The shape of stardust/current/_brand-extraction.json and the procedure that produces it. Run once per extract invocation, after the capture phases have finished. Aggregation may accumulate incrementally as concurrent page captures complete (extract/SKILL.md § Concurrency); the file itself is written once, when every page has landed.

Some fields are sourced from the home page only (logo, voice samples, hero-specific copy). Others must be aggregated across all extracted pages to avoid the home-page bias documented in § Aggregation scope.

The brand surface is the descriptive ground truth for what the existing site looks and feels like at the brand level. It is the input both to direct authoring of stardust/current/PRODUCT.md and DESIGN.md (Phase 4 of extract) and to the direct sub-command when it later reasons about how far to move from the current brand.

Stardust does not invent brand-surface values. Every value is either captured directly from the page or aggregated from captured values, and carries a source citation.

File shape

{
  "_provenance": {
    "writtenBy": "stardust:extract",
    "writtenAt": "2026-04-25T13:50:00Z",
    "readArtifacts": [
      "https://example.com/",
      "stardust/current/pages/home.json"
    ],
    "synthesizedInputs": [],
    "stardustVersion": "0.10.0"
  },
  "site": {
    "name": "Example",
    "tagline": "...",
    "originUrl": "https://example.com"
  },
  "origins": [ /* see § Origins */ ],
  "logo": { /* see § Logo */ },
  "palette": [ /* see § Palette */ ],
  "type": { /* see § Type */ },
  "spacing": { /* see § Spacing */ },
  "motifs": { /* see § Motifs */ },
  "componentStyle": { /* see § Component style */ },
  "systemComponents": [ /* see § System components */ ],
  "iconFont": null,                // null | { /* see § Icon font */ }
  "voice": { /* see § Voice */ },
  "voiceTable": { /* see § Voice table */ },
  "crossPromo": { /* see § Cross-promo */ },
  "register": "brand"             // "brand" | "product" | "ambiguous"
}

§ Aggregation scope

Each field below is sourced from one of two scopes:

scope	fields	rationale
home-only	`logo`, `voice.heroHeadline`, `voice.heroSubcopy`, `voice.heroImage`, `voice.heroMedium`, `voice.primaryCTALabel`, `voice.firstParagraph`, `register`	These are the landing surface; aggregating across pages would dilute, not clarify
cross-page	`palette`, `type` (sizes, weights, families used), `spacing`, `motifs.borderRadius`, `motifs.shadows`, `motifs.gradients`, `componentStyle`, `voice.ctaSamples`, `voice.navItems`, `voice.footerHeadings`	Home-only readings are biased by hero/CTA-heavy markup; the dominant motif on the site as a whole is what `direct` needs

Cross-page aggregation rules:

Read every stardust/current/pages/<slug>.json produced in Phase 2, plus every brand-sources/<host>/pages/<slug>.json when brand sources were captured (§ Origins).
For each numeric/categorical value (border-radius, shadow string, font-size, padding, etc.), build a frequency table weighted by element count, not page count — so a page with 30 cards contributes 30 radii, not 1.
Pick the mode (most frequent) as the primary value; second mode as secondary; pill-radius (≥9999px or ≥50% on a square element) captured separately as pill if present.
For each captured value, record the top 3 source pages in sources so the agent can verify and direct can reason about "where this motif lives."

If a field's mode-on-home and mode-cross-page disagree, prefer the cross-page value and surface the divergence in _provenance.notes (e.g. "home suggested borderRadius=150px (pill, buttons-only); cross-page mode is 3px (cards/inputs/chips, 122 occurrences)"). This divergence note is the single most important hint for direct when deciding whether to keep, soften, or replace the existing motif.

§ Origins

Which origins contributed evidence to this file. Always present: single-origin runs emit a one-entry array. Multi-origin entries appear when extract runs with --brand-source (extract/SKILL.md § Cross-site brand sources).

