zugzbot 1.0.21 → 1.0.23

package/.opencode/oh-my-design/spec/omd-v0.1.md

@@ -1,382 +0,0 @@
-# OmD Spec v0.1
-
-| Field | Value |
-|---|---|
-| Spec version | `0.1.0` |
-| Compatible with | `DESIGN.md` (Google Stitch format, 9 sections) |
-| Adds | Brand Philosophy Layer (6 sections) |
-| Target consumer | Claude Code, Cursor, Gemini CLI, Codex, other AI coding agents |
-| Filename | `DESIGN.md` at project root |
-| Status | Draft — feedback welcome at https://github.com/kwakseongjae/oh-my-design/issues |
-
----
-
-## 0. Why this spec exists
-
-`DESIGN.md` as defined by Google Stitch is a **token document** — color, typography, component, layout rules. That is necessary but not sufficient.
-
-When an AI coding agent generates UI from tokens alone, the output is coherent but **branded like nobody** — it defaults to the statistical center of its training distribution. The Claude Design system prompt names this failure mode explicitly: gradient backgrounds, Inter-on-white, purple on white, unjustified emojis. The fix is not more tokens. The fix is **brand context** — what the brand sounds like, who it's for, what it refuses to do, and how it moves.
-
-OmD v0.1 extends the Google 9-section format with 6 brand-philosophy sections that feed an agent the context it needs to generate on-brand output. The 9 base sections stay untouched. Existing Google-Stitch `DESIGN.md` files remain valid OmD files — the philosophy layer is **additive**.
-
-**OmD is a spec, not a code generator.** The artifact OmD produces and governs is `DESIGN.md` itself — the spec document. OmD intentionally does not emit `theme.css`, `tokens.json`, Tailwind configs, or other stack-specific code artifacts. Agents are expected to read the spec and translate it to whatever stack the project uses. This separation is deliberate: stacks change (shadcn/ui → something else → something after that); brand philosophy doesn't. A spec that outlives any one framework is more useful than one that bakes a framework in.
-
----
-
-## 1. File anatomy
-
-```
-DESIGN.md
-├── Base layer (Google Stitch, required)
-│   ├── 1. Visual Theme & Atmosphere
-│   ├── 2. Color Palette & Roles
-│   ├── 3. Typography Rules
-│   ├── 4. Component Stylings
-│   ├── 5. Layout Principles
-│   ├── 6. Depth & Elevation
-│   ├── 7. Do's and Don'ts
-│   ├── 8. Responsive Behavior
-│   └── 9. Agent Prompt Guide
-└── Brand Philosophy layer (OmD v0.1, recommended)
-    ├── 10. Voice & Tone
-    ├── 11. Brand Narrative
-    ├── 12. Principles
-    ├── 13. Personas
-    ├── 14. States
-    └── 15. Motion & Easing
-```
-
-Required vs optional:
-
-| Section | Status | Rationale |
-|---|---|---|
-| 1–5 | **Required** | Agents cannot generate anything coherent without theme, color, type, component, layout. |
-| 6–9 | **Recommended** | Needed for production-quality output. A file without section 9 is consumable but weaker. |
-| 10–15 | **Recommended** | The OmD layer. A file without these is valid but generic. |
-
-An OmD file MUST begin with a YAML frontmatter block declaring the spec version:
-
-```markdown
----
-omd: 0.1
-brand: <brand name>
----
-
-# <Title>
-
-## 1. Visual Theme & Atmosphere
-...
-```
-
-Files without frontmatter are treated as plain Google-Stitch `DESIGN.md` (valid but not OmD-annotated).
-
----
-
-## 2. Base layer (sections 1–9)
-
-These match the [Google Stitch DESIGN.md format](https://stitch.withgoogle.com/docs/design-md/overview/). This spec does not redefine them; it only enforces two conventions the Stitch docs recommend but don't mandate:
-
-1. **Descriptive names + exact hex values together.** `Toss Blue (#3182f6)`, not `Primary` alone, not `#3182f6` alone. This is non-negotiable — agents use the name to *reason*, and the hex to *render*.
