zugzbot 1.0.21 → 1.0.22

package/.opencode/oh-my-design/data/architecture-proposals/2026-05-13-thin-install-fresh-fetch.md

@@ -1,189 +0,0 @@
-# Thin install + fresh fetch (WebFetch + MCP)
-
-**Status**: proposed — not implemented
-**Author context**: 2026-05-13 conversation, after KR-10 batch
-**Motivation**: npm publish on every brand addition is high-friction and produces stale installs
-
----
-
-## Problem statement
-
-Current `oh-my-design-cli` package bundles 88 brand DESIGN.md files (~12MB) and ships them via npm.
-
-**Pain points observed**:
-1. Every batch (10 brands) requires `npm publish` to reach end users
-2. Users installed at count N keep seeing N brands forever, even after we add more
-3. Skill descriptions historically baked the count (`"67개 ..."`) — required re-edit per batch, easy to forget (we forgot in the 78 release; observed again at 88)
-4. Bundle grows linearly; cold install time scales
-5. Single brand typo fix = minor version bump + publish
-
-**Decided 2026-05-13**: skill descriptions are now **count-agnostic** ("실제 기업 레퍼런스" without a number). Won't regress this kind of bug again. But the fundamental delivery model still needs reform.
-
----
-
-## Proposed architecture
-
-Hybrid: **thin npm baseline + fresh fetch at runtime**.
-
-```
-npm package (thin, ~500KB)
-├─ skills/                       (routing logic, count-agnostic)
-├─ data/reference-fingerprints.json   (metadata only — id, category, primary, tone keywords)
-├─ data/snapshot-date.txt        ("baseline 2026-05-13")
-├─ hooks/, agents/, bin/         (unchanged)
-└─ (no DESIGN.md bundled — fetched at runtime)
-
-Runtime (skill execution in user's agent)
-1. Match brand via local fingerprints.json (offline-capable)
-2. Fetch full DESIGN.md from:
-   - Primary:  WebFetch  https://oh-my-design.kr/api/references/<id>/design-md
-   - MCP path: omd_get_designmd(id)        (if MCP server connected)
-   - Fallback: bundled (if shipped — optional Phase 3 decision)
-3. Apply, write to project DESIGN.md
-```
-
-**Outcome**:
-- New brand batch → `git push` web + git commit fingerprints; **no npm publish**
-- Users see new brands on next skill call (no reinstall)
-- npm publish frequency drops to "skill behavior changes" (<1/month)
-- Marketing copy (READMEs) still updates count per batch — that's intentional
-
----
-
-## Implementation phases
-
-### Phase 1 — Backend HTTP endpoint (~1h)
-
-`web/src/app/api/references/[id]/design-md/route.ts`:
-```ts
-import { readFileSync } from 'fs';
-import { join } from 'path';
-import { NextRequest, NextResponse } from 'next/server';
-
-export async function GET(_req: NextRequest, { params }: { params: { id: string } }) {
-  try {
-    const md = readFileSync(join(process.cwd(), 'references', params.id, 'DESIGN.md'), 'utf-8');
-    return new NextResponse(md, {
-      headers: {
-        'Content-Type': 'text/markdown; charset=utf-8',
-        'Cache-Control': 'public, max-age=300, s-maxage=3600',
-        'Access-Control-Allow-Origin': '*',
-        'ETag': `"${md.length}-${md.slice(-32)}"`,
-      },
-    });
-  } catch {
-    return NextResponse.json({ error: 'not found' }, { status: 404 });
-  }
-}
-```
-
-Smoke test: `curl https://oh-my-design.kr/api/references/kakaopay/design-md | head -20`
-
-### Phase 2 — MCP server (~3h, parallel with Phase 1)
-
-New package `oh-my-design-mcp` (separate npm package).
-
-Tools exposed:
-| Tool | Args | Returns |
-|---|---|---|
-| `omd_search_references` | `{ query, top_k? }` | top-N brand fingerprints (id, score, why) |
-| `omd_get_designmd` | `{ id }` | full DESIGN.md markdown |
-| `omd_list_categories` | — | category → ids map |
-| `omd_get_snapshot_date` | — | "2026-05-13" (for staleness check) |
-
-Resources exposed (MCP resources, browsable):
-- `omd://references` — list all
-- `omd://references/<id>` — single DESIGN.md
-
-Server reads from same HTTP endpoint internally (single source of truth) OR direct disk if self-hosted.
