- refetch.js: fetch one article via its source recipe on the site-shared
  browser instance; refresh 原文.md in place for already-archived articles
  (registry ownership-checked, index rebuilt from registry) and re-push
  via push.js --url; new URLs archive to today's dir pending summaries.
  Busy site locks are respected; only self-launched instances are recycled.
- push.js: --url <articleUrl> re-pushes a single article (query/hash
  normalized match, ledger ignored for it); recordResult now appends a
  ledger entry when the article postdates the ledger instead of silently
  dropping a successful push result (refetch-into-finalized-day case).
- harvest.js: export downloadImage/normalizeUrl for refetch.

- login-tools.js: 无 steps 的检测型配方(reuters/wsj)在 login 命令下不再抛
  'recipe.steps is not iterable',改为提示走面板人工登录(exit 6)
- keepalive.js: 用站点共享实例访问一次首页刷新反爬(DataDome)cookie,
  维持常驻访客画像;实例离线则拉起并在结束后回收(有采集锁时不回收);
  检测撞墙写状态文件(NH_KEEPALIVE_STATUS),撞墙 exit 3

Polls site login state for the panel's 人工登录 flow: opens the login page
(PUT /json/new with Page.navigate fallback — this Chromium build ignores
the ?url= param), probes via existing recipe or heuristics (visible
sign-in elements absent + account-testid element present), and on first
successful login can auto-generate a detection-only recipe (--write,
unique data-testid candidates only). JSON line output; exit 0 logged-in,
3 timeout.

Detection-only (checkUrl reuters.com + loggedInSelector
[data-testid=AccountButton], measured on a live logged-in session;
no steps — reuters sign-in lives on account.reuters.com, manual login
in the site instance window persists in the site-shared profile)

- logins/wsj.json: detection-only recipe (checkUrl + loggedOutSelector
  sso.accounts.wsj.com link, no steps — DataDome blocks auto-login); preflight
  now reports real wsj login state instead of 无登录配方
- scripts/instance-quit.js: idempotent browser instance recycle (close all
  tabs → Browser.close fallback); called by harvest.sh EXIT trap

- recognize Auth0-style type=text username fields (not just type=email)
- fill via native value setter so React controlled inputs register the
  value (Auth0 'email required' despite filled field)
- bootstrap checkUrl tab when no site tab open before running steps
- entry picker prefers exact 'Sign In'/'Log In' over long matches
  (was clicking 'Bloomberg Anywhere Remote Login')
- iframe context matching skips the main frame first (same-host login
  iframes like bloomberg.com/account/iframe/signin were executing in
  the main frame and 'seeing' no form)
- new site recipes: statista.json (Auth0 two-step, verified live),
  bloomberg.json (two-step; auto-login still needs trusted-click
  engine support, signin iframe swallows synthetic DOM clicks)

- login-tools: loadRecipe falls back to site recipe (logins/<site>.json);
  discover writes site-level recipe so nikkei-* channels share one recipe
- rename recipes to site names: nikkei.json, digitimes.json (ft.json already)

- discover <source>: open homepage → find login entry by visible text
  (fallback to common login paths) → click/open → detect main-frame vs
  iframe form (captcha/challenge iframes excluded) → enumerate fields →
  generate recipe (single-page / two-step / password-only) with
  remember-me checkbox ticking → write logins/<source>.json → optional
  live login verification (--no-login to skip)
- loggedOutText: login-state check by visible button/link text
  (regex), for sites whose login control has no stable class
- honest boundaries: CAPTCHA/OTP/perimeter protection exits with a
  clear human-required message and writes nothing (validated against
  wsj's DataDome layer); validated green path against nikkei Piano ID

- engine: step.frame executes fill/submit in the matching iframe's
  execution context (cross-origin auth widgets, e.g. Piano ID)
- engine: step.click {text} clicks a main-frame button by visible text
  (opens login modals); loggedOutSelector = logged-in when absent
- logins/nikkei-world.json: Piano two-step form (email -> password in
  auth.asia.nikkei.com iframe), validated with a real login

The only checkbox on join.asp is SaveID (remember email); extend the
tick pattern with save|. Session TTL on this site is short and there is
no keep-me-signed-in option, so preflight-before-harvest is the primary
mechanism; form login frequency stays at most once per harvest.

- submit button now resolved within the filled field's form first
  (document-wide query hit the header search form's identical
  button[type=submit] — clicked Search instead of Next)
- fill/checks/submit merged into one page-side step; checkbox ticking
  with label/name/id pattern filter (keep|remember|signed|记住)
- tab tracking by target id (open→use→close), immune to stale/concurrent
  same-site tabs; verification opens its own tab and closes it
- digitimes recipe: explicit Next button selector + keep-me-signed-in
  checkbox for long-lived session cookie

- check <source>: open checkUrl, assert loggedInSelector (exit 3 when
  logged out); login <source>: recipe-driven multi-step form flow with
  $USER/$PASS substitution from NH_LOGIN_* env, tab-relocation hardening
  against concurrent harvests, post-login verification
- recipes live in logins/<source>.json; ships digitimes-feed.json
- honors CDP_PORT_OVERRIDE / CDP_PROFILE_OVERRIDE (per-source instances)

Pair with CDP_PORT_OVERRIDE so parallel harvests can each drive a
dedicated Chromium instance (own port + own user-data-dir).

- pack.js: --pattern (articleUrlPattern) for direct sources; recipe no
  longer required (and forbidden) for --method direct, since a recipe
  forces the browser path at harvest time
- source-manage.js install: carry bundle articleUrlPattern into config
- fetcher-direct.js: decode HTML entities in title/img alt
- new wccftech-news direct bundle (server-rendered WordPress, no login,
  no Chromium); verified end-to-end via harvest --preview
- docs: SKILL.md, recipe-guide.md, source-management.md updated