, botKind: verdict.botKind || null, detections }};
@@ -60,24 +85,41 @@ window.__botd = null; window.__fp = null; window.__err = "";
     const errored = keys.filter(k => r.components[k] && "error" in r.components[k]);
     window.__fp = {{ visitorId: r.visitorId, componentKeys: keys, erroredComponents: errored }};
   }} catch (e) {{ window.__err += " fp:" + e; }}
+  try {{
+    const M = await import("/{_FPSCANNER}");
+    const scanner = new M.default();
+    const fp = await scanner.collectFingerprint({{ encrypt: false }});
+    window.__fps = {{ fastBotDetection: fp.fastBotDetection, details: fp.fastBotDetectionDetails }};
+  }} catch (e) {{ window.__err += " fps:" + e; }}
   document.getElementById("state").textContent = "done";
 }})();
 </script></body></html>"""
 
+# CreepJS gets its own page: creep.js is a plain `defer` script that runs on load
+# and populates window.Fingerprint. A minimal DOM is enough (the rich report DOM
+# is only for the visual page, not the computation).
+_CREEP_PAGE = f"""<!doctype html><html><head><meta charset="utf-8"><title>creep</title></head>
+<body><div id="fingerprint-data"></div><script src="/{_CREEPJS}" defer></script></body></html>"""
+
 
 class _DetectorSite:
-    """Localhost server: `/` → the page; `/<lib>` → the vendored bundle."""
+    """Localhost server: `/` → BotD+FPJS+fpscanner page, `/creepjs` → CreepJS page,
+    `/<file>` → the vendored bundle."""
 
     def __init__(self):
         page = _PAGE.encode()
+        creep_page = _CREEP_PAGE.encode()
         vendor = _VENDOR
 
         class H(http.server.BaseHTTPRequestHandler):
             def do_GET(self):  # noqa: N802
-                if self.path == "/" or self.path.startswith("/?"):
+                p = self.path.split("?")[0]
+                if p == "/":
                     body, ctype = page, "text/html; charset=utf-8"
+                elif p == "/creepjs":
+                    body, ctype = creep_page, "text/html; charset=utf-8"
                 else:
-                    f = vendor / Path(self.path.lstrip("/")).name
+                    f = vendor / Path(p.lstrip("/")).name
                     if not f.is_file():
                         self.send_error(404); return
                     body = f.read_bytes()
@@ -99,6 +141,10 @@ class _DetectorSite:
     def url(self):
         return f"http://127.0.0.1:{self.port}/"
 
+    @property
+    def creep_url(self):
+        return f"http://127.0.0.1:{self.port}/creepjs"
+
     def close(self):
         self._srv.shutdown()
 
@@ -111,7 +157,7 @@ def detector_site():
 
 
 def _run_detectors(firefox_binary, url):
-    """Launch the binary, load the page, return (botd, fp, err)."""
+    """Launch the binary, load the page, return (botd, fp, fps, err)."""
     with InvisiblePlaywright(seed=42, binary_path=firefox_binary) as browser:
         page = browser.new_page()
         page.goto(url, wait_until="load", timeout=45000)
@@ -121,16 +167,48 @@ def _run_detectors(firefox_binary, url):
         )
         botd = page.evaluate("() => window.__botd")
         fp = page.evaluate("() => window.__fp")
+        fps = page.evaluate("() => window.__fps")
         err = page.evaluate("() => window.__err")
-    return botd, fp, err
+    return botd, fp, fps, err
+
+
+def _run_creepjs(firefox_binary, creep_url):
+    """Launch the binary, run CreepJS fully offline, return its headless result."""
+    _EV = """() => {
+      const f = window.Fingerprint;
+      if (!f || !f.headless) return { ready: false };
+      const h = f.headless;
+      return {
+        ready: true,
+        headlessRating: h.headlessRating,
+        stealthRating: h.stealthRating,
+        likeHeadlessRating: h.likeHeadlessRating,
+        headless: h.headless || {},
+        stealth: h.stealth || {},
+        totalLies: (f.lies && f.lies.totalLies) || 0,
+      };
+    }"""
+    with InvisiblePlaywright(seed=42, binary_path=firefox_binary) as browser:
+        page = browser.new_page()
+        # truly offline: abort every non-loopback request (CreepJS's optional
+        # crowd-comparison POST to arh.antoinevastel.com never runs).
+        page.route(
+            "**/*",
+            lambda r: r.abort() if "127.0.0.1" not in r.request.url else r.continue_(),
+        )
+        page.goto(creep_url, wait_until="domcontentloaded", timeout=45000)
+        page.wait_for_function(
+            "() => !!(window.Fingerprint && window.Fingerprint.headless)",
+            timeout=60000,
+        )
+        return page.evaluate(_EV)
 
