@d-zero/beholder 2.1.6 → 3.1.0

package/CHANGELOG.md

@@ -3,6 +3,50 @@
 All notable changes to this project will be documented in this file.
 See [Conventional Commits](https://conventionalcommits.org) for commit guidelines.
 
+# [3.1.0](https://github.com/d-zero-dev/tools/compare/@d-zero/beholder@3.0.0...@d-zero/beholder@3.1.0) (2026-06-17)
+
+### Features
+
+- **beholder:** expose extractMetaFromDocument for jsdom-backed meta extraction ([a56e21c](https://github.com/d-zero-dev/tools/commit/a56e21c17dcc1e542595a596074c5d8e659c1168))
+
+# [3.0.0](https://github.com/d-zero-dev/tools/compare/@d-zero/beholder@2.1.6...@d-zero/beholder@3.0.0) (2026-06-16)
+
+### Bug Fixes
+
+- **beholder:** warn loudly and tripwire-test puppeteer Page.\_client() coverage ([97a07ea](https://github.com/d-zero-dev/tools/commit/97a07ea273e90d50bfede1d68f594ddee9c33268))
+
+- feat(beholder)!: expand meta extraction with frontmatter-keys schema and Wappalyzer tag detection ([6ee7861](https://github.com/d-zero-dev/tools/commit/6ee78617aac3fe3d5c022ccfd0df265de0c5310b))
+
+### Features
+
+- **beholder:** rewrite getAnchorList with single AX tree + parallel describeNode ([#876](https://github.com/d-zero-dev/tools/issues/876)) ([7e5b089](https://github.com/d-zero-dev/tools/commit/7e5b089695bd1e605d63c6faef2e8bf927bd861f))
+
+### BREAKING CHANGES
+
+- `Meta` is restructured from flat keys (`noindex`, `canonical`,
+  `'og:type'`, `'twitter:card'`, ...) into a nested shape backed by
+  `frontmatter-keys.md`. New required fields: `title`, `jsonLd`,
+  `speculationRules`, `originTrial`, `tags`, `others`. `getMeta(page)` now takes
+  a context object `getMeta(page, { url, html?, statusCode?, headers? }, timeout?)`.
+  Old top-level shortcuts (`canonical`, `alternate`, `noindex`, `nofollow`,
+  `noarchive`, `'og:*'`, `'twitter:card'`) are removed; values move to
+  `meta.link.canonical`, `meta.robots.*`, `meta.og.*`, `meta.twitter.*` etc.
+
+Changes:
+
+- New `src/meta/` module: `types.ts`, `keys.ts`, `parsers.ts`, `classify.ts`,
+  `id-extractors.ts`, `tag-detection.ts`, plus ambient `simple-wappalyzer.d.ts`
+- Browser-side `collectHead()` serializes every `<meta>`, `<link>`, structured-data
+  `<script>`, `<base>`, `<iframe>` plus a curated set of `window` globals into
+  `RawHeadEntry[]`; Node-side `classify()` maps these to typed Meta fields
+- `simple-wappalyzer` (MIT) added as a dependency for technology detection;
+  detected providers run through `id-extractors.ts` for real ID extraction
+  (GA4, GTM, UA, FB Pixel, Hotjar, Clarity, ...)