-
-Distribution: `npx -y oh-my-design-mcp` in MCP config:
-```json
-{
-  "mcpServers": {
-    "oh-my-design": { "command": "npx", "args": ["-y", "oh-my-design-mcp"] }
-  }
-}
-```
-
-### Phase 3 — Skill updates (~1h)
-
-`skills/omd-init/SKILL.md`, `skills/omd-harness/SKILL.md`:
-
-```diff
-- Phase 4: fetch reference DESIGN.md from .claude/data/references/<id>/DESIGN.md
-+ Phase 4: fetch reference DESIGN.md (priority order):
-+   a) If MCP `oh-my-design` connected: call omd_get_designmd(id)
-+   b) Else WebFetch https://oh-my-design.kr/api/references/<id>/design-md
-+   c) Else (offline): bundled .claude/data/references/<id>/DESIGN.md if present
-+   d) Else: report "id not in catalog, network unavailable, no bundle"
-```
-
-Skill must also surface the snapshot date to the user when fetching offline:
-> "오프라인 모드 — 2026-05-13 기준 카탈로그 사용 중. 최신 brand 보려면 인터넷 연결 후 재시도."
-
-### Phase 4 — Bundle diet (~30m, optional)
-
-`package.json`:
-```diff
-"files": [
-  "dist",
-  "skills/*",
-  "agents",
-  "data",
-- "web/references/*/DESIGN.md",
-+ "data/snapshot-date.txt",
-  ".claude/hooks/*.cjs",
-  ".claude/settings.json"
-]
-```
-
-`scripts/postinstall.cjs`: stop copying DESIGN.md files into `.claude/data/references/`. Skills will fetch.
-
-**Trade-off**: completely offline-broken if API down. Mitigation: optional flag `--bundle-references` for users who want offline guarantee.
-
-### Phase 5 — `omd:batch-launch` SKILL update (~15m)
-
-Phase 3 SYNC:
-- Remove `npm publish` from suggested follow-ups
-- Add note: "After git push, brand reaches users automatically via API"
-- Marketing surfaces (README count, llms.txt count) still need batch sync since they're static
-
----
-
-## Migration plan
-
-1. **Phase 1 first** (backend endpoint) — non-breaking, just adds capability
-2. **Phase 2 in parallel** (MCP) — non-breaking, separate package
-3. **Phase 3** (skill update) — backward-compat: tries MCP → WebFetch → bundle, so old installs still work
-4. **Phase 4 last** (bundle diet) — only after Phase 3 has shipped for 1-2 weeks and we confirm fetch reliability
-5. **Phase 5** (omd:batch-launch update) — final cleanup after migration is stable
-
----
-
-## Open questions
-
-- **Authentication**: keep API public read? (Yes — DESIGN.md files are intended for public consumption, that's the product.)
-- **Rate limit**: needed? (Probably basic per-IP at edge — Vercel/Cloudflare. Not critical for v1.)
-- **Versioning of DESIGN.md content**: should the API return a `?version=YYYY-MM-DD` parameter so installs can pin? (Defer — KISS for v1.)
-- **Caching**: should the skill cache fetched DESIGN.md locally per session? (Yes — agent runtime should cache the parsed result for the session.)
-- **What about agent that has no WebFetch tool?** (Edge case. Codex CLI has it. Cursor has it via tool config. If a user's agent really has neither WebFetch nor MCP, they install the `--bundle-references` flag at install time.)
-
----
-
-## When to do this
-
-**Trigger**: next time the user mentions "10개 더" or any brand addition. Pause Phase 2 build, ship the architecture first, then resume.
-
-**Or**: after 2 more batches (100, 110 brands), npm package bundle pain becomes hard to ignore — natural breakpoint.
-
-**Not now**: this conversation closed with count-agnostic skill descriptions as the immediate fix. Architecture is tracked but not implemented yet.
-
----
-
-## Reading list / inspirations
-
-- Linear's hosted catalog approach (linear.app/docs API)
-- Tailwind v4 CDN-first delivery model
-- shadcn/ui's "components are not a package, they're snippets you copy" — opposite direction but instructive
-- MCP spec: `https://modelcontextprotocol.io/`

package/.opencode/oh-my-design/data/issues/2026-05-13-multi-surface-schema-rfc.md

@@ -1,67 +0,0 @@
-# RFC: Multi-surface brand spec (deferred)
-
-**Status**: deferred / not blocking
-**Filed**: 2026-05-13
-**Trigger**: Banksalad mismatch — canonical DESIGN.md captured "advisor app" idiom, but `banksalad.com` is a "marketing landing" idiom. Generated landing.html was schema-faithful but didn't resemble the actual brand homepage.
-**Decision**: do NOT do this schema change now. Solve at the skill level first (see `data/issues/2026-05-13-multi-surface-schema-rfc.md` companion fix in commit history). Revisit only if skill-level fix proves insufficient after ~10 more brand additions or repeated user complaints about surface mismatch.
-
----
-
-## Problem
-
-One brand can have visually distinct surfaces (marketing landing, in-app, docs, onboarding). The current `DESIGN.md`-per-brand schema implicitly picks ONE surface as canonical. Generators following the spec faithfully will mismatch when user requests a different surface.
-
-## Why we're deferring the schema change
-
-1. **Both prior-art systems we surveyed keep schema single**:
-   - Google Stitch treats variants/modes/devices as runtime workflows (sidebar items in their docs), not separate spec documents.
-   - getdesign.md allows N entries per brand with catalog-level tags, but actual catalog data shows 1 entry per brand in practice (e.g. Apple has 1 entry).