 
 @pytest.mark.e2e
 def test_botd_no_detector_flags_automation(firefox_binary, detector_site):
     """The real BotD must not flag the build — aggregate AND every one of its
-    individual detectors (webDriver/userAgent/appVersion/plugins/process/... ).
-    """
-    botd, _fp, err = _run_detectors(firefox_binary, detector_site.url)
+    individual detectors (webDriver/userAgent/appVersion/plugins/process/...)."""
+    botd, _fp, _fps, err = _run_detectors(firefox_binary, detector_site.url)
     assert botd is not None, f"BotD produced no result (err:{err!r})"
     assert botd.get("bot") is False, (
         f"BotD aggregate flagged a bot: botKind={botd.get('botKind')!r}"
@@ -146,8 +224,8 @@ def test_fingerprintjs_visitorid_stable_across_launches(firefox_binary, detector
     """FingerprintJS visitorId must be present and identical across two fresh
     launches with the same seed — a real browser is stable; an over-randomized
     spoof drifts (and a drifting fingerprint is itself a bot tell)."""
-    _b1, fp1, err1 = _run_detectors(firefox_binary, detector_site.url)
-    _b2, fp2, err2 = _run_detectors(firefox_binary, detector_site.url)
+    _b1, fp1, _f1, err1 = _run_detectors(firefox_binary, detector_site.url)
+    _b2, fp2, _f2, err2 = _run_detectors(firefox_binary, detector_site.url)
     assert fp1 and fp1.get("visitorId"), f"no visitorId on run 1 (err:{err1!r})"
     assert fp2 and fp2.get("visitorId"), f"no visitorId on run 2 (err:{err2!r})"
     assert fp1["visitorId"] == fp2["visitorId"], (
@@ -159,13 +237,58 @@ def test_fingerprintjs_visitorid_stable_across_launches(firefox_binary, detector
 @pytest.mark.e2e
 def test_fingerprintjs_collects_rich_fingerprint(firefox_binary, detector_site):
     """FingerprintJS must collect a RICH component surface (a real browser
-    exposes many signals; a stripped/blocked surface is itself suspicious).
-    We don't assert zero errored components (some are legitimately unsupported
-    per browser), only that the surface is substantial and the id computed."""
-    _b, fp, err = _run_detectors(firefox_binary, detector_site.url)
+    exposes many signals; a stripped/blocked surface is itself suspicious)."""
+    _b, fp, _f, err = _run_detectors(firefox_binary, detector_site.url)
     assert fp and fp.get("visitorId"), f"FingerprintJS produced no id (err:{err!r})"
     keys = fp.get("componentKeys") or []
     assert len(keys) >= 15, (
         f"FingerprintJS collected only {len(keys)} components — surface too thin "
         f"(suppressed signals are themselves a tell): {keys}"
     )
+
+
+@pytest.mark.e2e
+def test_fpscanner_no_automation_rules(firefox_binary, detector_site):
+    """fpscanner's engine-agnostic bot rules (webdriver/selenium/bot-UA/platform/
+    timezone/language) must all be clean. The Chrome/GPU-only rules are ignored
+    on purpose (see module docstring) — they false-red on a software-WebGL host."""
+    _b, _fp, fps, err = _run_detectors(firefox_binary, detector_site.url)
+    assert fps is not None, f"fpscanner produced no result (err:{err!r})"
+    details = fps.get("details") or {}
+    assert details, f"fpscanner returned no detection details (err:{err!r})"
+    flagged = [
+        k for k in _FPSCANNER_AGNOSTIC
+        if details.get(k) and details[k].get("detected")
+    ]
+    assert not flagged, (
+        f"fpscanner flagged automation on engine-agnostic rules: {flagged} "
+        f"(full details: { {k: v for k, v in details.items() if v.get('detected')} })"
+    )
+
+
+@pytest.mark.e2e
+def test_creepjs_headless_and_proxy_clean(firefox_binary, detector_site):
+    """CreepJS (Firefox-aware) must see no headless tell and no JS-proxy stealth
+    tell. ``headlessRating`` aggregates webDriverIsOn + headless-UA checks (all
+    GPU-independent). The proxy/runtime stealth sub-signals (hasIframeProxy,
+    hasToStringProxy, hasBadChromeRuntime) must be false — a spoof implemented
+    with a JS Proxy is exactly what CreepJS catches. stealthRating/totalLies/
+    likeHeadlessRating are GPU/theme-sensitive, so we log them, not assert."""
+    r = _run_creepjs(firefox_binary, detector_site.creep_url)
+    assert r and r.get("ready"), f"CreepJS never populated window.Fingerprint: {r!r}"
+    print(
+        f"[creepjs] headlessRating={r['headlessRating']} stealthRating={r['stealthRating']} "
+        f"likeHeadlessRating={r['likeHeadlessRating']} totalLies={r['totalLies']} "
+        f"headless={r['headless']} stealth={r['stealth']}"
+    )
+    assert r["headlessRating"] == 0, (
+        f"CreepJS headless tells fired: headless={r['headless']} "
+        f"(headlessRating={r['headlessRating']})"
+    )
+    stealth = r.get("stealth") or {}
+    proxy_tells = {
+        k: stealth.get(k)
+        for k in ("hasIframeProxy", "hasToStringProxy", "hasBadChromeRuntime")
+        if stealth.get(k)
+    }
+    assert not proxy_tells, f"CreepJS JS-proxy stealth tells fired: {proxy_tells}"