+- Unknown markup is preserved under `Meta.others` (meta/property/httpEquiv/
+  itemprop/link/script/iframe buckets) so nothing is silently dropped
+- Tests: parsers/classify/id-extractors/tag-detection units + getMeta
+  error/timeout fallback
+
 ## [2.1.6](https://github.com/d-zero-dev/tools/compare/@d-zero/beholder@2.1.5...@d-zero/beholder@2.1.6) (2026-06-15)
 
 ### Bug Fixes

package/README.md

@@ -32,3 +32,29 @@ if (result.type === 'success') {
 ```
 
 設計判断（イベントではなく戻り値で返す理由、`page` のライフサイクル責務、リトライ機構など）は `src/scraper.ts` の JSDoc を参照。
+
+## DOM 文字列からメタ抽出（Puppeteer なし）
+
+HTML 文字列を jsdom などでパースしてから `Meta` を取り出したい場合、`extractMetaFromDocument` を使う。`Scraper` が内部で呼ぶ `collectHead → detectTags → classify` パイプラインと同じ実装を再利用するため、戻り値の `Meta` 形状は `scrapeStart` と同一。DOM ライブラリ（jsdom 等）はユーザランドの責務。
+
+```ts
+import { extractMetaFromDocument } from '@d-zero/beholder';
+import { JSDOM } from 'jsdom';
+
+const url = 'https://example.com/';
+const html = await (await fetch(url)).text();
+const dom = new JSDOM(html, { url });
+
+// `as unknown as Window` は jsdom の `DOMWindow` 型が lib.dom の `Window` と
+// 構造的に完全一致しないための型キャスト。ランタイムでは互換。
+const meta = await extractMetaFromDocument(dom.window as unknown as Window, {
+	url,
+	html,
+});
+
+console.log(meta.title);
+console.log(meta.og?.image);
+console.log(meta.tags.entries);
+```
+
+`context.html` を省略すると `window.document.documentElement.outerHTML` がフォールバックされる。ただし Wappalyzer の HTML パターンはスクリプト実行前の生 HTML に合わせて作られているので、可能なら取得直後の HTML 文字列を明示的に渡す方が検出が安定する。

package/dist/dom-evaluation.d.ts

@@ -19,8 +19,12 @@ import type { ElementHandle, Page } from 'puppeteer';
  * Default timeout (ms) applied to DOM evaluation operations when the caller does not
  * specify one. Bounds how long a single `page.evaluate` / property read may hang on a
  * page whose main thread is unresponsive.
+ *
+ * WHY 180s: Aligned with the upstream `Scraper#fetchData` retryable timeout (3 min) so
+ * a single phase does not exceed the retry budget while still tolerating large pages
+ * (e.g., 1000+ anchors) and slow main threads.
  */
-export declare const DEFAULT_DOM_EVALUATION_TIMEOUT = 30000;
+export declare const DEFAULT_DOM_EVALUATION_TIMEOUT = 180000;
 /**
  * Parameters for {@link getProp}.
  * @template T - The expected type of the property value.
@@ -65,35 +69,79 @@ export declare function getImageList(page: Page, viewportWidth: number, timeout?
  * the accessible name (from the accessibility tree, falling back to `textContent`),
  * and filters out non-HTTP links.
  *
- * WHY this keeps per-element CDP calls (unlike {@link getMeta} / {@link getImageList}):
- * the accessible name comes from Chrome's computed accessibility tree
- * (`page.accessibility.snapshot`), which is a CDP-only feature unavailable to in-page
- * DOM APIs. Each {@link getProp} read is still bounded by `timeout`.
+ * WHY Strategy F (single AX-tree fetch + parallel `DOM.describeNode`): the old
+ * implementation called `page.accessibility.snapshot({ root })` per anchor, which
+ * triggers a CDP round-trip *and* a Chrome-side AX subtree computation (~42ms
+ * each). On a page with 1181 anchors that compounded to ~53s. By fetching the
+ * full AX tree once and using `DOM.describeNode` in parallel to map element
+ * handles back to AX nodes by `backendDOMNodeId`, the same data is collected in
+ * ~150ms on the same page — a ~350× speed-up while preserving the original
+ * accessible-name semantics. See issue #876 for measurements.
+ *
+ * WHY the whole operation is wrapped in `raceWithTimeout`: even with bounded
+ * per-CDP-call timeouts, a degenerate page (blocked main thread, thousands of
+ * anchors, runaway describeNode latency) could chain enough sub-timeouts to
+ * exceed the caller's `timeout` budget. The outer race guarantees the function
+ * returns within `timeout`, surfacing whatever anchors were collected so far so
+ * the upstream scrape phase can continue rather than tripping a retryable retry.
  * @param page - The Puppeteer page to extract anchors from.
  * @param options - Optional URL parsing options (e.g., `disableQueries`).
- * @param timeout - Timeout in ms per property read. Defaults to {@link DEFAULT_DOM_EVALUATION_TIMEOUT}.
+ * @param timeout - Total time budget in ms for the whole extraction. Defaults to {@link DEFAULT_DOM_EVALUATION_TIMEOUT}.