-   - W3C Design Tokens Community Group spec deliberately does NOT scope tokens by surface — leaves it to vendor `$extensions`.
-2. **Authoring cost scales poorly**: 88 brands × N surfaces is real maintenance overhead for marginal cases. Only ~10-15% of brands appear to have visually distinct landing/app idioms.
-3. **The Banksalad case is a generator priority bug, not a schema gap**. We already had `tokens.json#live_overrides` and screenshots — the generator just didn't weight them.
-4. **Backward compatibility**: 88 published references would all need migration. High blast radius for unclear payoff.
-
-## What might warrant revisiting
-
-- ≥3 distinct user reports of "this brand's generated output doesn't look like the real brand"
-- A new batch where ≥30% of brands exhibit landing-vs-app idiom split
-- AI coding agents adopting a multi-surface convention (W3C tokens spec, Stitch, etc. adding `surface` as a first-class field)
-
-## Sketch (if/when we do this)
-
-Probable shape based on the survey:
-
-- `web/references/<id>/DESIGN.md` stays — essence-leaning (voice, principles, motion philosophy, raw palette)
-- Optional `surfaces/<surface>.md` overlay files — delta only, frontmatter `surface: marketing|product|docs|onboarding`, `extends: ../DESIGN.md`
-- Generator reads essence + (one) overlay → composes
-- Closed enum chosen for AI determinism (Spectrum/Carbon precedent — productive vs expressive moment)
-- Only ~10-15% of brands need overlays — most stay single-doc
-
-## What we're doing instead (now)
-
-Skill-level fix shipping in v1.4.0 or similar:
-
-1. `omd:reference-capture` writes `structure.json` recording observable composition cues (hero type, CTA shape, nav structure, ornament style) — facts, not copyrighted content.
-2. `omd:apply` / `omd:harness` instructed to Read hero screenshots as images for visual grounding before composing.
-3. Token-resolution priority order documented: `tokens.json#live_overrides` > canonical DESIGN.md > `delta_set`. Brand essence (voice, principles, motion philosophy) stays canonical-only.
-4. `omd:harness` Step 4 master prompt includes a `surface_signal` field derived from task keywords (`랜딩`→marketing, `대시보드`→product, etc.). Master uses it to weight reference-capture artifacts vs canonical spec.
-
-The skill-level fix achieves the same Banksalad-landing outcome without any schema or migration work.
-
-## References (survey notes)
-
-- Stitch docs at `stitch.withgoogle.com/docs/design-md/overview` — sidebar nav includes `Device Types`, `Design Modes`, `Generate design variations` as separate workflows
-- getdesign.md catalog (`getdesign.md/apple`) — N-entries-per-brand with category tags at catalog level, but actual data has single entries
-- W3C Design Tokens spec `tr.designtokens.org` — `$extensions` escape hatch, no surface primitive
-- Material 3 (`m3.material.io`) — three-tier `ref/sys/comp` is vertical, not surface-horizontal
-- Carbon (`carbondesignsystem.com`) — theme axis (`white/g10/g90/g100`) is the closest to surface, but it's a single coarse dimension
-- Adobe Spectrum (`spectrum.adobe.com`) — `productive/expressive` moment is a 2-bucket idiom split; precedent for closed enum
-- shadcn/ui (`ui.shadcn.com`) — explicitly refuses the surface problem by shipping unstyled primitives
-
----
-
-## Linked actions
-
-- [ ] Implement skill-level fix (4 items above) — target v1.4.0
-- [ ] After 3 more batches, re-evaluate whether skill fix is sufficient
-- [ ] If not, draft formal `omd: 0.2` schema following the surfaces/<name>.md sketch

package/.opencode/oh-my-design/data/opt-out-corpus.json

@@ -1,141 +0,0 @@
-{
-  "version": 1,
-  "description": "Korean + English opt-out / fast-exit / frustration signals. Used by signal-classifier to detect when the user wants the harness to stop probing.",
-  "categories": {
-    "opt_out_full": {
-      "description": "User wants out completely — go to FAST_EXIT.",
-      "ko": [
-        "그만",
-        "stop",
-        "그만해",
-        "그만 좀",
-        "아 됐어",
-        "됐어",
-        "그만 됐어",
-        "이제 그만",
-        "충분",
-        "충분해",
-        "이만하면 됐어",
-        "끝내자"
-      ],
-      "en": [
-        "stop",
-        "stop it",
-        "enough",
-        "i'm done",
-        "no more",
-        "we're done",
-        "that's enough",
-        "leave it"
-      ]
-    },
-    "opt_out_skip": {
-      "description": "Skip current question — defaults OK.",
-      "ko": [
-        "skip",
-        "그냥",
-        "그냥 가",
-        "알아서",
-        "알아서 해줘",
-        "넘어가",
-        "넘기자",
-        "패스",
-        "건너뛰",
-        "다음",
-        "다음으로",
-        "나중에",
-        "지금 말고"
-      ],
-      "en": [
-        "skip",
-        "skip it",
-        "just go",
-        "you decide",