tests/test_fingerprint_surface.py

@@ -27,6 +27,7 @@ Run only this file:
 from __future__ import annotations
 
 import re
+import sys
 
 import pytest
 
@@ -296,6 +297,12 @@ def test_webgl_readpixels_no_masking_signature(page):
     ~300+ 'spikes' and pixelscan flagged it as masking; the gamma remap leaves
     the gradient smooth (~0 spikes). Regression guard for the gamma fix."""
     res = _ev(page, _WEBGL_MASKING_PROBE)
+    if res.get("error") == "no-webgl" and sys.platform == "darwin":
+        pytest.skip(
+            "macOS CI runners expose no WebGL (no software-GL fallback); the gamma "
+            "readPixels remap is platform-agnostic C++ and is exercised by the Linux "
+            "(Xvfb/llvmpipe) and Windows (WARP) gates."
+        )
     assert "error" not in res, f"WebGL probe failed: {res}"
     # genuine / gamma -> ~0; the rejected +-1 algorithm produced ~320.
     assert res["spikes"] < 30, (

tests/test_integration.py

@@ -289,8 +289,10 @@ def test_windows_virtual_display_with_socks_proxy(monkeypatch):
     assert prefs["security.sandbox.gpu.level"] == 0  # virtual_display branch
     assert prefs["network.proxy.type"] == 1          # SOCKS branch
     assert prefs["network.proxy.socks"] == "127.0.0.1"
-    # Windows still has the renderer cleared.
-    assert prefs["zoom.stealth.webgl.renderer"] == ""
+    # Windows exposes a validated persona renderer (calibrated clean bucket),
+    # not empty/native — see _webgl_personas.
+    assert prefs["zoom.stealth.webgl.renderer"].startswith("ANGLE (")
+    assert prefs["zoom.stealth.webgl.renderer"].rstrip().endswith(", D3D11)")
 
 
 # ──────────────────────────────────────────────────────────────────────

tests/test_mouse.py

@@ -132,12 +132,9 @@ def test_mouse_move_outside_viewport_does_not_raise(firefox_binary):
 # ────────────────────────────────────────────────────────────────────
 
 
-@pytest.mark.e2e
-def test_humanize_emits_intermediate_moves(firefox_binary):
-    """A long mouse.move from one corner to another should fire several
-    mousemove events on the page when the humanize hook is enabled (which
-    is the StealthFox default)."""
-    with InvisiblePlaywright(seed=42, binary_path=firefox_binary) as browser:
+def _humanize_move_count(firefox_binary, humanize):
+    """Count page mousemove events fired by ONE long mouse.move."""
+    with InvisiblePlaywright(seed=42, binary_path=firefox_binary, humanize=humanize) as browser:
         page = browser.new_page()
         page.goto(_data_url(
             "<div id=d style='width:600px;height:400px' "
@@ -146,8 +143,27 @@ def test_humanize_emits_intermediate_moves(firefox_binary):
         page.mouse.move(10, 10)
         page.evaluate("window.__n = 0")
         page.mouse.move(500, 300)
-        moves = page.evaluate("window.__n")
-        assert moves >= 1, f"expected at least 1 mousemove event, got {moves}"
+        return page.evaluate("window.__n")
+
+
+@pytest.mark.e2e
+def test_humanize_emits_intermediate_moves(firefox_binary):
+    """A long mouse.move must expand into MANY intermediate mousemove events when
+    humanize is on (Bezier), and ~1 (a teleport) when off. We assert the on/off
+    CONTRAST: `moves >= 1` alone was a false-green — a teleport already fires 1 —
+    and that false-green hid a pref-namespace bug (wrapper wrote
+    `invisible_playwright.humanize`, the binary's Juggler reads `stealthfox.humanize`)
+    that left humanize silently dead in production. This test now fails if the
+    pref ever stops reaching the binary."""
+    on = _humanize_move_count(firefox_binary, True)
+    off = _humanize_move_count(firefox_binary, False)
+    assert off <= 2, f"humanize OFF should ~teleport (<=2 moves), got {off}"
+    assert on >= 4, (
+        f"humanize ON must expand into many intermediate moves (Bezier); got {on} "
+        f"(off={off}). moves==1 means the cursor teleports — the exact automation "
+        f"tell humanize exists to remove, and a sign the stealthfox.* pref isn't "
+        f"reaching the binary's Juggler."
+    )
 