  * @returns An array of {@link AnchorData} objects for all HTTP(S) links found on the page.
  */
 export declare function getAnchorList(page: Page, options?: ParseURLOptions, timeout?: number): Promise<AnchorData[]>;
 /**
- * Extracts comprehensive meta information from the page's `<head>`.
+ * Required context for {@link getMeta}. Provided by the scraper from data it
+ * already has on hand (URL it navigated to, response status/headers it received).
+ *
+ * `html` is optional: when omitted, `getMeta` falls back to `page.content()`
+ * to obtain the rendered HTML for the third-party tag detection pass.
+ */
+export type GetMetaContext = {
+    /** The fully resolved URL of the page (after redirects). */
+    readonly url: string;
+    /** Rendered HTML. Falls back to `page.content()` when omitted. */
+    readonly html?: string;
+    /** Response status code, surfaced to the Wappalyzer driver. */
+    readonly statusCode?: number;
+    /** Response headers; case is preserved by the caller, lowercased internally. */
+    readonly headers?: Record<string, string | string[] | undefined>;
+    /**
+     * When `true`, the returned `Meta` includes `_raw: RawHeadEntry[]` for
+     * debugging. Default `false` to keep the serialized payload small.
+     */
+    readonly includeRaw?: boolean;
+};
+/**
+ * Extracts comprehensive metadata from the page.
  *
- * Collects all metadata in a single `page.evaluate` call (14 CDP round-trips
- * collapsed into 1) wrapped in {@link raceWithTimeout}. On timeout (an unresponsive
- * page) a minimal `{ title: '' }` is returned rather than hanging.
+ * Two passes happen in parallel:
+ * 1. Browser-side `collectHead()` serializes every `<meta>`, `<link>`,
+ *    relevant `<script>`, `<base>`, `<noscript>`/`<iframe>` and a curated
+ *    set of `window` globals into a `RawHeadEntry[]`. Node-side `classify()`
+ *    then maps those entries to typed `Meta` fields using the lookup tables
+ *    in `./meta/keys.ts`, with unknown entries preserved in `Meta.others`.
+ * 2. `detectTags()` runs `simple-wappalyzer` over the page HTML to produce
+ *    `Meta.tags` (technology detection + real-ID extraction).
  *
- * Collected metadata:
- * - `title` - The document title.
- * - `lang` - The `lang` attribute of the `<html>` element.
- * - `description` - The `<meta name="description">` content.
- * - `keywords` - The `<meta name="keywords">` content.
- * - `noindex` / `nofollow` / `noarchive` - Parsed from the `<meta name="robots">` directives.
- * - `canonical` - The `<link rel="canonical">` content.
- * - `alternate` - The `<link rel="alternate">` content.
- * - Open Graph tags: `og:type`, `og:title`, `og:site_name`, `og:description`, `og:url`, `og:image`.
- * - `twitter:card` - The Twitter Card type.
- * @param page - The Puppeteer page to extract meta information from.
- * @param timeout - Timeout in ms for the evaluation. Defaults to {@link DEFAULT_DOM_EVALUATION_TIMEOUT}.
- * @returns An object containing all extracted meta properties.
+ * The whole call is wrapped in `raceWithTimeout`. On timeout an empty `Meta`
+ * (with `title: ''` and empty required arrays/objects) is returned.