-        "your call",
-        "next",
-        "later",
-        "not now"
-      ]
-    },
-    "frustration": {
-      "description": "User is frustrated with current proposal — ALL options off.",
-      "ko": [
-        "다 별로",
-        "다 이상",
-        "이건 아닌데",
-        "별로야",
-        "별로",
-        "이상해",
-        "이상한데",
-        "마음에 안 들어",
-        "마음에 안 들",
-        "다시",
-        "다시 해",
-        "다 다시"
-      ],
-      "en": [
-        "all bad",
-        "none of these",
-        "neither",
-        "not what i want",
-        "i don't like",
-        "redo",
-        "start over",
-        "try again"
-      ]
-    },
-    "approval": {
-      "description": "Affirmative — proceed.",
-      "ko": [
-        "go",
-        "ok",
-        "좋아",
-        "좋네",
-        "그래",
-        "응",
-        "ㅇㅇ",
-        "그렇게",
-        "이대로",
-        "맞아",
-        "맞아요",
-        "괜찮",
-        "할게"
-      ],
-      "en": [
-        "go",
-        "ok",
-        "okay",
-        "yes",
-        "sure",
-        "good",
-        "looks good",
-        "lgtm",
-        "let's go",
-        "proceed",
-        "ship it"
-      ]
-    },
-    "deep_engagement": {
-      "description": "User wants more depth/iteration — opposite of skip.",
-      "ko": [
-        "더 깊게",
-        "한번 더",
-        "더 보여줘",
-        "다른 옵션",
-        "더 다듬",
-        "디테일",
-        "조금 더",
-        "이 부분"
-      ],
-      "en": [
-        "more options",
-        "show more",
-        "go deeper",
-        "more detail",
-        "what about",
-        "let's refine"
-      ]
-    }
-  }
-}
-

package/.opencode/oh-my-design/data/reference-audits/2026-05-13-kr10.md

@@ -1,132 +0,0 @@
-# Batch Audit — 2026-05-13 / KR-10
-
-**Theme**: Korean IT companies with publicly observable design system signal
-**Brands added**: 10 (78 → 88)
-**Skill**: `omd:batch-launch` (Phase 2 build pipeline = `omd:add-reference` CREATE)
-**Verification date**: 2026-05-13
-
----
-
-## How the 10 were researched (pipeline applied to each)
-
-Every brand went through the standard `omd:add-reference` CREATE 3-tier pipeline. Each `omd-batch-launch` subagent was constrained to:
-
-1. **Tier 1 — Live inspect (required)**: playwright MCP `getComputedStyle` on hero CTA / nav / footer / input / cards across **≥1 surface** of the brand's production site. Where possible, also `curl` of the production CSS bundle (more authoritative than runtime sampling because it exposes the entire token system).
-2. **Tier 2 — Cross-check (both attempted)**: `getdesign.md/<id>` + `styles.refero.design/?q=<brand>` (also `?q=<한글명>`). Either source success = OK; both empty = explicitly logged in §4 footer.
-3. **Tier 3 — Reconcile + write**: 9 product sections (Stripe schema) + §10-15 philosophy (Toss §10-15 shape) + `_research.md` source map + `_promo.json` highlight + verified footer.
-
-Every brand also has:
-- `web/references/<id>/DESIGN.md` (the spec, frontmatter `omd: 0.1`)
-- `web/references/<id>/_research.md` (Tier-by-Tier source list + dates + confidence per section)
-- `web/references/<id>/_promo.json` (id, Korean name, logo URL, single "most striking" highlight for promo video)
-- §4 footer with `**Verified:** 2026-05-13` + Tier 1 sources + Tier 2 status + Conflicts unresolved
-
----
-
-## Systemic finding
-
-**Tier 2 directories (`getdesign.md`, `styles.refero.design`) have weak Korean coverage**. All 10 brands returned **empty** on both 3rd-party indexes. This is not a brand-quality issue — it's a directory-bias issue (both are English-tooling-oriented). Documented per-brand in §4 footers.
-
-**But brand-published DS / design narrative resources DO exist for most of the 10** — they were just not in those 3rd-party indexes. 7/10 have public DS-related surfaces that were used in Tier 1 reconcile:
-
-| brand | brand-published DS surface | type | live (2026-05-13) |
-|---|---|---|---|
-| socar | `design.socar.kr` | docs hub | ✅ 200 |
-| wanted | `montage.wanted.co.kr` + `wanted.co.kr/brandcenter` | full DS | ✅ 200 |
-| remember | `dramancompany.github.io/remember-ui` | Storybook | ✅ 200 |
-| banksalad | `github.com/banksalad` (org + styleguide repos) | source | ✅ 200 |
-| zigzag | `brunch.co.kr/@zigzag/73` + `devblog.kakaostyle.com` | brand articles (ZDS) | ✅ 200 |
-| gangnamunni | `blog.gangnamunni.com/post/welchis/` | DS architecture post | ✅ 200 |
-| kakaopay | `story.kakaopay.com` | brand / design narrative | ✅ 200 |
-| 29cm | — | — | ❌ no public DS |
-| ably | — | — | ❌ no public DS |
-| zigbang | — | — | ❌ no public DS |
-
-→ For future Korean batches, weight Tier 1 (live inspect + **brand-owned tech blogs / brunch / medium / GitHub orgs**) more heavily. Tier 2 (3rd-party directories) is documentation discovery, not validation.
