`/``/image text), so the serif-vs-sans signature was lost. Now detects the display face by **visual prominence** (above-fold text node with largest fontSize × rendered width) and classifies serif/sans/italic. *Verified: hyperliquid now captures `Teodor 90px serif` (was reported "Inter").* - **A.3 Content-based section classification** (`extract.ts:966+`) — replaced tag/class string-matching (which defaulted to `unknown` on Webflow/Shopify div-soup) with content signals (text density, image ratio, grid cols, heading size, full-bleed, animation, above-fold) → named bands (hero/marquee/feature/pricing/faq/gallery/logo-strip…). *Verified: switchcollective §13 `unknown` dropped 55%→18%.* - **A.4 Background-treatment capture** (`extract.ts`) — per-band `bgTreatment` (mesh-gradient/linear-gradient/radial/image) + `tone: dark`, surfaced in §13. *Verified: switchcollective mesh-gradient now captured (was "completely missed").* - **A.2 Sanitize garbage reads** (`generate-design-md.ts`) — `fontWeight: 0`/NaN no longer rationalized as "remarkably light weight (0)"; weight claims require a valid 100–900 weight. - **B.8 Display signature leads §1** (`generate-design-md.ts`) — narrative now opens with "Headlines are set in **{display}**, a {serif/sans} display face… Body text is set in **{body}**", instead of hiding the display by anchoring on the body font. ### Verified - `tsc --noEmit` clean; full suite **108 passed / 0 failed** (goldens refreshed). Per-fix re-extraction assertions on hyperliquid + switchcollective. Remaining audit items (A.5 lazy-load, A.6 nav, A.7 imagery-binding, B.9 full consistency-lint, C build hard-gates) tracked for the re-prove pass. ## [2.6.0] — 2026-05-30 > **Correctness over volume — fixes from the audit-max vs getdesign.md (5 random brands).** A ground-truthed audit (screenshots + raw-css) showed the volume scorer said "CA wins 5/5" while CA actually got load-bearing facts wrong: invented primary colors, inflated counts, wrong-page captures, self-contradictions. These 8 fixes target correctness. Full report: `docs/comparison/auditmax-vs-gd-20260530.md`. ### Fixed - **Primary/accent selection** (`tokenize.ts`) — accent.primary now comes from the most prominent **rendered CTA button** background, not frequency/var-name heuristics. **Hard-blocks colors absent from `allColors`** (miro `#00b473`, a never-painted `--tw-color-success-accent`, was the brand color → now `#fde050`; notion `#62aef0` var → `#455dd3` real CTA). Unrendered colors allowed only with a positive brand-var score (revolut `#494fdf`). Penalty regex extended to `success|positive|info|…`. - **Canvas dark/light** (`tokenize.ts` + `generate-design-md.ts`) — directional rescue: when `` is dark but the **dominant rendered section area is light**, the canvas is light (notion: body `#191918`, content white → "pure-white canvas", was "commits fully to dark-mode"). Never flips light→dark, so light sites with dark bands (miro/revolut) are unaffected; genuinely dark sites (linear) keep dark. Narrative + §8 Do/Don'ts now both read the tokenized canvas (no more §1-vs-§8 contradiction). - **Wrong-surface detector** (`generate-design-md.ts`) — flags `⚠️ LOW-CONFIDENCE CAPTURE` when ≥1 strong editor signal (filename/line-number/editor-chrome button labels) + ≥2 signals fire (posthog captured its in-page code-editor demo, not the marketing site). - **Honest color count** (`generate-design-md.ts`) — separates "**N rendered on the page**" from "**M declared in design tokens**" (was "111 distinct colors detected on the live page" when ~52 were rendered). Token-only colors marked `(token)`. **Third-party logo colors