+ * @param page
+ * @param context
+ * @param timeout
+ * @example
+ * ```ts
+ * const meta = await getMeta(page, {
+ *   url: 'https://example.com/',
+ *   html: await page.content(),
+ *   statusCode: response.status,
+ *   headers: response.headers,
+ * });
+ * console.log(meta.title);                         // <title> text
+ * console.log(meta.og?.image);                     // og:image[] array
+ * console.log(meta.robots?.noindex);               // parsed robots
+ * console.log(meta.tags.detected.Analytics);       // Wappalyzer hits
+ * console.log(meta.tags.entries.find(e => e.provider === 'Google Analytics')?.id);
+ * ```
  */
-export declare function getMeta(page: Page, timeout?: number): Promise<Meta>;
+export declare function getMeta(page: Page, context: GetMetaContext, timeout?: number): Promise<Meta>;

package/dist/dom-evaluation.js

@@ -15,6 +15,9 @@
  */
 import { raceWithTimeout } from '@d-zero/shared/race-with-timeout';
 import { domDetailsLog, domLog } from './debug.js';
+import { classify, emptyMeta } from './meta/classify.js';
+import { WINDOW_GLOBALS_TO_CHECK, collectHeadFromDocument } from './meta/collect-head.js';
+import { detectTags } from './meta/tag-detection.js';
 import { parseUrl } from './parse-url.js';
 const pid = `${process.pid}`;
 const log = domLog.extend(pid);
@@ -23,8 +26,12 @@ const dLog = domDetailsLog.extend(pid);
  * Default timeout (ms) applied to DOM evaluation operations when the caller does not
  * specify one. Bounds how long a single `page.evaluate` / property read may hang on a
  * page whose main thread is unresponsive.
+ *
+ * WHY 180s: Aligned with the upstream `Scraper#fetchData` retryable timeout (3 min) so
+ * a single phase does not exceed the retry budget while still tolerating large pages
+ * (e.g., 1000+ anchors) and slow main threads.
  */
-export const DEFAULT_DOM_EVALUATION_TIMEOUT = 30_000;
+export const DEFAULT_DOM_EVALUATION_TIMEOUT = 180_000;
 /**
  * Retrieves a DOM property value from a Puppeteer element handle with a timeout.
  *
@@ -108,6 +115,114 @@ export async function getImageList(page, viewportWidth, timeout = DEFAULT_DOM_EV
     dLog('Images are: %O', imageList.map((i) => i.src));
     return imageList;
 }
+/**
+ * One-shot warning latch: only the first time `_client()` is missing in a
+ * process do we log the degradation. Subsequent calls stay silent to avoid
+ * spamming logs while every page in a crawl re-enters the fallback path.
+ */
+let warnedAboutMissingClient = false;
+/**
+ * Returns puppeteer's internal CDP session for the page, or `null` if it is
+ * unreachable (e.g., test mocks, puppeteer wrappers that hide the internal API,
+ * or a future puppeteer release that renames `_client`).
+ *
+ * WHY a warning log: callers transparently fall back to textContent-only mode
+ * when this returns `null`, which masks a silent perf regression if a
+ * puppeteer update removes `_client`. The warning makes the degraded state
+ * observable in production logs so a maintainer can patch the access path.
+ *
+ * Callers fall back to a textContent-only path when this returns `null`.
+ * @param page - The Puppeteer page.
+ */
+function getInternalCDPClient(page) {
+    try {
+        const client = page._client?.();
+        if (!client) {
+            if (!warnedAboutMissingClient) {
+                warnedAboutMissingClient = true;
+                log('WARN: puppeteer Page._client() returned no session — getAnchorList ' +
+                    'falls back to textContent-only mode. Verify the installed puppeteer ' +
+                    'version still exposes the internal _client() accessor.');
+            }
+            return null;
+        }
+        return client;
+    }
+    catch (error) {
+        if (!warnedAboutMissingClient) {
+            warnedAboutMissingClient = true;
+            log('WARN: puppeteer Page._client() threw — getAnchorList falls back to ' +
+                'textContent-only mode. Error: %O', error);
+        }
+        return null;
+    }
+}
+/**
+ * Fetches the full accessibility tree once and builds a `backendDOMNodeId → accessibleName`
+ * map covering every AX node that exposes a backend DOM id.
+ *
+ * WHY include every non-ignored node (not just `role === 'link'`): the original
+ * `page.accessibility.snapshot({ root })` returned whatever AX node represented
+ * the anchor — including anchors whose computed role was overridden via ARIA
+ * (e.g., `<a role="button">`). Mapping every non-ignored node preserves that.
+ *
+ * WHY skip `ignored === true`: puppeteer's high-level snapshot uses
+ * `interestingOnly: true` by default and returns `null` for ignored nodes
+ * (aria-hidden, display:none, visibility:hidden). The old code then fell back
+ * to `textContent.trim()`. Including ignored nodes here would short-circuit
+ * that fallback with the AX tree's empty name and silently drop link text.