 
 # ────────────────────────────────────────────────────────────────────
@@ -185,8 +201,11 @@ def test_hover_triggers_mouseenter(firefox_binary):
         # Wait for the event rather than reading immediately: under load / on a
         # virtual display the mouseenter can land a beat after hover() returns,
         # which made an instant read flaky. wait_for_function still fails (times
-        # out) if mouseenter genuinely never fires.
-        page.wait_for_function("() => window.__h === true", timeout=5000)
+        # out) if mouseenter genuinely never fires. Timeout is generous (10s) so a
+        # busy full-suite run — where browser startup + CPU contention can push
+        # the event past a tight 5s window — doesn't flake; the event itself fires
+        # in well under a second when run in isolation.
+        page.wait_for_function("() => window.__h === true", timeout=10_000)
 
 
 # ────────────────────────────────────────────────────────────────────

tests/test_prefs.py

@@ -15,12 +15,18 @@ from invisible_playwright.prefs import (
 
 @pytest.mark.unit
 def test_translate_includes_gpu_renderer_windows(monkeypatch):
-    """On Windows, renderer/vendor are cleared so ANGLE reports native hardware."""
+    """On Windows we falsify the GPU to one of the calibrated CLEAN buckets (FP Pro
+    tampering_ml<=0.5 on every seed; sweep 2026-06-14). Only Radeon R9 200 Series and
+    Intel Arc A750 ship — every NVIDIA/iGPU/945 bucket is penalized. See _webgl_personas."""
     monkeypatch.setattr(sys, "platform", "win32")
+    _CLEAN = {
+        "ANGLE (AMD, AMD Radeon R9 200 Series Direct3D11 vs_5_0 ps_5_0, D3D11)",
+        "ANGLE (Intel, Intel(R) Arc(TM) A750 Graphics Direct3D11 vs_5_0 ps_5_0, D3D11)",
+    }
     p = generate_profile(seed=42)
     prefs = translate_profile_to_prefs(p)
-    assert prefs["zoom.stealth.webgl.renderer"] == ""
-    assert prefs["zoom.stealth.webgl.vendor"] == ""
+    assert prefs["zoom.stealth.webgl.renderer"] in _CLEAN
+    assert prefs["zoom.stealth.webgl.vendor"] in {"Google Inc. (AMD)", "Google Inc. (Intel)"}
 
 
 @pytest.mark.unit
@@ -82,10 +88,15 @@ def test_accept_language_underscore_normalized():
 
 
 @pytest.mark.unit
-def test_font_metrics_windows_returns_empty(monkeypatch):
-    # FM2: Windows never applies width-scale factors.
+def test_font_metrics_windows_applies_named_factors(monkeypatch):
+    # FM2: Windows/mac apply the per-NAMED-font factors (so whitelisted named
+    # families don't collapse to the list-head width on the canvas measureText
+    # path), but WITHOUT the Linux generic-family compensation (generics bypass
+    # the whitelist and render native there).
     monkeypatch.setattr(sys, "platform", "win32")
-    assert _font_metrics_for_platform("Arial|1.0,Verdana|0.9,") == ""
+    out = _font_metrics_for_platform("Arial|1.0,Verdana|0.9,")
+    assert out == "Arial|1.0,Verdana|0.9,"
+    assert "sans-serif|" not in out  # no generic compensation on Windows
 