[
  {
    "origin": "https://example.com",
    "role": "primary",              // "primary" | "brand-source" | "design-source"
    "pagesCaptured": 5,
    "contributedSignals": []        // primary owns everything not attributed elsewhere
  },
  {
    "origin": "https://example.co.uk",
    "role": "brand-source",
    "pagesCaptured": 3,
    "contributedSignals": [
      "motifs.patterns[duotone-photography]",
      "palette[accent-2]",
      "voice.ctaSamples[+2]"
    ]
  }
]

Merge rules:

Conflicts resolve toward the primary. When a brand-source's evidence disagrees with the primary's for the same slot (a palette role, the heading family, the signature radius), the primary origin's value wins; the losing value is noted in _provenance.notes, not silently dropped.
Brand-source evidence widens, never overrides. A motif, photography treatment, or voice pattern the primary site underuses may enter the surface from a sibling — as an additional entry, attributed in contributedSignals[] so downstream direct / uplift can amplify it as captured evidence with an origin citation, not as invention.
Weighting. Brand-source pages join the cross-page frequency tables (§ Aggregation scope), but a value observed only on brand-source pages can never become a token's primary / mode — it lands as secondary, accent-N, or an additional pattern.
Scope exclusions. Brand-source pages are excluded from § System components, § Voice table, and § Cross-promo detection — those describe the primary site's IA. Home-only fields (§ Aggregation scope) always read from the primary origin's home page.
role: "design-source" appears only in stardust/canon-source/_brand-extraction.json (the donor's own capture, per extract/SKILL.md § Cross-site brand sources); donor evidence never aggregates into the primary file.

§ Logo

{
  "source": "inline-svg",          // one of: inline-svg | img | apple-touch-icon | og-image | favicon | synthesized
  "sourceSelector": "header svg",  // CSS selector or URL — null only for synthesized
  "localPath": "stardust/current/assets/logo.svg",
  "format": "svg",                 // svg | png | jpg | ico
  "intrinsicWidth": 180,
  "intrinsicHeight": 32,
  "synthesized": false,
  "synthesizedBasis": null         // e.g. "Brand initials EX, derived from page title"
}

Locator priority chain in playwright-recipe.md § Logo locator chain. First hit wins.

§ Palette

Aggregated computed colors from across all extracted pages (see § Aggregation scope). Frequency-sorted, near-duplicates clustered, role-named.

[
  {
    "role": "background",
    "value": "#ffffff",
    "occurrences": 3421,           // pixels weighted by element area
    "sourceSelectors": ["body", ".container", ".hero"]
  },
  {
    "role": "text-primary",
    "value": "#0f1217",
    "occurrences": 894,
    "sourceSelectors": ["h1", "h2", "p"]
  },
  {
    "role": "primary",
    "value": "#147aff",
    "occurrences": 412,
    "sourceSelectors": [".btn-primary", "a"],
    "sources": ["home", "pricing", "donate"],
    "usedAs": ["background", "border", "fill"]
  },
  {
    "role": "surface",
    "value": "#f7f8fa",
    "occurrences": 280,
    "sourceSelectors": [".card", "section.alt"]
  },
  {
    "role": "border",
    "value": "#e5e7eb",
    "occurrences": 156,
    "sourceSelectors": [".card", "input"]
  }
]

Aggregation rules:

Discard non-colors before aggregation. rgba(*, 0), rgba(0, 0, 0, 0), the literal string transparent, and any CSS color whose alpha channel is 0 are not colors — they are "no color." If a section's background-color resolves to one of these, the section inherits its parent's effective background; the transparent value must not enter the palette aggregation. Picking up rgba(255, 255, 255, 0) as a surface role with high occurrence count (because most elements default to transparent backgrounds in CSS) is a real bug observed on a nonprofit shelter site.
Cluster colors within ΔE < 5 (CIE76 in Lab space) and pick the most frequent member as the cluster representative.
Pure #000 and #fff are kept verbatim — do not silently tint them. Stardust's job here is descriptive, not corrective. The divergence toolkit (in direct) will flag pure black/white when the redesign target rejects them.
Role names are heuristic: most-frequent background-color → background; most-frequent text color → text-primary; most-frequent background-color on [role="button"] and .btn-* → primary; etc. When a role can't be assigned, use accent-N with N ascending.
Cap the palette at 8 entries. If the site uses more, keep the top 8 by occurrences and record the dropped colors in _provenance.notes.
Track usage context per color in usedAs: a deduped list drawn from { "background", "text", "border", "fill", "stroke", "outline" } reflecting which CSS properties this color appeared as across the crawl. A color that appears only as text and never as background/border/fill is the signal the Tensions detector uses for color-imbalance flagging (brand-review-template.md § Detectors).