+ *
+ * On timeout or CDP failure, an empty map is returned so callers transparently
+ * fall back to `textContent.trim()` for every anchor.
+ * @param client - The CDP session attached to the page.
+ * @param timeout - Maximum time to wait for the AX tree fetch.
+ */
+async function buildAccessibleNameMap(client, timeout) {
+    const { result, timeout: timedOut } = await raceWithTimeout(() => client
+        .send('Accessibility.getFullAXTree')
+        .then((res) => res)
+        .catch((error) => {
+        log('Accessibility.getFullAXTree failed: %O', error);
+        return null;
+    }), timeout);
+    const map = new Map();
+    if (timedOut) {
+        log('Accessibility.getFullAXTree timed out after %dms', timeout);
+        return map;
+    }
+    if (!result?.nodes) {
+        return map;
+    }
+    for (const node of result.nodes) {
+        if (node.backendDOMNodeId == null || node.ignored === true) {
+            continue;
+        }
+        const name = typeof node.name?.value === 'string' ? node.name.value : '';
+        map.set(node.backendDOMNodeId, name);
+    }
+    return map;
+}
+/**
+ * Resolves a CDP backend node id for a given element handle.
+ *
+ * Wrapped in {@link raceWithTimeout} so a single hung `DOM.describeNode` cannot
+ * stall the outer `Promise.all` over every anchor on the page.
+ * @param client - The CDP session attached to the page (must be the same session
+ *                 that owns the handle's `objectId`).
+ * @param objectId - The remote object id of the element handle.
+ * @param timeout - Maximum time to wait for the describeNode call.
+ * @returns The backend node id, or `null` if unavailable / timed out / failed.
+ */
+async function resolveBackendNodeId(client, objectId, timeout) {
+    const { result, timeout: timedOut } = await raceWithTimeout(() => client
+        .send('DOM.describeNode', { objectId })
+        .then((res) => res)
+        .catch(() => null), timeout);
+    if (timedOut || !result) {
+        return null;
+    }
+    return result.node?.backendNodeId ?? null;
+}
 /**
  * Extracts all anchor (`<a>` and `<area>`) elements with `href` attributes from the page.
  *
@@ -115,104 +230,215 @@ export async function getImageList(page, viewportWidth, timeout = DEFAULT_DOM_EV
  * the accessible name (from the accessibility tree, falling back to `textContent`),
  * and filters out non-HTTP links.
  *
- * WHY this keeps per-element CDP calls (unlike {@link getMeta} / {@link getImageList}):
- * the accessible name comes from Chrome's computed accessibility tree
- * (`page.accessibility.snapshot`), which is a CDP-only feature unavailable to in-page
- * DOM APIs. Each {@link getProp} read is still bounded by `timeout`.
+ * WHY Strategy F (single AX-tree fetch + parallel `DOM.describeNode`): the old
+ * implementation called `page.accessibility.snapshot({ root })` per anchor, which
+ * triggers a CDP round-trip *and* a Chrome-side AX subtree computation (~42ms
+ * each). On a page with 1181 anchors that compounded to ~53s. By fetching the
+ * full AX tree once and using `DOM.describeNode` in parallel to map element
+ * handles back to AX nodes by `backendDOMNodeId`, the same data is collected in
+ * ~150ms on the same page — a ~350× speed-up while preserving the original
+ * accessible-name semantics. See issue #876 for measurements.
+ *
+ * WHY the whole operation is wrapped in `raceWithTimeout`: even with bounded
+ * per-CDP-call timeouts, a degenerate page (blocked main thread, thousands of
+ * anchors, runaway describeNode latency) could chain enough sub-timeouts to
+ * exceed the caller's `timeout` budget. The outer race guarantees the function
+ * returns within `timeout`, surfacing whatever anchors were collected so far so
+ * the upstream scrape phase can continue rather than tripping a retryable retry.
  * @param page - The Puppeteer page to extract anchors from.
  * @param options - Optional URL parsing options (e.g., `disableQueries`).
- * @param timeout - Timeout in ms per property read. Defaults to {@link DEFAULT_DOM_EVALUATION_TIMEOUT}.