 
 @pytest.mark.unit
@@ -100,13 +111,14 @@ def test_font_metrics_empty_input_returns_empty():
 
 
 @pytest.mark.unit
-def test_gpu_renderer_empty_on_windows(monkeypatch):
-    # PG2
+def test_gpu_renderer_persona_on_windows(monkeypatch):
+    # PG2: Windows exposes a validated persona renderer (well-formed ANGLE bucket, NOT empty/native).
     monkeypatch.setattr(sys, "platform", "win32")
     p = generate_profile(seed=42)
     prefs = translate_profile_to_prefs(p)
-    assert prefs["zoom.stealth.webgl.renderer"] == ""
-    assert prefs["zoom.stealth.webgl.vendor"] == ""
+    r = prefs["zoom.stealth.webgl.renderer"]
+    assert r and r.startswith("ANGLE (") and r.rstrip().endswith(", D3D11)")
+    assert prefs["zoom.stealth.webgl.vendor"].startswith("Google Inc. (")
 
 
 @pytest.mark.unit
@@ -143,13 +155,16 @@ def test_canvas_noise_mask_windows_uses_intel_path(monkeypatch):
 
 
 @pytest.mark.unit
-def test_webgl_extensions_cleared_on_windows(monkeypatch):
-    # WE2
+def test_webgl_extensions_persona_on_windows(monkeypatch):
+    # WE2: with a persona active on Windows, extensions are FORCED to the persona's native-order
+    # list (host-independent), NOT cleared. Order is load-bearing (must match the persona verbatim).
     monkeypatch.setattr(sys, "platform", "win32")
+    from invisible_playwright._webgl_personas import select_persona
     p = generate_profile(seed=42)
     prefs = translate_profile_to_prefs(p)
-    assert prefs["zoom.stealth.webgl.extensions"] == ""
-    assert prefs["zoom.stealth.webgl2.extensions"] == ""
+    persona = select_persona(42)
+    assert prefs["zoom.stealth.webgl.extensions"] == persona["ext1"]
+    assert prefs["zoom.stealth.webgl2.extensions"] == persona["ext2"]
 
 
 # ──────────────────────────────────────────────────────────────────────

tests/test_sampler.py

@@ -45,13 +45,28 @@ def test_classify_gpu_intel_hd_old_buckets(renderer):
     "ANGLE (Intel, Intel(R) HD Graphics 530 Direct3D11)",
     "ANGLE (Intel, Intel(R) UHD Graphics 630 Direct3D11)",
     "ANGLE (Intel, Intel(R) Iris Xe Graphics Direct3D11)",
-    "ANGLE (Intel, Intel(R) Arc A750 Direct3D11)",
+    # Integrated Arc iGPUs (Core Ultra "Arc 130T/140T/Graphics") stay integrated_modern.
+    "ANGLE (Intel, Intel(R) Arc(TM) 140T GPU Direct3D11)",
 ])
 def test_classify_gpu_intel_modern(renderer):
-    """CG4-CG7 [DT]: modern Intel HD/UHD/Iris/Arc → integrated_modern."""
+    """CG4-CG7 [DT]: modern Intel HD/UHD/Iris + integrated Arc → integrated_modern."""
     assert classify_gpu(_gpu(renderer)) == "integrated_modern"
 
 
+@pytest.mark.unit
+@pytest.mark.parametrize("renderer,expected", [
+    # Discrete Intel Arc DESKTOP cards are NOT integrated: A5xx/A7xx/Bxxx ~ mid-range
+    # discrete (RTX 3060 tier); A3xx are entry discrete → low_end.
+    ("ANGLE (Intel, Intel(R) Arc(TM) A750 Graphics Direct3D11 vs_5_0 ps_5_0)", "mid_range"),
+    ("ANGLE (Intel, Intel(R) Arc(TM) A770 Graphics Direct3D11)", "mid_range"),
+    ("ANGLE (Intel, Intel(R) Arc(TM) B580 Graphics Direct3D11)", "mid_range"),
+    ("ANGLE (Intel, Intel(R) Arc(TM) A380 Graphics Direct3D11)", "low_end"),
+])
+def test_classify_gpu_intel_arc_discrete(renderer, expected):
+    """Discrete Intel Arc desktop SKUs map to a discrete-GPU class, not integrated."""
+    assert classify_gpu(_gpu(renderer)) == expected
+
+
 @pytest.mark.unit
 @pytest.mark.parametrize("renderer", [
     "ANGLE (AMD, AMD Radeon Graphics Direct3D11)",