-2. **Explain the "why" at least once per section.** "Shadow-as-border replaces traditional `1px solid` throughout to enable smoother transitions and rounded corners without clipping" is correct. "Uses shadow-bordered cards" is not.
-
-Anti-patterns to avoid (taken directly from Claude Design's anti-slop list):
-
-- `rounded-xl` or other framework-specific shorthand without the underlying pixel value
-- Generic palette names (`primary-500`) without a brand-specific descriptor
-- Font choices of Inter, Roboto, Arial, or system-ui as a "safe default" — these are the default AI-slop fonts. If your brand uses Inter, state that it *deliberately* uses Inter and explain why.
-- Gradient backgrounds, purple-on-white, unjustified emojis, left-accent cards as decorative choices
-
----
-
-## 3. Brand Philosophy layer (sections 10–15)
-
-This is what OmD adds. Each section is a few hundred words max — the goal is not a brand bible, it's a brief that a coding agent can fit in its context window and apply consistently.
-
-### 10. Voice & Tone
-
-**What to write.** How the product speaks, in 3–5 lines of prose. Then a table mapping contexts to tone shifts.
-
-**Why agents need it.** Microcopy is where brands die. A Toss-style fintech that pushes a blunt "Delete" button where "Remove this card" would fit has lost the brand regardless of how correct its color tokens are.
-
-**Format.**
-
-```markdown
-### 10. Voice & Tone
-
-**Voice** — Toss speaks like a friend who happens to be a fiduciary:
-calm, unhurried, zero jargon, no emoji in financial contexts.
-Positive statements, never hedged. "You have ₩1,240,000" — not
-"Your current balance is approximately ₩1,240,000".
-
-| Context | Tone |
-|---|---|
-| CTAs | Imperative, short (송금하기, 확인) |
-| Error messages | Specific, blameless, action-oriented |
-| Success confirmations | Past tense, one sentence, no emoji |
-| Onboarding | Second-person, present tense, one idea per screen |
-| Financial amounts | Bare numerals, currency after, never "approximately" |
-
-**Forbidden phrases.** "Please note that", "Unfortunately", "Oops",
-any sentence starting with "I'm sorry".
-```
-
-### 11. Brand Narrative
-
-**What to write.** The "why" of the brand in under 200 words. Where it came from, what it refuses to be, what category it's trying to move.
-
-**Why agents need it.** Generation quality correlates with contextual depth. An agent that understands Toss is "fintech that rejected institutional-blue formality for an optimistic cerulean" will make better color decisions in ambiguous cases than one that only has hex codes.
-
-**Format.**
-
-```markdown
-### 11. Brand Narrative
-
-Toss started in 2015 as a single-feature money-transfer app in a
-Korean banking market dominated by legacy institutions with
-institutional-blue websites, 12-digit account numbers, and Active-X
-plug-ins. Its founding rejection was of that entire aesthetic vocabulary:
-the bright, optimistic blue (#3182f6) was chosen specifically because
-it was **not** the indigo of KB, Shinhan, or Woori.
-
-Toss is not a neo-bank — it's a super-app. The interface is calm
-because the product is complex: one app holds transfers, investments,
-credit scoring, insurance. The design's job is to flatten that
-complexity into one gesture per screen. "Your money is in good hands,
-and we'll make it easy" is the brand's entire ask.
-
-What Toss refuses: the seriousness of legacy finance, the playfulness
-of consumer apps, the data-viz density of Bloomberg. Toss occupies a
-specific middle.
-```
-
-### 12. Principles
-
-**What to write.** 5–10 first-principles rules, each a single sentence, derivable to concrete UI decisions.
-
-**Why agents need it.** When a novel case appears that tokens don't cover, the agent needs a rule to fall back on. "Breathing room for money — financial amounts get 1.5× the surrounding spacing of normal text" tells the agent what to do for a balance screen that wasn't in the component list.
-
-**Format.**
-
-```markdown
-### 12. Principles
-
-1. **Breathing room for money.** Financial amounts get ≥1.5× the
-   surrounding spacing of normal text. A balance at 30px with 32px
-   margins is correct; the same balance at 16px margins looks cheap.
-2. **Progressive density.** Summary screens are spacious; detail
-   screens are dense. The deeper the user goes, the more information
-   per pixel — they've committed to the context.