+ * @param timeout - Total time budget in ms for the whole extraction. Defaults to {@link DEFAULT_DOM_EVALUATION_TIMEOUT}.
  * @returns An array of {@link AnchorData} objects for all HTTP(S) links found on the page.
  */
 export async function getAnchorList(page, options, timeout = DEFAULT_DOM_EVALUATION_TIMEOUT) {
     log('Getting anchors');
     const $anchors = await page.$$('a[href], area[href]');
-    const anchorList = [];
-    for (const $anchor of $anchors) {
-        const $href = await getProp({ $el: $anchor, propName: 'href', fallback: '' }, timeout);
-        const hrefVal = $href.toString();
-        const href = parseUrl(hrefVal, options);
+    if ($anchors.length === 0) {
+        log('Got 0 anchors');
+        return [];
+    }
+    const collected = [];
+    let axHits = 0;
+    let textFallbacks = 0;
+    // Set after the overall race trips so in-flight `resolveAnchor` calls can
+    // short-circuit instead of continuing to consume CDP capacity and pushing
+    // late entries into the already-returned `collected` array.
+    let cancelled = false;
+    const work = async () => {
+        const client = getInternalCDPClient(page);
+        if (cancelled)
+            return;
+        const nameByBackendId = client
+            ? await buildAccessibleNameMap(client, timeout)
+            : new Map();
+        if (cancelled)
+            return;
+        await Promise.all($anchors.map(async ($anchor) => {
+            if (cancelled)
+                return;
+            const resolved = await resolveAnchor($anchor, client, nameByBackendId, options, timeout);
+            if (cancelled || !resolved) {
+                return;
+            }
+            if (resolved.source === 'ax') {
+                axHits++;
+            }
+            else {
+                textFallbacks++;
+            }
+            collected.push(resolved.anchor);
+        }));
+    };
+    const { timeout: timedOut } = await raceWithTimeout(work, timeout);
+    cancelled = true;
+    if (timedOut) {
+        log('getAnchorList timed out after %dms; returning %d anchors collected so far', timeout, collected.length);
+    }
+    // Snapshot so post-return mutations from any in-flight Promise.all callback
+    // (already gated by `cancelled`, but not synchronously cancellable) cannot
+    // alter the array the caller now holds.
+    const result = [...collected];
+    log('Got %d anchors (%d via AX, %d via textContent)', result.length, axHits, textFallbacks);
+    dLog('Anchors are: %O', result.map((a) => a.href.href));
+    return result;
+}
+/**
+ * Resolves a single anchor handle into an {@link AnchorData} entry, or `null`
+ * if the anchor's href is not an HTTP(S) URL.
+ *
+ * Fires `getProp(href)` and `DOM.describeNode` in parallel, then looks up the
+ * accessible name from the pre-built AX map. If the anchor is not represented
+ * in the AX map (or CDP is unavailable), falls back to a lazy `textContent`
+ * fetch — only paying the extra CDP round-trip when actually needed.
+ * @param $anchor - The Puppeteer element handle for an anchor element.
+ * @param client - The shared CDP session, or `null` if unavailable.
+ * @param nameByBackendId - Map from `backendDOMNodeId` to accessible name.
+ * @param options - URL parsing options.
+ * @param timeout - Per-CDP-call timeout in ms.
+ * @returns The resolved anchor along with the name source, or `null` when the
+ *          anchor's href is not crawlable.