§ Type

{
  "headingFamily": {
    "name": "Söhne",
    "stack": "\"Söhne\", system-ui, sans-serif",
    "weights": [400, 600, 700],
    "sizes": ["clamp(2rem, 5vw, 3.5rem)", "2.25rem", "1.75rem"],
    "lineHeights": [1.1, 1.15, 1.2],
    "letterSpacing": ["-0.02em", "-0.015em", "-0.01em"],
    "sourceSelectors": ["h1", "h2"]
  },
  "bodyFamily": {
    "name": "Inter",
    "stack": "\"Inter\", system-ui, sans-serif",
    "weights": [400, 500],
    "sizes": ["1rem", "0.9375rem", "0.8125rem"],
    "lineHeights": [1.5, 1.55, 1.4],
    "letterSpacing": ["normal"],
    "sourceSelectors": ["p", "li", "label"]
  },
  "monoFamily": null,
  "scaleRatio": 1.25,              // see § Modular-scale audit below; null if ad-hoc
  "scaleAudit": {
    "kind": "modular",             // "modular" | "ad-hoc"
    "ratios": [1.25, 1.25, 1.20],
    "matchedScale": "major-third"  // null when kind == "ad-hoc"
  },
  "loadStrategy": "swap",           // detected from font-display in @font-face rules
  "files": [                        // captured font files (per playwright-recipe.md § Capture list 16)
    {
      "url": "https://example.com/fonts/sohne-buch.woff2",
      "family": "Söhne",
      "weight": 400,
      "style": "normal",
      "unicodeRange": "U+0000-00FF",
      "localPath": "stardust/current/assets/fonts/sohne-buch.woff2",
      "sourceCssRule": "@font-face { font-family: 'Söhne'; src: url('/fonts/sohne-buch.woff2') format('woff2'); ... }",
      "licensingFlag": "private"     // null | "private" | "open"
    }
  ]
}

files[] is required when at least one captured page references a @font-face rule. Empty array when the site uses only system-stack fallbacks. licensingFlag defaults to null; set to "private" when the family name does not appear in the known-open-licence list (Google Fonts, Adobe Fonts free tier, fontsource.org); "open" when it does. The flag is heuristic — a "private" flag means "verify before redeploying with prototype output", not "this font is non-free."

Identify heading vs body by which family appears in the heading outline (pages/<slug>.json § Headings) most often. If only one family is in use, set both headingFamily and bodyFamily to it with disjoint weights.

Type-size aggregation

Per-level heading sizes are picked by weighted score, not mode-of-frequency:

score = pixelSize × (fontWeight / 400) × √count

For each heading level present in the crawl, group captured headings by (fontSize, fontWeight) tuple, compute the score per group, and emit the top-scoring group's fontSize and fontWeight as that level's representative.

Rationale: WordPress / SEO-plugin patterns ship hidden semantic H1s on every page (e.g. position: absolute; clip: rect(0 0 0 0)). Mode-of-size lets 20 hidden 18px H1s outweigh 2 visible 60px hero H1s by raw count. The pixel-and-weight scoring biases toward the visually-dominant variant — the display H1 is large and heavy and repeated visibly across the site, so it always outranks accessibility scaffolding under this rule.

This rule applies to per-level size selection in type.headingFamily.sizes[]. Mode-of-frequency continues to apply to motif aggregation (border-radius, shadow) per the rules in § Aggregation scope.

Modular-scale audit

After heading sizes are collected, compute the ratio between every consecutive pair (largest → smallest, in px). Compare each ratio to the canonical scale set:

name	ratio
minor-second	1.067
major-second	1.125
minor-third	1.200
major-third	1.250
perfect-fourth	1.333
augmented-fourth	1.414
perfect-fifth	1.500
golden	1.618

If every observed ratio is within ±0.025 of the same canonical ratio, set scaleAudit.kind = "modular", matchedScale = <name>, and scaleRatio = <ratio>.

Otherwise set scaleAudit.kind = "ad-hoc", matchedScale = null, scaleRatio = null, and list the observed ratios verbatim in scaleAudit.ratios. Page-builder sites (Elementor, Webflow, Squarespace) frequently end up here — large uneven jumps like 14 → 18 → 20 → 32 → 45 → 60.

direct reads scaleAudit.kind to decide whether the redesign target should adopt a modular scale by default.

§ Spacing

{
  "baseUnit": 4,                   // 4 | 8 — inferred from mode of paddings/gaps
  "scale": [4, 8, 12, 16, 24, 32, 48, 64, 96],
  "sectionPadding": "96px",
  "containerMaxWidth": "1280px",
  "gridGap": "24px"
}

If the existing site lacks rhythm (paddings are arbitrary, no detectable scale), set scale: [] and baseUnit: null — the divergence toolkit will flag this in direct.