-
-→ The 7 brand-published DS URLs are also registered in `web/src/lib/design-systems.ts` so they surface as DS buttons in Browse modal and Builder Step 3 preview header. The 3 brands without a public DS correctly return null and the DS button stays hidden.
-
----
-
-## Per-brand audit
-
-Confidence rubric:
-- **High** — ≥2 Tier 1 surfaces OR production CSS bundle captured; no unresolved hex; ≥3 supporting brand-owned sources (blog/medium/brunch); §10-15 sourced
-- **Medium** — 1 Tier 1 surface; some inferred values; brand-owned philosophy sources present
-- **Low** — Tier 1 partially blocked; key tokens marked `(unverified live)` or `(illustrative)`; needs UPDATE pass
-
-| # | id | confidence | Tier 1 depth | Tier 2 | Known gaps | Follow-up |
-|---|----|------------|--------------|--------|------------|-----------|
-| 1 | socar | **Medium** | 1 surface (home full); design.socar.kr partial (browser session contended) | empty (getdesign no record · refero no isolated set) | SOCAR Blue **exact hex unpublished** — brand-name-only reference. Brand fonts Sandoll Gothic Neo2 / Avenir (print) vs live Pretendard (web) — documented as intentional print-vs-web split. | UPDATE pass on `design.socar.kr` for canonical SOCAR Blue hex + Space Frame system tokens |
-| 2 | gangnamunni | **High** | 3 (home live + production CSS bundle `d246c5b2edcb00b6.css` with full Cell token map + blog.gangnamunni.com Cell+Welchis architecture post) | empty | none material | — |
-| 3 | kakaopay | **Medium** | 1 surface (home full live inspect); /payment/online second surface blocked by browser tab contention | empty (both verified empty) | none unresolved. KPDS = narrative-only on `story.kakaopay.com`, no public docs site. | UPDATE pass on /payment/online or /transfer for second surface verification |
-| 4 | zigzag | **High** | 4 (home live + production CSS bundle 184KB with full ZDS canonical tokens via `--zds-color-*` vars + devblog.kakaostyle.com ZDS rearchitecture + brunch.co.kr/@zigzag/73 brand-owned) | empty | Two-pink architecture (`#fa6ee3` brand-anchor vs `#f55dd6` interactive primary) — documented as intentional role split | — |
-| 5 | 29cm | **Medium** | 1 surface (home via isolated Playwright `browser_run_code_unsafe` workaround for session contention) | empty | PDP / checkout flows not inspected this pass | UPDATE pass on PDP + checkout (editorial product story modules) |
-| 6 | ably | **Low** | 1 navigation only — live `evaluate` calls blocked by Playwright session contention; §4 token hexes flagged `(unverified live)` in footer | empty / inconclusive (refero JS-rendered, capture blocked) | §4 token values need live re-capture. Brand disambiguation note added (Ably Corp KR commerce ≠ ably.com realtime API). | **Priority**: full Phase U2 live inspect re-run on a clean Playwright session |
-| 7 | banksalad | **High** | 5 (home live + 865KB production CSS bundle with `border-radius:2px` ×81 confirming system default + BPL tech blog + 2× GitHub orgs) | empty (both verified empty) | One nav exception documented (sign-in pill at 16px radius vs system 2px default) — already noted as exception, not conflict | — |
-| 8 | zigbang | **Medium** | 2 (home full + /home/apt/map partial — map-overlay listing card tokens inferred due to second-pass contention) | empty | (a) **brand-orange exact hex unpublished** — `~#FF6600` flagged illustrative in §2 + footer; (b) map-overlay listing-card / price-chip tokens inferred | UPDATE pass on /home/apt/map for canonical orange hex + listing-card token verification |
-| 9 | wanted | **High** | 8 sources (live DOM on `wdlist/518` 22 samples + brandcenter + Montage docs + wanted-sans GitHub + wanteddev org + wantedlab.com + WDS creation blog) | empty (refero blocked by session contention, getdesign verified empty) | none unresolved. `#14191E` (marketing) vs `#171719` (product) is documented two-surface split, not a conflict. | (optional) refero re-attempt on clean session |
-| 10 | remember | **High** | 5 (career page primary live inspect + `dramancompany.github.io/remember-ui/` Storybook root + corp x2 + redirect) | empty (3rd-party empty + GitHub UI repo 404 — source private/deleted, Storybook deploy still live) | (a) filter-chip **selected/active** state not observed (only default captured); (b) error/warning semantic tokens inferred from KR fintech convention | UPDATE pass capturing filter chip selected/active + error/warning states |