-3. **One action per screen.** If a screen has two primary buttons, it
-   is two screens.
-4. **Blue is interaction, not decoration.** Toss Blue (#3182f6) only
-   appears where the user can tap. It never decorates.
-5. **Restraint communicates trust.** Shadows are single-layer, pure
-   black, low opacity. No colored shadows, no multi-layer stacks. In
-   finance, visual noise undermines credibility.
-6. **Korean and Latin are co-equal.** Never assume one is primary.
-   Typography stacks, optical weights, tabular numerals all assume
-   both are rendering simultaneously.
-```
-
-### 13. Personas
-
-**What to write.** 2–4 persona sketches, each 3–5 lines. Who the product is for in *concrete* terms — name, age bracket, context of use, the sentence they'd say about it.
-
-**Why agents need it.** Personas ground generated UI in a real use case. A Toss screen designed "for 28-year-old Seongjae who checks his balance on the subway at 8:47am before a meeting" is different from one designed "for the user".
-
-**Format.**
-
-```markdown
-### 13. Personas
-
-**정민 (Jeongmin), 28, Seoul.** Software engineer. Checks balance
-once a day on the subway. Expects the app to open to the balance
-screen and load in <1s. If she has to tap twice to see her money,
-she's already annoyed.
-
-**이선생님 (Mr. Lee), 54, Busan.** Small business owner. Uses Toss
-because his daughter set it up. Primary use: transferring to
-suppliers. Needs one-tap repeat transfer. Distrusts anything that
-looks like an ad. Would uninstall the app before clicking a banner.
-
-**예은 (Yeeun), 21, Daegu.** University student. Toss is her
-primary banking app — she's never used a legacy bank's interface.
-Expects Toss Blue to be "banking blue". If another financial app
-uses cerulean, she assumes it's imitating Toss.
-```
-
-### 14. States
-
-**What to write.** Empty, loading, error, success, skeleton patterns. One paragraph each or a table mapping state × surface → treatment.
-
-**Why agents need it.** The Claude Design prompt says *"Placeholder > bad attempt"* — agents will invent bad states if you don't supply good ones. This section prevents the generic gray-box skeleton and the sad "No data" illustration.
-
-**Format.**
-
-```markdown
-### 14. States
-
-| State | Treatment |
-|---|---|
-| **Empty** | Single line of body-gray text explaining *why* it's empty, plus one suggested action as a secondary button. Never an illustration. Never "No data". |
-| **Loading** | Skeleton screens matching the final layout's block structure at `#f2f4f6` (grey100). No spinners on primary surfaces. Spinners only inside already-pressed buttons. |
-| **Error (inline)** | `#f04452` (red500) border, error text below in `red500` 13px. One actionable sentence. Never "An error occurred." |
-| **Error (toast)** | `#191f28` background, white text, 3-second auto-dismiss. One sentence. No icons. |
-| **Success** | Brief flash of `#e8f3ff` (blue50) background behind the updated element, 300ms fade to default. No toast for routine actions. |
-| **Skeleton** | Grey100 blocks at exact final dimensions. 1.2s shimmer at `linear-gradient` with 8% white highlight. Never on financial amounts — those show `--` until loaded. |
-```
-
-### 15. Motion & Easing
-
-**What to write.** Duration scale, easing functions by intent (enter / exit / attention / microinteraction), and one or two signature motions.
-
-**Why agents need it.** Motion is the one dimension Google's 9 sections don't cover. Without named durations and easings, agents produce either no motion or Material Design defaults — neither is "on-brand".