+ */
+async function resolveAnchor($anchor, client, nameByBackendId, options, timeout) {
+    try {
+        const objectId = $anchor.remoteObject().objectId;
+        const [hrefVal, backendNodeId] = await Promise.all([
+            getProp({ $el: $anchor, propName: 'href', fallback: '' }, timeout),
+            client && objectId != null
+                ? resolveBackendNodeId(client, objectId, timeout)
+                : Promise.resolve(null),
+        ]);
+        const href = parseUrl(hrefVal.toString(), options);
         if (!href || !href.isHTTP) {
-            continue;
+            return null;
+        }
+        const axName = backendNodeId == null ? undefined : nameByBackendId.get(backendNodeId);
+        if (axName !== undefined) {
+            return { anchor: { href, textContent: axName }, source: 'ax' };
         }
-        const axNode = await page.accessibility.snapshot({ root: $anchor });
         const textContent = await getProp({ $el: $anchor, propName: 'textContent', fallback: '' }, timeout);
-        const accessibleName = axNode ? axNode.name || '' : textContent.trim();
-        const link = {
-            href,
-            textContent: accessibleName,
-        };
-        anchorList.push(link);
-    }
-    log('Got %d anchors', anchorList.length);
-    dLog('Anchors are: %O', anchorList.map((a) => a.href.href));
-    return anchorList;
+        return { anchor: { href, textContent: textContent.trim() }, source: 'text' };
+    }
+    catch (error) {
+        // `remoteObject()` (and other synchronous handle accesses) can throw when
+        // the handle is disposed (page navigated mid-extraction). Drop just this
+        // anchor rather than poisoning the Promise.all over every other anchor.
+        dLog('resolveAnchor failed for an anchor: %O', error);
+        return null;
+    }
 }
 /**
- * Extracts comprehensive meta information from the page's `<head>`.
- *
- * Collects all metadata in a single `page.evaluate` call (14 CDP round-trips
- * collapsed into 1) wrapped in {@link raceWithTimeout}. On timeout (an unresponsive
- * page) a minimal `{ title: '' }` is returned rather than hanging.
- *
- * Collected metadata:
- * - `title` - The document title.
- * - `lang` - The `lang` attribute of the `<html>` element.
- * - `description` - The `<meta name="description">` content.
- * - `keywords` - The `<meta name="keywords">` content.
- * - `noindex` / `nofollow` / `noarchive` - Parsed from the `<meta name="robots">` directives.
- * - `canonical` - The `<link rel="canonical">` content.
- * - `alternate` - The `<link rel="alternate">` content.
- * - Open Graph tags: `og:type`, `og:title`, `og:site_name`, `og:description`, `og:url`, `og:image`.
- * - `twitter:card` - The Twitter Card type.
- * @param page - The Puppeteer page to extract meta information from.
- * @param timeout - Timeout in ms for the evaluation. Defaults to {@link DEFAULT_DOM_EVALUATION_TIMEOUT}.
- * @returns An object containing all extracted meta properties.
+ * Extracts comprehensive metadata from the page.
+ *
+ * Two passes happen in parallel:
+ * 1. Browser-side `collectHead()` serializes every `<meta>`, `<link>`,
+ *    relevant `<script>`, `<base>`, `<noscript>`/`<iframe>` and a curated
+ *    set of `window` globals into a `RawHeadEntry[]`. Node-side `classify()`
+ *    then maps those entries to typed `Meta` fields using the lookup tables
+ *    in `./meta/keys.ts`, with unknown entries preserved in `Meta.others`.
+ * 2. `detectTags()` runs `simple-wappalyzer` over the page HTML to produce
+ *    `Meta.tags` (technology detection + real-ID extraction).
+ *
+ * The whole call is wrapped in `raceWithTimeout`. On timeout an empty `Meta`
+ * (with `title: ''` and empty required arrays/objects) is returned.