-
----
-
-## Confidence distribution
-
-- **High** (5): gangnamunni, banksalad, wanted, zigzag, remember
-- **Medium** (3): socar, 29cm, kakaopay
-- **Low** (1): ably
-- *(zigbang was kept at Medium even though 2 surfaces inspected, because of 2 unresolved illustrative tokens — orange hex + map overlay)*
-
-**Aggregate**: 5/10 are publishable as-is. 3/10 are publishable with marked uncertainty. **2/10 (ably + zigbang) warrant a follow-up UPDATE pass before being cited in external blog posts or comparisons.**
-
----
-
-## Known shared limitation — Playwright session contention
-
-The Phase 2 build ran 10 subagents in parallel, all sharing the host Playwright MCP browser. Several agents reported that `browser_evaluate` calls landed on *other agents' tabs* (kakaopay/zigzag/banksalad/29cm appear in multiple reports as "the tab that hijacked my session"). Mitigations applied per agent:
-
-- Fall back to direct `curl` of the production CSS bundle (banksalad, gangnamunni, zigzag) — actually higher signal than runtime sampling for these
-- Use `browser_run_code_unsafe` in isolated newPage context (29cm)
-- Document and explicitly flag affected surfaces in `_research.md` (ably, kakaopay, zigbang)
-
-**Process fix for next batch**: serialize Playwright sessions per brand (run 1-2 in parallel, not 10), OR isolate to `--user-data-dir` per agent if available. Cost: longer wall-clock. Benefit: no contention, clean §4 token values for every brand.
-
----
-
-## Promo highlight selection (auditable)
-
-Each brand's `_promo.json` records a single "most striking" element — chosen by the build subagent and used in the promo MP4. Type distribution:
-
-- **palette** (7): socar, gangnamunni, zigzag, ably, banksalad, zigbang, wanted
-- **voice** (2): kakaopay, 29cm
-- **cta** (1): remember
-
-Justifications are written into each `_promo.json` (`label` field) and re-verified during Phase 4 composition build.
-
----
-
-## Follow-up TODO (prioritized)
-
-1. **ably** — full Phase U2 live inspect re-run (clean session). §4 tokens currently `(unverified live)`.
-2. **zigbang** — UPDATE pass on `/home/apt/map` for canonical orange hex + map-overlay tokens.
-3. **socar** — UPDATE pass on `design.socar.kr` for canonical SOCAR Blue hex + Space Frame token system.
-4. **remember** — UPDATE pass capturing filter-chip selected/active + error/warning semantic tokens.
-5. **29cm** — UPDATE pass on PDP + checkout flows.
-6. **kakaopay** — UPDATE pass on a transactional second surface (e.g., /transfer).
-7. **reader UI**: long variant headings overlap preview area (observed on 29cm §5). Fix in `web/src/components/reference-preview.tsx` (or equivalent).
-8. **Process**: serialize Playwright sessions per build agent for future batches.
-
----
-
-## Schema for this file (future batches use the same shape)
-
-Required top-level sections (so they're greppable across batches in `data/reference-audits/`):
-- `## How the 10 were researched`
-- `## Systemic finding`
-- `## Per-brand audit` (table with id, confidence, Tier 1 depth, Tier 2, Known gaps, Follow-up)
-- `## Confidence distribution`
-- `## Known shared limitation`
-- `## Promo highlight selection`
-- `## Follow-up TODO`
-
-Filename: `data/reference-audits/<YYYY-MM-DD>-<batch-slug>.md`

package/.opencode/oh-my-design/data/reference-audits/2026-05-14-kr10.md

@@ -1,72 +0,0 @@
-# Batch Audit — 2026-05-14 / KR finance
-
-**Theme**: Korean digital-first banks (extending the KR-10 set with live-token capture)
-**Skill**: `omd:add-reference` CREATE — with live CDP token harness
-**Verification date**: 2026-05-14
-
----
-
-## Audit rows
-
-| id | name | country | category | url | tier1_live | tier1_official_ds | tier2_indexes | raw_samples | css_vars | flags |
-|---|---|---|---|---|---|---|---|---|---|---|
-| kbank | K bank (케이뱅크) | KR | finance | https://www.kbanknow.com | ✓ CDP :9222 / getComputedStyle on `/ib20/mnu/PBKMAN000000` | ✗ negative — no public DS site (design.kbanknow.com no DNS; no GitHub design-system repo; no Figma Community kit). Documented as authoritative negative result in §11/§12. | ✗ getdesign.md empty; ✗ styles.refero.design empty (kbank + 케이뱅크 both) | 6 | 43 | a11y: no landmark elements; H1 uses font-size:0 image-replacement; fixed 1280px canvas — all logged in §8/§10 |