-
-**Format.**
-
-```markdown
-### 15. Motion & Easing
-
-**Durations** (named, not raw milliseconds):
-
-| Token | Value | Use |
-|---|---|---|
-| `motion-instant` | 0ms | State changes that should feel immediate (toggle flips) |
-| `motion-fast` | 150ms | Micro-interactions — hover, focus, small reveals |
-| `motion-standard` | 250ms | The default — sheet opens, page transitions, card expands |
-| `motion-slow` | 400ms | Emphasized transitions — success checkmarks, onboarding |
-| `motion-page` | 600ms | Full-screen transitions between top-level sections |
-
-**Easings:**
-
-| Token | Curve | Use |
-|---|---|---|
-| `ease-enter` | `cubic-bezier(0.0, 0.0, 0.2, 1)` | Things appearing — fast start, settle |
-| `ease-exit` | `cubic-bezier(0.4, 0.0, 1, 1)` | Things leaving — slow start, accelerate out |
-| `ease-standard` | `cubic-bezier(0.4, 0.0, 0.2, 1)` | Two-way transitions |
-| `ease-spring` | `cubic-bezier(0.34, 1.56, 0.64, 1)` | Delightful overshoot — confirmations, achievement states |
-
-**Signature motions.**
-
-1. **Money-moves.** When a balance updates, the old number slides
-   up 20px and fades out (`motion-fast / ease-exit`), the new number
-   slides in from below 20px (`motion-standard / ease-enter`). Never
-   cross-fade money.
-2. **Sheet presentation.** Bottom sheets rise from y+40px with
-   `motion-standard / ease-enter` and a synchronized backdrop
-   fade-in. Dismissal uses `motion-fast / ease-exit`.
-3. **Reduce motion.** If `prefers-reduced-motion: reduce`, all
-   `motion-*` tokens collapse to `motion-instant`. No exceptions.
-```
-
----
-
-## 4. Consumption contract — what agents must do
-
-An AI coding agent that loads an OmD-compliant `DESIGN.md` is expected to:
-
-1. **Parse frontmatter.** If `omd: 0.1` is present, load sections 10–15 as authoritative brand context *before* generating any UI.
-2. **Treat section 12 (Principles) as constraints.** A generated component that violates a principle must be regenerated, not shipped.
-3. **Treat section 10 (Voice & Tone) as microcopy source.** Generated button labels, error messages, and empty states MUST match the voice spec. Never invent a tone.
-4. **Treat section 14 (States) as the default.** Do not generate ad-hoc empty/error states. Use the table.
-5. **Apply section 15 (Motion) tokens.** If the framework exposes transition APIs, wire them to the named tokens.
-6. **Fail loud on missing sections.** If a required section (1–5) is missing, the agent should refuse to generate and surface the specific gap to the user.
-7. **Do not hallucinate across sections.** A value not declared in the file is not authorized to be used.
-
-This contract is what makes OmD a *spec* rather than a template. The file is readable by humans and parseable by tools, but its primary consumer is an LLM that needs explicit constraints to produce on-brand output.
-
----
-
-## 5. Compatibility
-
-| Tool | Support | Notes |
-|---|---|---|
-| Google Stitch | ✅ Base 1–9 | OmD files are valid Stitch inputs. Sections 10–15 are ignored as unknown. |
-| Claude Code | ✅ via `.claude/skills/omd/` | See OmD repo for the skill bundle. Once installed, `DESIGN.md` at project root is auto-consumed. |
-| Cursor | ✅ Base 1–9 | Add `DESIGN.md` to `.cursor/rules/` for lazy loading. |
-| Gemini CLI / Antigravity | ✅ Base 1–9 | Reads root `DESIGN.md` natively. |
-| Copilot | ⚠ Partial | Reads root `DESIGN.md` if present; sections 10–15 support depends on extension. |
-| VoltAgent/awesome-design-md | ✅ Base 1–9 | OmD files can be contributed as-is; philosophy layer is additive. |
-
----
-
-## 6. Versioning policy
-
-- **Minor versions (0.1 → 0.2)** may add new recommended sections or new required fields inside existing sections. Files at version N continue to validate at version N+minor.
-- **Major versions (0.x → 1.0, 1.x → 2.0)** may make previously recommended sections required, or reshape the section tree. A major bump will always ship with a migration guide.
-- **Patch versions (0.1.0 → 0.1.1)** are documentation and anti-pattern list updates only. No schema change.
-- Files declare the exact version they target (`omd: 0.1`). An agent MAY parse newer versions with best-effort, but MUST NOT fail a v0.1 file for lacking a v0.2 section.
-
----
-
-## 7. Open questions for v0.2
-
-These are intentionally left out of v0.1 to keep the first version small:
-
-- **Data-viz primitives** — chart palette, axis rules, annotation patterns. Too domain-specific for v0.1; may become a separate companion spec.