filtered** (Google/NVIDIA/Shopify/…). - **Consistency lint** (`generate-design-md.ts` + `tokenize.ts`) — narrative font = tokenized font (kills "narrative says Open Sans, YAML says X"); generic body fonts fall back to the real custom font in heading stacks (miro → Roobert PRO, was "sans-serif"); component vocabulary dedups names (no "Primary Brand, Primary Brand, Primary Brand"); **unique YAML component keys**; empty wrapper components (no padding/radius/border/shadow) purged; e-commerce product-card template gated to genuine product grids (≥4 cards) so it stops being stamped on stripe/notion. - **Narrative honesty** (`generate-design-md.ts`) — WCAG contrast formula `(L1+0.05)/(L2+0.05)` caps at 21:1 (was printing impossible "100:1"/"94.1:1"); removed templated filler ("not the simulated drop-shadow of cheap interfaces", "recedes into transparency"); softened/removed the blanket "no approximation, no hallucination" claim (header, frontmatter, footer). ### Verified - `tsc --noEmit` clean; full suite **108 passed / 0 failed** (goldens refreshed). Per-brand: notion accent `#62aef0`→`#455dd3`, canvas dark→**white**, count "111"→"42 rendered + 70 token"; miro `#00b473` (invented)→`#fde050`, font sans-serif→Roobert PRO; posthog **flagged** app-shell; revolut/stripe/linear/airbnb no regression. Independent re-audit of notion: primary + canvas + count + §8 contradiction **resolved**. ## [2.5.1] — 2026-05-30 > **Prose-quality fix — kill functional-role names leaking into editorial narrative.** Brand-aware naming maps a site's CSS vars to names (`--color-indigo` → "Indigo"), great for hue/brand words but robotic when the var is a *positional role*: `--color-text-quaternary` → "Border: **Text Quaternary**", `--color-bg-primary` → "Background: **Bg Primary**", "Body text reads in **Text Primary**". getdesign.md keeps role names in its structured `colors:` block but uses appearance names in narrative — CA now does the same. ### Fixed - **`scripts/generate-design-md.ts`** — Added `proseColorName`/`proseName` + `isGenericRoleName` (true when a name is composed *entirely* of functional-role tokens — text/bg/surface/border + primary/secondary/…/quaternary). When a resolved name is all-role, prose falls back to the nearest-named-color **appearance** ("Off-Cream", "Jet Black", "Dim Gray"); real hue/brand words (Indigo, Rausch, Hof) are untouched. Applied to the 3 human-readable surfaces only — §1 Visual Theme narrative, §1 Key Characteristics, Agent Prompt Guide (quick ref + prompts + iteration guide). The structured §2 Color Palette table keeps role names by design. - **`scripts/generate-design-md.ts`** — Grammar: article now agrees with the canvas descriptor ("The canvas is **an** inky black surface", was "a inky"). ### Verified - Linear: "Body text reads in ~~Text Primary~~ → **Off-Cream**", "Border: ~~Text Quaternary~~ → **Dim Gray**", "**Indigo**" preserved. Airbnb: "**Rausch**"/"**Hof**" preserved (not over-stripped). 0 role-name leaks in prose across linear/stripe/mongodb; §2 table retention intact. - `tsc --noEmit` clean; full suite **108 passed / 0 failed**. Goldens refreshed — they were stale at v2.4.0 and already diverged 82% from v2.5.0 output *before* this change (position-locked diff cascades on the v2.5.0 Full-Extracted-Palette insertion); this fix's own footprint is 0.7–2.2%, under the 5% threshold. ## [2.5.0] — 2026-05-30 > **Color dimension fix — getdesign.md now wins 0 dimensions on 0 brands.