-| classum | Classum (클라썸) | KR | education | https://business.classum.com | ✓ CDP :9222 / getComputedStyle on `business.classum.com` + production Webflow CSS bundle (13,904 lines) frequency analysis | ✗ negative — no public DS / brand portal; `classum.com/brand` title-only; no Figma library, no design.classum.com, no GitHub design-system repo. Closed B2B SaaS as of 2026-05-14. | ✗ getdesign.md `classum` → "No designs found"; ✗ styles.refero.design empty | 6 | 0 (no `:root` token vars — Webflow-class-driven theming) | 14px Korean-density body sits at WCAG floor; primary CTA `#FF4438` on white ≈4.3:1 (borderline AA); single-color CTA palette shares brand red with destructive — downstream should split semantic vs brand |
-| lunit | Lunit (루닛) | KR | healthcare | https://www.lunit.io | ✓ CDP :9222 / getComputedStyle on `www.lunit.io/` — 34 + 23 targeted samples across hero H1, 7 sections, stats band, 5 CTA candidates, 4 nav links, body paragraphs, footer | ✗ negative — `design.lunit.io` no DNS; GitHub org `lunit-io` has 10 public repos, all medical-imaging / ML research (`archon`, `dicom-rs`, `mmg-model-nia`, `nlst-nodule-detection`, `ocelot23algo`, `openslide`, `spec-cxr`, …), zero design-system / Storybook / tokens repo; HTML keyword scan for `design-system / brand-guidelines / lunit-ui / storybook` returns no hits. Documented as authoritative negative result in §15. | ✗ getdesign.md `lunit` not present; ✗ styles.refero.design empty (lunit + 루닛) | 8 | 0 Lunit-namespaced (only WordPress `--wp-*` defaults) — corroborates "no public DS" finding | Stack = WordPress + Oxygen Builder + AOS. Signature blue `#1032cf` is **withheld from primary CTA** (CTA is black pill `#000` / radius 100px) — unusual brand-color discipline; primary geometric thesis is binary radius `0px` everywhere except CTA pill; LIGHT-weight typography defaults (Lexend 300 / ClashGrotesk 400) anti-pharma. |
-| flex | flex (플렉스) | KR | saas / hr | https://flex.team | ✓ CDP :9222 / getComputedStyle on `flex.team/` — 40 samples across header, nav, hero (`.l2024_dark`), manifesto cards, service filter pills, h1/h2/h3, primary/secondary buttons; 9 curated `raw_samples` in proof | ✗ negative — no `design.flex.team` DNS; `flex.team/design` and `/brand` 404; no public Storybook / Figma library; no GitHub org-level design-system repo. Documented as authoritative negative result in `_research.md`. Corporate data sourced from `thevc.kr/Flex` (founder 장해남, 2019-05 founding, Seongnam HQ, Series B, ~213 employees as of 2026-03). | ✗ getdesign.md `flex` → "No entry"; ✗ styles.refero.design `?q=flex.team` no result cards (also tried `?q=flex` — collides with Flex Ltd. manufacturing, irrelevant) | 9 in `.live-inspect-proof.json` (40 in raw) | 0 Flex-namespaced (Stitches/CSS-in-JS `l2024-c-*` class system, not `:root` custom props) | **One-ink discipline**: `#1D1D1F` + alpha-stepped (0.04/0.10/0.24/0.72/0.96) — no semantic blue/green/red on marketing surface. **Pill rings, not borders**: `box-shadow 0 0 0 1px inset` instead of `border` (no layout shift on hover). **Inverted h1/h2 size**: hero h1=28px, in-page h2=52px with -1.56px tracking (the visual signature). **Theme flip via class**: `.l2024_dark` / `.l2024_light` toggles surfaces section-by-section, not by token swap. Motion tokens partially inferred (entry fade observed, exact durations not captured in single CDP pass) — flagged in §15. |
-| upbit | Upbit (업비트) | KR | finance / crypto-exchange | https://upbit.com | ✓ CDP :9222 / getComputedStyle on **two surfaces** — `/home` (41 samples) + `/exchange?code=CRIX.UPBIT.KRW-BTC` (80 samples) = **121 raw samples** across buttons, anchors, inputs, role=button, role=tab, nav children, market/coin classes, table rows/cells/headers, h1-h3, price classes, Emotion `css-*` classes | ✗ negative — `design.upbit.com` DNS 000; `upbit.com/brand` 301 → marketing root; `design.dunamu.com` DNS 000; `dunamu.com` 200 (corporate vision/IR/ESG, no DS surface); Naver blog `dunamupr` is PR not design; WebSearch / GitHub no canonical hit. Documented as authoritative negative in `_research.md`. | ✗ getdesign.md explicit text "No designs found for 'upbit'"; ✗ styles.refero.design `?q=upbit` no result cards | 121 (12 shipped in `.live-inspect-proof.json`, ≥5 floor) | 263 page-level CSS vars captured (Emotion-generated, not Upbit-namespaced design tokens — corroborates "no public DS") | **Locale-critical** rise/fall convention (KR red=up / blue=down) — any port outside KR/JP/TW must invert. Hit-target trade-off: 28px quick-fill tags below WCAG 2.5.5 (44px) — explicit speed-over-floor decision called out in §9. Legacy IE-era font fallback chain (Dotum / 돋움) not pruned by Upbit — documented, not "fixed". State / motion tokens partially inferred (no live error-state capture this pass) — flagged as follow-up. |