-- **i18n / RTL** — locale-specific typography stacks, mirroring rules. v0.2 candidate.
-- **Accessibility contract** — WCAG target declaration, contrast thresholds, focus-ring tokens. Currently lives inside section 4 as ad-hoc notes; v0.2 will hoist it.
-- **Iconography system** — icon family, weight, optical sizing. Partial coverage in section 4; v0.2 may promote.
-
-Feedback on which of these should be required in v0.2 is welcome.
-
----
-
-## 8. JSON Schema (informative)
-
-A non-normative JSON Schema for the file's section skeleton:
-
-```json
-{
-  "$schema": "http://json-schema.org/draft-07/schema#",
-  "title": "OmD DESIGN.md section manifest",
-  "type": "object",
-  "required": ["omd", "brand", "sections"],
-  "properties": {
-    "omd": { "type": "string", "pattern": "^0\\.\\d+(\\.\\d+)?$" },
-    "brand": { "type": "string" },
-    "sections": {
-      "type": "object",
-      "required": [
-        "visual_theme",
-        "color_palette",
-        "typography",
-        "components",
-        "layout"
-      ],
-      "properties": {
-        "visual_theme":        { "type": "string" },
-        "color_palette":       { "type": "string" },
-        "typography":          { "type": "string" },
-        "components":          { "type": "string" },
-        "layout":              { "type": "string" },
-        "depth":               { "type": "string" },
-        "dos_and_donts":       { "type": "string" },
-        "responsive":          { "type": "string" },
-        "agent_prompt_guide":  { "type": "string" },
-        "voice_and_tone":      { "type": "string" },
-        "brand_narrative":     { "type": "string" },
-        "principles":          { "type": "string" },
-        "personas":            { "type": "string" },
-        "states":              { "type": "string" },
-        "motion":              { "type": "string" }
-      }
-    }
-  }
-}
-```
-
-This is informative rather than normative — the file's authoritative form is the markdown document. The schema exists as a target for future tooling (`omd validate`).
-
----
-
-## 9. Credits
-
-OmD v0.1 builds on:
-
-- [Google Stitch DESIGN.md](https://stitch.withgoogle.com/docs/design-md/overview/) for sections 1–9.
-- [Anthropic Claude Design](https://claude.com/claude) for the architectural principle that brand context, not tokens, is what agents most lack — sections 10–15 exist to supply that context.
-- [Anthropic `frontend-design` skill](https://github.com/anthropics/skills) for the anti-slop list informing section 2's conventions.
-- [VoltAgent awesome-design-md](https://github.com/VoltAgent/awesome-design-md) for demonstrating the demand for a community-curated library of DESIGN.md files.
-
-OmD is MIT-licensed and openly contributable. File issues and PRs at <https://github.com/kwakseongjae/oh-my-design>.

package/.opencode/oh-my-design/spec/regional-sources.yaml

@@ -1,77 +0,0 @@
-# Regional source registry — KR / TW reference verification
-#
-# WHY THIS EXISTS
-# The Tier-2 cross-check catalogs we lean on (getdesign.md, styles.refero.design)
-# are Western-biased: they under-cover Korean and Taiwanese brands, and when they
-# DO list one the values are often stale or guessed. So for KR/TW references the
-# pipeline cannot fall back on Tier 2 — Tier 1 (the brand's OWN surface) has to
-# carry the proof. This file is two things:
-#   (a) a DISCOVERY index — where to find regional brand-owned eng/design surfaces
-#   (b) the POLICY the catalog-integrity proof gate enforces for KR/TW brands
-#
-# LIVENESS / HONESTY RULE
-# The URLs below are SEED pointers, not asserted-live facts. During CREATE/UPDATE
-# the agent MUST HEAD-check the surface it actually cites (same rule as Tier-1
-# liveness). A dead seed → find the brand's current official surface; do not cite a
-# URL you did not open this turn (see spec/verification-pipeline.md "Honesty rules").