§ Motifs

Signature visual moves the existing site uses repeatedly. borderRadius, shadows, and gradients are aggregated cross-page per § Aggregation scope; patterns is observed where it occurs and labelled with the page list in evidence.

{
  "borderRadius": {
    "primary": "8px",              // mode of non-zero border-radii across all pages, weighted by element count
    "secondary": "16px",
    "pill": "9999px",
    "primarySources": ["home", "about", "stories"],
    "occurrences": { "8px": 122, "16px": 53, "9999px": 38, "2px": 30 }
  },
  "shadows": [
    { "value": "0 1px 2px rgba(0,0,0,0.06)", "uses": "buttons" },
    { "value": "0 4px 16px rgba(0,0,0,0.08)", "uses": "cards" },
    { "value": "0 24px 48px rgba(0,0,0,0.12)", "uses": "modals" }
  ],
  "gradients": [
    { "value": "linear-gradient(135deg, #147aff 0%, #6b21ff 100%)", "uses": "hero-background" }
  ],
  "patterns": [
    { "name": "card-grid",         "evidence": "3 sections use 3-column repeat-cards" },
    { "name": "hero-with-image",   "evidence": "home hero uses split-half image+copy" },
    { "name": "social-proof-strip", "evidence": "logo strip after hero" }
  ]
}

Patterns is the open-ended one. Common values to watch for: card-grid, hero-with-image, hero-with-illustration, feature-3up, feature-list, stat-row, pricing-3up, social-proof-logos, testimonial-carousel, cta-band, footer-mega, nav-mega. Add new pattern names freely; this is descriptive.

§ Component style

The v1 fields, preserved so nothing is lost when DESIGN.json's extensions block carries them forward.

{
  "buttons": {
    "primary": {
      "background": "#147aff",
      "color": "#ffffff",
      "borderRadius": "8px",
      "padding": "12px 24px",
      "fontWeight": 600,
      "shadow": "0 1px 2px rgba(0,0,0,0.06)",
      "hoverDelta": "lighten 6%"
    },
    "secondary": { /* ... */ },
    "ghost": { /* ... */ }
  },
  "dualCTAPattern": "primary-then-secondary-link",  // observed if both appear together; null otherwise
  "cards": { "background": "#f7f8fa", "borderRadius": "12px", "padding": "24px", "shadow": null, "border": "1px solid #e5e7eb" },
  "inputs": { "borderRadius": "8px", "padding": "10px 12px", "border": "1px solid #d1d5db", "focusRing": "0 0 0 3px rgba(20,122,255,0.2)" }
}

§ System components

Cross-page repeated DOM blocks. These are almost always the most load-bearing surfaces of the site (site header, site footer, cross-promo strips, persistent CTAs, breadcrumbs). Missing them at the brand-surface stage means direct cannot decide deliberately whether to keep, move, or kill them — they silently disappear from the redesign target.

[
  {
    "name": "site-header",
    "kind": "header",                // header | footer | cross-promo | nav-secondary | sidebar | cta-band | breadcrumb | background-motif | other
    "occurrences": 12,                // pages where it appears (out of pages crawled)
    "headingSequence": ["About", "Stories", "Donate"],
    "ctaLabels": ["Donate now"],
    "domFingerprintHash": "sha256:...",
    "exampleSlug": "home",            // representative page; the verbatim block lives on this slug
    "exampleSelector": "header > nav.primary",
    "examplePages": ["home", "about", "stories"]
  }
]

Detection algorithm (heading-sequence diff, the cheap version that captures ~80% of the value):

For each extracted page, build a fingerprint from each landmark: (landmark tag, ordered list of immediate-child heading texts + CTA labels).
Group fingerprints across pages. Any fingerprint that appears on ≥ 3 pages (or ≥ 50% of pages crawled, whichever is smaller — so small crawls still surface obvious system blocks) is a system component.
Classify with the kind heuristic:
- landmark header / [role="banner"] → header
- landmark footer / [role="contentinfo"] → footer
- sequence contains 3+ navigation links + 1+ CTA on interior pages but not home → cross-promo
- landmark nav not in header/footer → nav-secondary
- landmark aside / [role="complementary"] → sidebar
- large CTA-only block repeated above the footer → cta-band
- ordered links matching Home > Section > Page → breadcrumb
- else → other
Capture one verbatim example block (HTML serialised from exampleSlug at exampleSelector) so direct and prototype can reason about the actual content, not just the structure.

