feat(douyin): add search command for keyword video search

[Daily-AC]

Motivation

The douyin adapter ships 13 subcommands today (activities, hashtag, publish, stats, ...) but no keyword video search. hashtag search --keyword returns topic/challenge rows, not videos, so anything that wants "find me douyin videos for a query" has nowhere to go. This came up while wiring omnireach — its multi-source search needed a tiktok-mirror command for the China-region equivalent (see issue Daily-AC/omnireach#12).

This PR adds opencli douyin search <query> against www.douyin.com.

Output shape (tiktok-aligned)

{"rank": 1, "desc": "...", "author": "...", "url": "...", "plays": 0, "likes": 0, "comments": 0, "shares": 0}

Same eight columns as tiktok search, same order. Downstream tools that already normalize tiktok search rows can ingest douyin rows without per-adapter mapping.

Caveat on counters: plays, comments, shares are exposed as 0 because the search result card markup only surfaces the like count. Clients that need the full counter set should call /aweme/v1/web/aweme/detail/?aweme_id=<id> for the relevant id parsed from url. Surfacing 0 is preferable to fabricating values or dropping the columns (which would break tiktok-shape consumers).

Prerequisites

• Chrome profile bound to OpenCLI must be logged in to https://www.douyin.com. Anonymous visitors get a results page with no rendered cards (no visible login prompt either — it just silently empty); the adapter detects this and raises AuthRequiredError. Log in once via Chrome and the cookies persist.

Implementation strategy: DOM extraction (post-rewrite)

The initial draft used Strategy.INTERCEPT to capture the SPA's signed /aweme/v1/web/general/search/single/ XHR. Live testing on a logged-in profile revealed this never works:

• wait xhr "general/search/single" times out at 20s even though [data-e2e="scroll-list"] li has 16 rendered result cards in the DOM.
• Direct synthesis of the XHR from page context (with the full SPA param set: device_platform=webapp, aid=6383, pc_client_type=1, version_code, etc.) returns status_code: 0, data: [], search_nil_info.search_nil_type: "verify_check" — the endpoint validates an a_bogus signature that lives in the SPA bundle, and bare URLs miss it.
• The SPA renders results into <ul data-e2e="scroll-list"> server-side during initial navigation. The data is in the HTML at navigation time; no client-side fetch is required.

Final approach (clis/douyin/search.js):

1. Navigate to https://www.douyin.com/search/<urlencoded>?type=video.
2. page.evaluate a MutationObserver-backed waiter that resolves to one of {state: 'rendered', cards} / {state: 'login_wall'} / {state: 'timeout'} within a 15s budget.
3. For each <li> inside [data-e2e="scroll-list"], harvest the row's <a href*="/video/"> and the ordered list of leaf-element textContents. The classnames are obfuscated and churn between Douyin builds, so we pin only data-e2e selectors and identify fields by shape:
   • ^\d{1,2}:\d{2}(:\d{2})?$ → duration (skipped)
   • ^\d+(\.\d+)?[万亿]?$ → likes (with parseDouyinCount handling 万/亿)
   • text equal to @, followed by → author
   • longest remaining text → desc
4. Project to the tiktok-aligned schema; plays/comments/shares stay 0 per the caveat above.

This survives classname obfuscation, doesn't need request signing, and matches the pattern xiaohongshu/rednote use for the same problem class.

Live test evidence: VERIFIED on logged-in profile

5 queries × --limit 5, all return 5 well-formed rows. Top-row samples (full JSON arrays validated locally; trimming to 1-2 rows per query for review):

Query 1: opencli douyin search "AI 编程" --limit 5 -f json

[
  {"rank": 1, "desc": "有了AI编程程序员还有出路吗？ #科技科普 #计算机 #AI #编程 #程序员", "author": "英雄编程", "url": "https://www.douyin.com/video/7573894888656293163", "plays": 0, "likes": 17000, "comments": 0, "shares": 0},
  {"rank": 4, "desc": "致全地球人，AI编程速成指南，代码改变命运！ #AI在抖音  #AI编程  #编程入门#编程学习   #计算机专业", "author": "数字游牧人Samuel", "url": "https://www.douyin.com/video/7462667529274592547", "plays": 0, "likes": 35000, "comments": 0, "shares": 0}
]