-
-policy:
-  applies_to: [KR, TW]
-  cutoff: "2026-06-01"           # refs verified on/after this date are gated; older refs exempt
-  regional_min: 2               # KR/TW gated ref must cite >= 2 brand-owned regional sources
-  rule: >
-    A KR or TW reference verified on/after the cutoff must cite >= regional_min
-    brand-owned regional sources in its §4 "Tier 1 sources" footer and back them
-    with a `## Proof` block in .verification.md (>= 5 raw computed-style samples +
-    source URL + inspect date). "Brand-owned" = the brand's own site, official
-    design-system docs, official engineering/design blog, or official GitHub org.
-    Western third-party catalogs (getdesign.md, styles.refero.design) are Tier-2
-    cross-check ONLY and DO NOT count toward regional_min.
-
-# ── Brand-owned engineering / design surfaces (authoritative Tier 1) ──────────
-# These are first-party. Citing the brand's own site PLUS one of these (or its
-# official GitHub org) satisfies regional_min for that brand.
-brand_owned:
-  kr:
-    - { brand: Toss,              url: "https://toss.tech" }
-    - { brand: 우아한형제들/배민,   url: "https://techblog.woowahan.com" }
-    - { brand: Kakao,             url: "https://tech.kakao.com" }
-    - { brand: "Kakao Ent. (FE)", url: "https://fe-developers.kakaoent.com" }
-    - { brand: Naver,             url: "https://d2.naver.com" }
-    - { brand: "Naver (design)",  url: "https://campaign.naver.com/design" }
-    - { brand: 당근 (Daangn),      url: "https://medium.com/daangn" }
-    - { brand: LINE,              url: "https://techblog.lycorp.co.jp/ko" }
-    - { brand: 마켓컬리 (Kurly),    url: "https://helloworld.kurly.com" }
-    - { brand: 뱅크샐러드,          url: "https://blog.banksalad.com" }
-    - { brand: 29CM,              url: "https://medium.com/29cm" }
-    - { brand: 무신사 (Musinsa),    url: "https://medium.com/musinsa-tech" }
-    - { brand: 직방 (Zigbang),      url: "https://medium.com/zigbang" }
-    - { brand: 오늘의집 (Bucketplace), url: "https://www.bucketplace.com/post" }
-    - { brand: Spoqa,             url: "https://spoqa.github.io" }
-    - { brand: KRDS (gov DS),     url: "https://www.krds.go.kr" }
-  tw:
-    - { brand: Dcard,             url: "https://medium.com/dcardlab" }
-    - { brand: Pinkoi,            url: "https://medium.com/pinkoi-engineering" }
-    - { brand: "Gogolook/Whoscall", url: "https://gogolook.com/en/blog" }
-    - { brand: KKBOX,             url: "https://engineering.kkbox.com" }
-    - { brand: 17LIVE,            url: "https://17.live" }
-    - { brand: KKday,             url: "https://www.kkday.com" }
-    - { brand: iCHEF,             url: "https://www.ichefpos.com" }
-    - { brand: "91APP",           url: "https://www.91app.com" }
-    - { brand: Cake,              url: "https://www.cake.me" }
-
-# ── Discovery aggregators (community media) ───────────────────────────────────
-# Use these to FIND a brand's first-party surface. They are NOT Tier 1 and DO NOT
-# count toward regional_min — a brunch/velog post about a brand is not the brand.
-discovery:
-  kr:
-    - "https://yozm.wishket.com"   # 요즘IT
-    - "https://velog.io"
-    - "https://disquiet.io"
-    - "https://brunch.co.kr"
-    - "https://eopla.net"          # 이오플래닛
-  tw:
-    - "https://www.ithome.com.tw"  # iThome
-    - "https://buzzorange.com/techorange"  # TechOrange
-    - "https://www.inside.com.tw"  # INSIDE
-    - "https://vocus.cc"

package/.opencode/oh-my-design/spec/verification-pipeline.md

@@ -1,154 +0,0 @@
-# DESIGN.md Verification Pipeline
-
-> Source-of-truth pipeline for updating `references/<id>/DESIGN.md` §4 Components (and §1–3 tokens when in conflict). Derived from user-stated pattern: 1차 공식 DS 리서치 → 2차 getdesign.md + styles.refero.design 교차검증 → 3차 DESIGN.md 업데이트.