Cross-page CSS-background reuse

Beyond DOM-fingerprint repeats, also detect shared CSS background images across pages. A background image used on home AND on about-us is almost always a system motif (a recurring banner, a brand-photo treatment, a section ground) — not page-specific content.

Algorithm:

Read every pages/<slug>.json#media.cssBackgrounds[].url across the crawl (after playwright-recipe.md capture step 11).
Group by URL (case-sensitive, full URL match).
For each URL appearing on ≥ 2 pages, emit a system-component entry with kind: "background-motif", occurrences = page count, examplePages = the page slugs.
The exampleBlock for a background-motif entry is the domPath + boundingClientRect + backgroundSize / backgroundPosition of one representative occurrence — enough for direct and prototype to reason about how the brand deploys the image.

This catches the case where a hero photo applied via background-image is shared across multiple pages but invisible to <img>-based capture. Without this aggregation, prototype routinely places brand photos at the wrong section position because it never saw them in the first place.

Detection thresholds

When pages crawled < 3, skip detection and emit systemComponents: [] with a _provenance.notes line ("system component detection requires ≥ 3 pages; crawl too small").

DOM-fingerprint diff (the thorough version) is out of scope for v0.2 — file an issue if the heading-sequence version misses important blocks.

§ Icon font

When the site uses an icon font (per playwright-recipe.md § Capture list 17), capture the family name, the saved file, and the iconClass → codepoint table so downstream prototype and migrate can render the brand's actual icons rather than emoji fallbacks.

{
  "family": "panynj-icons",
  "localPath": "stardust/current/assets/fonts/panynj-icons.woff2",
  "sourceCss": "@font-face { font-family: 'panynj-icons'; src: url('/fonts/panynj-icons.woff2') format('woff2'); ... }",
  "glyphCount": 19,                // distinct icon classes captured (subset of the font's glyph table — only the ones the site actually uses)
  "glyphs": [
    { "class": "icon-search",       "codepoint": "\\E928", "name": "search" },
    { "class": "icon-accessible",   "codepoint": "\\E852", "name": "accessible" },
    { "class": "icon-arrow-right",  "codepoint": "\\E806", "name": "arrow-right" }
  ]
}

Set iconFont: null when no icon font is detected (the common case for SVG-icon and inline-graphic sites). When detected, every glyph the captured pages used appears in glyphs[]; glyphs the font defines but the site doesn't deploy are out of scope.

name is a heuristic guess from the class suffix (icon-search → "search"); leave null when the suffix is opaque (icon-i7, glyph-22, etc.).

§ Voice

Sampled copy from the home page. Used by direct to reason about tone moves and to seed the voice.examples.do/dont arrays in DESIGN.json.

{
  "heroHeadline": "Build, ship, and own your work",
  "heroSubcopy": "...",
  "heroImage": {
    "url": "https://example.com/img/hero-2-engineers-tablet.avif",
    "alt": "Two engineers reviewing a tablet at a standing desk",
    "source": "css-background",       // "img" | "css-background" | "css-pseudo-background" | null
    "domPath": "main > section.hero",
    "localPath": "stardust/current/assets/media/hero-a3f9.avif",
    "rect": { "x": 0, "y": 80, "width": 1440, "height": 720 }
  },
  "heroMedium": {
    "kind": "video",                  // "video" | "canvas" | "lottie" | "animated-svg" | "scroll-motion" | null
    "mechanism": "hls",               // "video-file" | "hls" | "youtube/vimeo-embed" | "webgl" | "lottie" | "css-keyframes" | "scroll-driven"
    "src": "https://customer-xxxx.cloudflarestream.com/<id>/manifest/video.m3u8",
    "domPath": "div.home-video-wrapper > video#background-video",
    "loader": "hls.js@cdn",           // library the source uses to drive it, if any
    "rect": { "x": 0, "y": 0, "width": 1440, "height": 900 }
  },
  "primaryCTALabel": "Start free trial",
  "ctaSamples": ["Start free trial", "Talk to sales", "See pricing", "Read the docs"],
  "navItems": ["Product", "Pricing", "Customers", "Docs", "Sign in"],
  "footerHeadings": ["Product", "Company", "Resources", "Legal"],
  "firstParagraph": "...",
  "tone": {
    "guess": "professional-warm",  // descriptive guess; one of: professional-warm | professional-formal | playful-bright | playful-dry | technical-precise | aspirational | bold-direct | other
    "evidence": "short sentences, second-person address, no jargon"
  }
}

`heroImage` resolution

The hero photo is the home page's most load-bearing visual asset. Without this field elevated, downstream prototype has to re-derive "which captured image is the hero" from a noisy 16-image list every render — and frequently picks the og:image (a curated thumbnail optimised for social cards) instead of the actual visible hero. The 2026-05-04 logistics home shipped the wrong hero on three variants this way before the user noticed.