+ * @param page
+ * @param context
+ * @param timeout
+ * @example
+ * ```ts
+ * const meta = await getMeta(page, {
+ *   url: 'https://example.com/',
+ *   html: await page.content(),
+ *   statusCode: response.status,
+ *   headers: response.headers,
+ * });
+ * console.log(meta.title);                         // <title> text
+ * console.log(meta.og?.image);                     // og:image[] array
+ * console.log(meta.robots?.noindex);               // parsed robots
+ * console.log(meta.tags.detected.Analytics);       // Wappalyzer hits
+ * console.log(meta.tags.entries.find(e => e.provider === 'Google Analytics')?.id);
+ * ```
  */
-export async function getMeta(page, timeout = DEFAULT_DOM_EVALUATION_TIMEOUT) {
+export async function getMeta(page, context, timeout = DEFAULT_DOM_EVALUATION_TIMEOUT) {
     log('Getting Meta');
-    const { result, timeout: timedOut } = await raceWithTimeout(() => page
-        .evaluate(() => {
-        /* global document, HTMLMetaElement, HTMLLinkElement */
-        const content = (selector) => {
-            const el = document.querySelector(selector);
-            return el instanceof HTMLMetaElement ? el.content : '';
-        };
-        const linkHref = (selector) => {
-            const el = document.querySelector(selector);
-            return el instanceof HTMLLinkElement ? el.href : '';
-        };
-        return {
-            title: document.title,
-            lang: document.documentElement.lang,
-            description: content('meta[name="description"]'),
-            keywords: content('meta[name="keywords"]'),
-            robots: content('meta[name="robots"]'),
-            canonical: linkHref('link[rel="canonical"]'),
-            alternate: linkHref('link[rel="alternate"]'),
-            'og:type': content('meta[property="og:type"]'),
-            'og:title': content('meta[property="og:title"]'),
-            'og:site_name': content('meta[property="og:site_name"]'),
-            'og:description': content('meta[property="og:description"]'),
-            'og:url': content('meta[property="og:url"]'),
-            'og:image': content('meta[property="og:image"]'),
-            'twitter:card': content('meta[name="twitter:card"]'),
-        };
-    })
-        .catch(() => null), timeout);
+    const { result, timeout: timedOut } = await raceWithTimeout(() => runGetMeta(page, context), timeout);
     if (timedOut || result == null) {
         log('Meta extraction timed out or failed; returning fallback');
-        return { title: '' };
-    }
-    const { robots: robotsVal, ...rest } = result;
-    const robots = new Set(robotsVal.split(',').map((robot) => robot.trim().toLowerCase()));
-    const meta = {
-        ...rest,
-        noindex: robots.has('noindex'),
-        nofollow: robots.has('nofollow'),
-        noarchive: robots.has('noarchive'),
-    };
+        return emptyMeta();
+    }
     log('Got meta');
-    dLog('Meta data are: %O', meta);
-    return meta;
+    dLog('Meta data are: %O', result);
+    return result;
+}
+/**
+ *
+ * @param page
+ * @param context
+ */
+async function runGetMeta(page, context) {
+    try {
+        const rawPromise = collectHeadOnPage(page);
+        const htmlPromise = context.html === undefined
+            ? page.content().catch(() => '')
+            : Promise.resolve(context.html);
+        const [raw, html] = await Promise.all([rawPromise, htmlPromise]);
+        const tags = await detectTags({
+            url: context.url,
+            html,
+            ...(context.statusCode === undefined ? {} : { statusCode: context.statusCode }),
+            ...(context.headers === undefined ? {} : { headers: context.headers }),
+        });
+        return classify(raw, {
+            tags,
+            ...(context.includeRaw ? { includeRaw: true } : {}),
+        });
+    }
+    catch (error) {
+        log('runGetMeta failed: %O', error);
+        return null;
+    }
+}
+/**
+ * Collects raw `<head>` entries from a Puppeteer page by injecting
+ * {@link collectHeadFromDocument} into the page realm.
+ *
+ * WHY string-eval instead of `page.evaluate(fn, args)`: the shared
+ * implementation lives in this module (`collectHeadFromDocument`), and a
+ * `page.evaluate(() => collectHeadFromDocument(window, …))` wrapper cannot
+ * reach that module-scope binding inside the page realm — only the wrapper's
+ * own source crosses the CDP boundary. Serializing the implementation via
+ * `Function.prototype.toString` and invoking it through
+ * `page.evaluate(string)` is what keeps the Puppeteer path and the
+ * jsdom path on one source of truth.
+ *
+ * The same {@link collectHeadFromDocument} function is also exposed via
+ * {@link ../extract-meta.ts | extractMetaFromDocument} for jsdom/Node callers,
+ * so the two paths cannot drift apart.
+ * @param page - The Puppeteer page whose document will be inspected.
+ */
+async function collectHeadOnPage(page) {
+    const fnSource = collectHeadFromDocument.toString();
+    const globalsLiteral = JSON.stringify(WINDOW_GLOBALS_TO_CHECK);
+    const expr = `(${fnSource})(window, ${globalsLiteral})`;
+    const raw = await page.evaluate(expr).catch(() => []);
+    return raw;
 }