-
----
-
-## Tiers
-
-### Tier 1 — Primary (mandatory)
-The brand's **own** authoritative surface. Pick the highest-priority source available:
-
-1. **Official design system docs** (e.g. `developer.apple.com/design/human-interface-guidelines`, `tossmini-docs.toss.im/tds-mobile`, `airbnb.design`, `spotify.design`, `kakaocorp.com/page/service/service/design`)
-2. **Live brand site inspection** — DOM/CSS values for the actual component as rendered on the brand's flagship surface (e.g. apple.com/shop, airbnb.com search bar, spotify.com hero CTA)
-3. **Public component library / Storybook** if the brand publishes one (Atlassian DS, Polaris, Carbon, etc.)
-
-**Output:** raw observations table per component (variant name, bg, fg, radius, padding, height, font-size, font-weight, border, source URL, screenshot/snippet).
-
-### Tier 2 — Cross-check (mandatory if any Tier 2 source exists)
-Independent third-party catalogs:
-
-1. **getdesign.md** — `https://getdesign.md/<brand>` (YAML token format; reliable for color taxonomy + component spec)
-2. **styles.refero.design** — search the brand by name or use the direct style URL if known. Refero pages are paginated/scroll-loaded; if the brand isn't visible after one full scroll, **search by name in the search box** before declaring not-listed.
-
-**Output:** parallel observations table for each Tier 2 source.
-
-### Tier 3 — Reconciliation + Write
-1. **Build a conflict matrix** per (component, field): Tier 1 value | getdesign value | refero value.
-2. **Resolution rule:**
-   - All three agree → write that value.
-   - Tier 1 ≠ Tier 2 → Tier 1 wins **only if** the Tier 1 evidence is timestamped/inspectable now (live DOM or current docs page). Otherwise prefer the majority of Tier 2 sources.
-   - Two Tier 2 sources disagree and Tier 1 is silent → flag as `(unresolved)` in DESIGN.md and pick the more recent-looking source; document the conflict in the verification footer.
-3. **Update DESIGN.md** §4 with canonical schema (variant heading + bullet `Field: value` lines).
-4. **Append verification footer** to §4 listing all sources consulted with URLs and the date.
-5. **Run tests:** `npm test --silent` — Components extraction tests must pass.
-6. **Visual smoke check:** load `/reference/<id>` in browser; confirm spec cards render with no blank/white-on-white/elongated-circle regressions.
-
----
-
-## Per-component minimum depth (Option Y)
-
-For each ref, attempt to verify **all** that exist in the brand's surface (do not stop at 4):
-
-| Class | Required if present |
-|-------|---------------------|
-| Button | All semantic variants (primary, secondary, tertiary/text, destructive) × all visual styles (fill, outline, ghost, soft) × all sizes — list at least the canonical default + 2 sizes |
-| Input/TextField | Default + focused + error + disabled, with size variants (e.g. small/medium/large or compact/regular) |
-| Card | At least the headline card and one secondary card variant (e.g. media card, product card, list card) |
-| Badge/Tag | All semantic colors (info/success/warning/error/neutral) + size variants |
-| Tab/Segmented | Active + inactive states with explicit bg AND text color split |
-| Toggle/Switch | On/off |
-| Toast/Alert | Each semantic variant |
-| Dialog/Modal | Default if documented |
-| Avatar | Size + shape variants |
-| ListItem | If brand uses distinctive list patterns (Toss list rows, Apple navigation rows) |
-
-**Stop condition:** when Tier 1 surface has no more component classes documented OR when all Tier 1 components have been verified against ≥1 Tier 2 source.
-
----
-
-## Per-ref deliverable
-
-For each brand processed, produce:
-
-1. **Working notes + Proof block** in `web/references/<id>/.verification.md` (**tracked, committed** — this is the audit trail; see Proof Gate below).
-2. **DESIGN.md §4 update** — canonical schema variant blocks.