tests/unit/test_config_public.py

@@ -29,8 +29,8 @@ def test_get_default_stealth_prefs_random_seed_returns_dict():
     assert isinstance(prefs, dict)
     assert len(prefs) > 0
     # humanize toggle is always set explicitly
-    assert "invisible_playwright.humanize" in prefs
-    assert prefs["invisible_playwright.humanize"] is True
+    assert "stealthfox.humanize" in prefs
+    assert prefs["stealthfox.humanize"] is True
 
 
 def test_get_default_stealth_prefs_seed_is_deterministic():
@@ -50,22 +50,22 @@ def test_get_default_stealth_prefs_different_seeds_differ():
 def test_humanize_false_disables_prefs():
     """humanize=False removes the maxTime knob and flips the toggle to False."""
     prefs = get_default_stealth_prefs(seed=42, humanize=False)
-    assert prefs["invisible_playwright.humanize"] is False
-    assert "invisible_playwright.humanize.maxTime" not in prefs
+    assert prefs["stealthfox.humanize"] is False
+    assert "stealthfox.humanize.maxTime" not in prefs
 
 
 def test_humanize_default_sets_max_time_1_5():
     """humanize=True -> default maxTime is 1.5s, stored as string."""
     prefs = get_default_stealth_prefs(seed=42, humanize=True)
-    assert prefs["invisible_playwright.humanize"] is True
-    assert prefs["invisible_playwright.humanize.maxTime"] == "1.5"
+    assert prefs["stealthfox.humanize"] is True
+    assert prefs["stealthfox.humanize.maxTime"] == "1.5"
 
 
 def test_humanize_float_overrides_max_time():
     """Float for humanize is the explicit cap in seconds."""
     prefs = get_default_stealth_prefs(seed=42, humanize=3.0)
-    assert prefs["invisible_playwright.humanize"] is True
-    assert prefs["invisible_playwright.humanize.maxTime"] == "3.0"
+    assert prefs["stealthfox.humanize"] is True
+    assert prefs["stealthfox.humanize.maxTime"] == "3.0"
 
 
 def test_extra_prefs_overlay_takes_precedence():

tests/vendor/README.md

@@ -6,13 +6,19 @@ on a dev box and on a GitHub runner (no external CDN at test time — Firefox
 tracking-protection blocks the openfpcdn.io CDN anyway, and we want CI offline).
 
 They are served from a localhost HTTP server and loaded into the patched Firefox;
-the tests assert the REAL detectors don't flag the stealth build (BotD: `bot===false`)
-and that the fingerprint is stable (FingerprintJS: same `visitorId` across launches).
+the tests assert the REAL detectors don't flag the stealth build (BotD: `bot===false`;
+fpscanner: engine-agnostic rules clean; CreepJS: `headlessRating===0` + no JS-proxy
+tells) and that the fingerprint is stable (FingerprintJS: same `visitorId` across
+launches). CreepJS runs fully offline — the tests abort every non-loopback request,
+so its optional crowd-comparison POST never fires and the verdict is computed locally.
 
 | File | Package | Version | Source | License |
 |---|---|---|---|---|
 | `botd-2.0.0.esm.js` | `@fingerprintjs/botd` | 2.0.0 | https://cdn.jsdelivr.net/npm/@fingerprintjs/botd@2.0.0/dist/botd.esm.js | MIT |
 | `fingerprintjs-5.2.0.umd.min.js` | `@fingerprintjs/fingerprintjs` | 5.2.0 | https://cdn.jsdelivr.net/npm/@fingerprintjs/fingerprintjs@5.2.0/dist/fp.umd.min.js | MIT |
+| `fpscanner-1.0.6.es.js` | `fpscanner` | 1.0.6 | https://cdn.jsdelivr.net/npm/fpscanner@1.0.6/dist/fpScanner.es.js | MIT |
+| `creepjs-10aa672.js` | `abrahamjuliot/creepjs` | git `10aa6724` | https://raw.githubusercontent.com/abrahamjuliot/creepjs/10aa6724cd33a1015db1574211890518cd04f0cc/docs/creep.js | MIT |
 
-Both are MIT (Copyright © FingerprintJS, Inc.). To update: download the pinned
-dist from jsdelivr, drop it here, and bump the version in the filename + this table.
+All MIT (FingerprintJS Inc. / Antoine Vastel / Abraham Juliot). To update: download
+the pinned dist (jsdelivr for npm packages, raw.githubusercontent for CreepJS at a
+commit SHA), drop it here, and bump the version in the filename + this table.