Resolve from the home page's captured media (both pages/home.json#media.images[] and media.cssBackgrounds[] — the second covers background-image and ::before / ::after pseudo-element heroes per playwright-recipe.md § Capture list 11):

Filter to candidates with rect.top < 800 px (within the first viewport).
Filter to candidates with rect.width × rect.height ≥ 100 000 px² (excludes header logos, tiny chips, decorative columns).
Filter to candidates with rect.aspectRatio in [0.3, 3.0] — excludes thin banners and tall vertical strips.
Pick the candidate with the largest visible area. Tie- breaker: <img> over cssBackgrounds[] over cssBackgrounds[] with pseudo selector (slight preference for explicit DOM imagery, which downstream consumers can manipulate more reliably).

Set source to "img", "css-background", or "css-pseudo-background" based on which capture produced the winner. When no candidate survives the filters, set heroImage: null and surface in _provenance.notes (most non-marketing pages — search, 404, login — legitimately have no hero).

Downstream prototype reads heroImage.url / heroImage.localPath directly when composing the hero slot under Mode A's image-reuse contract (per skills/direct/SKILL.md § Mode A).

`heroMedium` resolution (signature hero medium)

The hero is often not a still image but a moving asset — a background video (file / HLS / Mux / Cloudflare Stream), a canvas/WebGL scene, a Lottie/Rive animation, an animated SVG, or a scroll-driven motion. This is the page's signature and is the most common thing a generic redesign silently flattens to a static hero. Elevate it here so downstream prototype preserves it under intent-dimensions.md § 8b (Signature preservation) instead of re-deriving it from the raw media inventory and missing it — the exact failure that dropped a fintech home's HLS brand-animation on the first migration pass.

Resolve from the home page's captured media (pages/home.json#media): the <video> / <iframe> inventory (playwright-recipe.md § Capture list — every <video>/<iframe> src), <canvas> presence, and any motion library detected among loaded scripts (hls.js, lottie, three.js, gsap, lenis).

Prefer a <video> / <canvas> / Lottie mount whose rect intersects the first viewport (rect.top < 800) and covers a large area (same area/aspect filters as heroImage).
Record kind + mechanism + src (the manifest/file/embed URL — reuse it the way image-fidelity reuses image URLs) + loader (the driving library, so prototype can wire it).
If the hero's signature is motion without a discrete asset (scroll-pinned, parallax, kinetic type), set kind: "scroll-motion" with mechanism naming the technique and src: null.
When the hero has no moving signature, set heroMedium: null (most pages — the still heroImage is then the hero).

A non-null heroMedium is a § 8b trigger: every Mode A variant must reproduce it (graceful fallback + prefers-reduced-motion + scrim), and doing so is exempt from the per-page surprise budget.

The tone.guess is a heuristic — never present it as ground truth in the user report. direct will use it as one of several signals, not as a fact.

§ Voice table

Cross-page aggregations that fuel the brand-review's data tables. The voice block above samples the home page; voiceTable aggregates across all extracted pages so the review can show frequency, not just examples.

{
  "ctaFrequency": [
    { "label": "Donate", "total": 35, "pageCount": 10, "pages": ["home", "about-us", ...] },
    { "label": "Read More", "total": 56, "pageCount": 5, "pages": [...] }
  ],
  "headingFrequency": [
    { "text": "Looking for immediate shelter?", "total": 22, "pageCount": 10, "level": 2 },
    { "text": "Get Involved",                   "total": 11, "pageCount": 9,  "level": 3 }
  ],
  "toneMetrics": {
    "headingsTotal": 312,
    "headingsUppercasePercent": 44,
    "distinctHeadings": 153,
    "distinctCtaLabels": 40
  }
}