Query 2: opencli douyin search "claude code" --limit 5 -f json

[
  {"rank": 1, "desc": "全网最全！60分钟全面掌握Claude Code～ 【附完整文档】\n#AI #秋芝2046 #ClaudeCode #AI教程 #前沿科技趋势发布月", "author": "秋芝2046", "url": "https://www.douyin.com/video/7636497165430394162", "plays": 0, "likes": 40000, "comments": 0, "shares": 0},
  {"rank": 3, "desc": "Claude Code 零基础终极教程！ ... #ai新星计划 #claudecode  #ai教程  #claude #智能体", "author": "木子不写代码", "url": "https://www.douyin.com/video/7636872072064470298", "plays": 0, "likes": 408000, "comments": 0, "shares": 0}
]

Query 3: opencli douyin search "周杰伦演唱会现场" --limit 5 -f json

[
  {"rank": 1, "desc": "#周杰伦甜甜的南宁 嘉年华2 南宁站 day2 I Do  4k超清#演唱会现场 #神级现场 #神级live现场 #live现场", "author": "SecretXSPEC", "url": "https://www.douyin.com/video/7640855326389263801", "plays": 0, "likes": 1671, "comments": 0, "shares": 0},
  {"rank": 4, "desc": "周杰伦2010年超时代世界巡回演唱会", "author": "清心影视", "url": "https://www.douyin.com/video/7032189958270045470", "plays": 0, "likes": 11000, "comments": 0, "shares": 0}
]

Query 4: opencli douyin search "#美食探店" --limit 5 -f json (hashtag form — works the same as plaintext)

[
  {"rank": 1, "desc": "特厨探店|不惊艳，但是吃着超级舒服的家常菜—老太婆家常菜 #美食探店 #隋坡 #美食 #黄山", "author": "特厨隋坡（重新出发版）", "url": "https://www.douyin.com/video/7643888049474063631", "plays": 0, "likes": 96000, "comments": 0, "shares": 0},
  {"rank": 2, "desc": "一锅三吃太夯了！被万象城这家贵州烙锅惊艳到了#美食探店#长沙美食#贵州烙锅湖南首店#长沙正宗贵州烙锅#黔寨寨贵州烙锅", "author": "Seven不七亏", "url": "https://www.douyin.com/video/7643712014329827950", "plays": 0, "likes": 6091, "comments": 0, "shares": 0}
]

Query 5: opencli douyin search "rust" --limit 5 -f json

[
  {"rank": 1, "desc": "【Rust腐蚀】有史以来最棒的开荒之旅！ #steam游戏 #多人联机 #生存游戏", "author": "Bone骨头碎片", "url": "https://www.douyin.com/video/7472103372993105161", "plays": 0, "likes": 57000, "comments": 0, "shares": 0},
  {"rank": 3, "desc": "Rust大讲堂 #1 开局的思路详解，新手必看，老手一起交流 #Rust #腐蚀 #代号#前哨 #新手教程", "author": "Jullseye", "url": "https://www.douyin.com/video/7481895809458474240", "plays": 0, "likes": 4552, "comments": 0, "shares": 0}
]

Coverage observed: English ("rust", "claude code"), CJK ("AI 编程", "周杰伦演唱会现场"), hashtag form ("#美食探店"). No verify_check rejections, no holdouts, no anti-bot drift in this run. Like counts span 8 to 408k, confirming the 万/亿 parser handles both small-creator and viral rows.

Tests

• 30 unit tests in clis/douyin/search.test.js covering: arg validation, parseSearchLimit bounds, parseDouyinCount over 万/亿/plain/empty/non-string, normalizeDouyinVideoUrl over scheme-relative/absolute/empty, projectCard over classname-agnostic leaf-text shapes (duration skip, like-count parsing, author-after-@ extraction, longest-text fallback for desc, fused @author prefix stripping, safe defaults for missing leafTexts), full func() flow including login-wall mapping, timeout-to-AuthRequired mapping, {session, data} envelope unwrap, malformed-payload CommandExecutionError, --limit cap, and Chinese URL encoding.
• npm test (442 files / 4664 tests / 1 skipped): all pass.
• npm run typecheck: clean.
• opencli validate: PASS (14 douyin commands, 0 errors, 0 warnings).

Files

• clis/douyin/search.js (new) — DOM-extraction adapter
• clis/douyin/search.test.js (new) — 30 unit tests
• cli-manifest.json — regenerated entry, column list [rank, desc, author, url, plays, likes, comments, shares]