-3. **Verification footer** at end of §4:
-   ```
-   ---
-   **Verified:** YYYY-MM-DD
-   **Tier 1 sources:** <list of URLs>
-   **Tier 2 sources:** getdesign.md/<id> | styles.refero.design/<style-id>
-   **Conflicts unresolved:** <list, or "none">
-   ```
-4. **Test additions** — extract-components.test.ts assertions for the canonical default variant of each component class.
-
----
-
-## Proof Gate (Tier 1) — enforced by catalog-integrity
-
-A footer that *says* `Verified: <date>` is not proof — earlier batches stamped
-footers without touching §4 (the "footer-only" anti-pattern). The proof gate makes
-the evidence machine-checkable and tracked. **Active for refs `verified >= 2026-06-01`;
-older refs are exempt** (grandfathered — the gate is forward-only).
-
-During Tier 1 live inspect, write a `## Proof` block into the (tracked)
-`web/references/<id>/.verification.md`:
-
-```markdown
-## Proof — Tier 1 live inspect
-
-**Inspected:** 2026-06-01
-**Method:** playwright getComputedStyle
-**Sources:**
-- https://<brand>.com/ — nav, hero CTA, footer
-- https://<brand>.com/<surface> — commerce/checkout (if separate)
-
-### Raw samples
-- `https://<brand>.com/` · hero CTA `.btn-primary` · bg `rgb(116, 92, 237)` = #745CED · radius 8px · padding 12px 24px · font 16px/600
-- `https://<brand>.com/` · nav link · color `rgb(17,17,17)` · 15px/500
-- … (>= 5 lines, each ONE real computed-style observation with a color or px value)
-```
-
-The catalog-integrity test asserts, for every gated ref:
-
-1. `.verification.md` exists and contains a `## Proof` heading.
-2. The proof block holds **>= 5 raw samples** (lines carrying `rgb(` / `#hex` / `px`).
-3. At least one `http(s)` source URL is present.
-4. The §4 footer `**Tier 1 sources:**` lists **>= 1** URL.
-
-This cannot be satisfied by stamping a footer — you have to paste the real DOM
-values you read. A real inspect produces 30–100 samples (see `apple/.verification.md`);
-fewer than 5 means the inspect did not happen.
-
-**Drift re-check (UPDATE mode):** re-inspect the same selectors and diff the new
-computed values against the stored `### Raw samples`. Any field that moved → update
-§4 and note the rollback reason in the conflict matrix. (Runtime step; not CI-gated.)
-
----
-
-## Regional sources — KR / TW (`spec/regional-sources.yaml`)
-
-getdesign.md and refero under-cover Korean and Taiwanese brands, so Tier 2 cannot
-carry a KR/TW ref. For `country: KR` or `country: TW` refs `verified >= 2026-06-01`:
-
-- The §4 `**Tier 1 sources:**` footer must list **>= 2 brand-owned regional sources**
-  (the brand's own site / official DS docs / official eng-design blog / official
-  GitHub org). See the `brand_owned` index in `spec/regional-sources.yaml`.
-- getdesign.md and styles.refero.design **do not count** toward the >= 2 — they are
-  Tier-2 cross-check only.
-- Use the `discovery` aggregators (요즘IT, velog, iThome, INSIDE…) to FIND the
-  brand's first-party surface, but cite the first-party surface, not the aggregator.
-
-catalog-integrity asserts the `>= 2 non-Western Tier-1 source` rule for gated KR/TW
-refs.
-
----
-
-## Honesty rules (self-imposed)
-
-- **Never claim "checked X" without a tool call in the same turn.** If the verification-footer says `getdesign.md/apple` was consulted, there must be a WebFetch in the transcript pointing at exactly that URL.
-- **Never declare "not listed" on refero from a single scroll.** Use the search box or direct style URL.
-- **When Tier 1 and Tier 2 conflict, surface the conflict** in chat to the user before resolving — do not silently pick.
-
----
-
-## Pilot order
-
-1. **Apple** (Tier 2 refero data already captured in this turn — `a4f123f2-cd4b-4d26-998f-a3d3ee158024`)
-2. **Toss** (Tier 1 TDS already partially captured; revisit Tier 2 refero/getdesign)
-3. **Airbnb**
-4. **Spotify**
-5. **Kakao**
-
-After Batch 1 (5/67) passes user review, scale to Batch 2.