Aggregation rules:

ctaFrequency — built from every pages/*.json#ctas[].label. To catch styled <a> links the visual-button heuristic missed (e.g. WordPress sites that style links as buttons via classes), also include link texts that match the closed list below. Match is case-insensitive after trimming whitespace; full-string equality only (no partial / fuzzy matching).
```
donate, donate now, give, give now, support us, contribute,
read more, learn more, more info, see more, view more, view, more,
discover more, explore, overview,
sign up, signup, subscribe, register, join, create account,
contact, contact us, get in touch, talk to us, reach out, say hello,
get help, get involved, find help,
volunteer, share, download, submit,
here, click here, read this, this
```
These mirror and slightly expand the CTA equivalence buckets in brand-review-template.md § CTA equivalence buckets; the buckets drive the T-cta-vocab Tensions detector, while this list drives the voiceTable.ctaFrequency aggregation.

Sort by total descending; on ties, sort by pageCount descending. Cap retained list at 8 entries; the rest are dropped (the brand-review table renders the top 8).

Closed list. The list is exhaustive for v0.2; do not extend at runtime. When a real run surfaces a missing CTA verb worth catching, add it here in spec, not in the renderer. Fuzzy / edit-distance matching is tracked as a v0.3 issue (brand-review-template.md § Open issues for v0.3).
headingFrequency — built from every pages/*.json#headings[].text. Filter to entries with pageCount >= 3. Sort by total, then pageCount. This is the strongest signal of repeated site-wide blocks (e.g. a "Looking for immediate shelter?" cross-promo).
toneMetrics.headingsUppercasePercent — fraction of headings whose text equals its uppercase form. The brand-review uses ≥ 25 as the threshold to apply text-transform: uppercase to H1/H2/H3 in the review chrome.
toneMetrics.distinctHeadings and distinctCtaLabels — set sizes across all pages. High distinct-CTA counts (>20) often signal vocabulary fragmentation; the Tensions detector (brand-review-template.md § Detectors, T-cta-vocab) consumes this.

§ Cross-promo

A cluster of repeated headings + CTAs that appears on most interior pages. Almost always the most load-bearing site-wide repeated block (call-to-action band, footer cross-promo, "looking for X?" panel). Detection runs after headingFrequency is built.

{
  "detected": true,
  "anchorHeading": "Looking for immediate shelter?",
  "cluster": [
    { "text": "Looking for immediate shelter?",  "total": 22, "pageCount": 10, "overlap": 10 },
    { "text": "Experiencing a housing crisis?",  "total": 12, "pageCount": 10, "overlap": 9  },
    { "text": "Please call: 211",                "total": 11, "pageCount": 10, "overlap": 9  }
  ],
  "pages": ["about-us", "donate", "events", ...],
  "pageCount": 10,
  "totalPages": 24
}

Detection algorithm:

Take the most-frequent entry in headingFrequency as the anchor.
Find all pages whose headings include the anchor text — call this the anchor-page set.
For each top-12 entry in headingFrequency, count how many anchor-pages also contain it (overlap). Keep entries where overlap / anchorPages.size >= 0.6.
If the resulting cluster has ≥ 2 entries, emit a cross-promo record.

When detected, the brand-review must render a cross-promo reproduction section (brand-review-template.md § Section ordering → Cross-promo reproduction). Set detected: false and emit null for the rest when no cluster meets the threshold.

§ Embed-dominated pages

After Phase 2, scan each page's embedDominance field (current-state-schema.md § Embed dominance). If any page has dominated: true, add a section to the generated DESIGN.md (and mention in the user report) labeled "Third-party embeds (opaque to extraction)" that lists the affected slugs and embed sources. The brand-surface tokens for those pages are not captured in computed styles — the screenshot is the only artifact direct and prototype can reason from.

Do not attempt recursive extraction of the iframe src URL in v0.2; that's tracked as a separate feature.

§ Register

"register": "brand"

One of brand, product, ambiguous. Heuristic:

brand if landing-page indicators dominate: hero with marketing copy, social proof, pricing, signup CTA above the fold, no authentication required to see the page.
product if tool indicators dominate: data tables, navigation optimized for known features, requires auth, contains widgets like filters and sort.
ambiguous if both appear or neither dominates. Stardust will ask the user in direct rather than guess.

What this file is not

A definition of what the brand should be. That's the target PRODUCT.md and DESIGN.md, written by direct.
A critique. No judgement, no scores. critique belongs to impeccable.
A migration spec. The brand surface describes; it does not prescribe.