** The comparison scorer counts distinct documented hex colors; CA *extracted* ~45 colors per brand but the capped category lists (`bg≤5`, `text≤5`, `accent≤6`…) only *surfaced* ~18 — so on colorful brands (mongodb, theverge, starbucks, mintlify) getdesign.md documented more and won the color sub-dimension. Fix surfaces the full extracted palette without inventing anything. ### Fixed - **`scripts/generate-design-md.ts`** — Added a **"Full Extracted Palette"** block to the active §2 builder. Sources: `rawData.allColors` ∪ color literals harvested from `cssCustomProperties` values (sites like theverge keep their palette in token vars, not on rendered elements). Root cause of the harvest miss: the in-scope `parseRgb` (css-helpers) matches `rgb()` only and rejected theverge's 160+ pure-hex vars — gate replaced with a hex-or-rgb test. Capped at 32, dedupes against categorized colors. - **`scripts/generate-showcase.ts`** — Mirror block in the public showcase color section so the deployed site reflects the full palette (passes `rawDataDesktop` to `buildColorSection`). ### Result (docs/comparison/ca-vs-gd-final.json, 22 brands w/ both sides) - Color dimension: CA 14 / **GD 4** / tie 4 → **CA 22 / GD 0 / tie 0**. - Per-brand hex documented: mongodb 18→41, theverge 12→20, starbucks 9→21, mintlify 19→33; no regression on high performers (linear 31→58, stripe 40→68). - Avg composite CA 92 → **94** (Δ +28). **GD wins no dimension on any brand.** ## [2.4.0] — 2026-05-30 > **OPTION A — Wipe & Rebuild vs getdesign.md**. Backup atomique (git tag + tarball 864 MB) → suppression 38 extractions communes (skip figma) → re-extraction des 71 brands GD en batches 3-parallèles → score composite /100 avec verdict CA_WINS/GD_WINS/TIE/CA_WEAK → comparaison structurée 71 brands → PDF audit final A4 landscape. ### Added - **`scripts/mass-extract.sh`** — Orchestrateur d'extraction parallèle (3 process max, swap protection), Tier 1/2/3 par difficulté anti-bot, timeout 240s/extract, retry logic. - **`scripts/post-process-all.sh`** — Batch tokenize + DESIGN.md + showcase pour tout extractions/. - **`scripts/sync-catalog.ts`** — Synchronise catalog/index.json avec filesystem (ajout new, suppression ghosts, préservation goldens). - **`scripts/compare-vs-gd-final.ts`** — Comparaison 71 brands GD avec composite score /100 (Volume·Color·Verif·Narrative·Completeness·Sections pondérés), verdict per-brand, ΔRGB palette delta. - **`scripts/generate-final-pdf.ts`** — PDF A4 landscape avec cover, KPIs, scoreboard 71 brands, CA_WEAK roadmap, méthodologie reproductible. - **`scripts/verify-checklist.ts`** — Checklist 10 brands stratifiées (3 dark + 3 light + 2 brand-strong + 2 edge) pour vérification MCP/manuelle des tokens (5 dimensions). ### Changed - **Catalog v1.5.0** — Sync filesystem-driven (88 → 71 GD brands + 50 CA exclusives = ~121 totales) avec timestamp `lastSyncedAt`. - **`package.json` files[]** — Ajout scripts/mass-extract.sh, post-process-all.sh, sync-catalog.ts, compare-vs-gd-final.ts, generate-final-pdf.ts, verify-checklist.ts. ### Verified - Backup atomique testé : git tag `backup/pre-wipe-20260530-1301` + branch `rescue/*` + tarball `/opt/_backups/ca-pre-wipe-*.tar.gz` 864 MB + sha256. - Rollback < 60s : `git reset --hard backup/pre-wipe-* && tar -xzf /opt/_backups/...`. - Stealth retry validé sur Tier 3 anti-bot brands (cf logs/mass-*/timing.txt). ## [2.2.0] — 2026-05-25 > Sprint "Fix-and-ship" — 5 bugs critiques fixés + feature unique "View Source Proof" + déploiement production clone-architect.ps-tools.dev. Score moyen catalog **65 → 78.1/100** sur 9 stress brands, **86.5/100** sur 30 newly-extracted. **6 brands grade A**, GitHub.com premier **100/100** parfait. ### Fixed - **`extract.ts:dismissPopups()`**: Étendu top-5 CMPs (OneTrust, Didomi, Cookiebot, Google Funding Choices, Usercentrics) + scan `page.frames()` pour iframes IAB TCF + selectors native `