-| wadiz | Wadiz (와디즈) | KR | commerce / crowdfunding | https://www.wadiz.kr | ✓ CDP :9222 attempted (handshake 403 — `--remote-allow-origins` guard); fallback = direct curl of production CSS bundle `static.wadiz.kr/main/main.css` (936473 bytes — authoritative full-token source) + `wui.css` reset + 3 HTML surfaces (`/`, `/web/wreward/category/289`, `/web/main`). All `#00c4c4` (111×), `#191f28` (208×), radius `8px` (112×), `font-weight:700` (358×), and named module classes (`Button_mint`, `Button_blue`, `Button_red`, `Button_yellow`, `Button_contained`, `Button_tint`, `LabelBadge_solid`/`outlined`/`tint`) directly observed via frequency analysis. Reproducible: `curl -sL https://static.wadiz.kr/main/main.css \| grep -c "#00c4c4"` → 111. | △ partial — Wadiz blog (`blog.wadiz.kr`) is known to publish "와디즈 디자인 시스템 구축하기" (referenced in Wadiz internal search snippet), but direct article URL 404 on this pass (likely renamed/moved). No `design.wadiz.kr` DNS, no `/web/wbrand` page, no GitHub org-level design-system repo. Documented as **partial-positive** (existence attested, canonical URL pending re-discovery) in §15. Followup: re-find article URL in UPDATE pass. | ✗ getdesign.md `wadiz` → "No designs found"; ✗ styles.refero.design `?q=wadiz` empty | 8 in `.live-inspect-proof.json` (≥5 floor) — selectors: filled-mint CTA, mint hover, tint mint button, checkbox-checked, yellow LabelBadge solid + tint, body font-family, CardThumbnail | 5 root CSS vars (`--dark`, `--white-68`, `--black-84`, `--black-54`, `--pale-grey-30`) — CSS Modules class system, not heavy token-var usage. Token system lives in named class variants (`Button_mint`, `LabelBadge_yellow`) rather than `:root --tokens`. | **A11y weak point** (documented, not fixed): filled mint CTA `#00c4c4` background + `#fff` text = 2.7:1 (borderline AA fail); outlined mint text `#00c4c4` on `#e6fafa` = 2.4:1 — system as shipped accepts the trade-off, called out in §9. **IP guardrail tight**: no verbatim Wadiz copy; gerund `-하기` verb pattern + factual funding-state vocabulary documented as voice *shape*, not reproduced. **Personas explicitly inferred** (§11 marked FILL-IN). **Detail-page + funding-sheet flows not inspected** — UPDATE pass recommended. CDP fallback to curl is *higher signal* than runtime DOM sampling because the 936KB bundle exposes the entire token system (not just rendered-DOM applied styles). |
-| kream | KREAM (크림) | KR | marketplace | https://kream.co.kr | ✓ CDP :9222 / getComputedStyle on **two surfaces** — `/` (40 raw samples; 50+ `--greyscale-*` CSS custom properties extracted from `:root`) + `/search?keyword=nike` (43 raw samples covering filter chip, category pill, shortcut row, search input, tab strip) + `/shop` 404 ghost-button chrome | ✗ negative — no `kream.design` / `design.kream.co.kr` / `kream.co.kr/design`; no public DS hub (parent Naver has mixed-public pattern — Naver Connect/Snow have docs, KREAM does not). Documented as authoritative negative result in §9. | ✗ getdesign.md `kream` "No record found"; ✗ styles.refero.design `?q=kream` no match (both verified 2026-05-14) | 13 raw_samples in `.live-inspect-proof.json` (over ≥5 floor; total 83 across two surfaces) | ~50 `--greyscale-dark-*` + `--greyscale-white-alpha-*` CSS custom properties on `:root` — formal grayscale token system with both numeric scale (0/2/4/6/8/10/20/30/35/40/50/60/70/75/80/90) and semantic role naming | **Grayscale-only chrome** — no brand red/blue/green in product chrome (single observed teal `rgb(17,161,151)` used twice as watch accent). **Four-step radius scale encodes card class** (16px banner / 30px chip / 8px ghost / 6px row). **Weight, not hue, signals state** (active tab 700 vs default 400, both `#222`). §10 voice samples are **OmD-original characterizations** — no verbatim KREAM marketing copy reproduced. §11 narrative is reported-fact based (no founder/exec quote at OmD attribution fidelity in EN-language public sources). State matrix and product-detail page not captured this pass — flagged for UPDATE. |
-| inflearn | Inflearn (인프런) | KR | education | https://www.inflearn.com | ✓ CDP :9222 / getComputedStyle on **two surfaces** — `/` (12 raw_samples, 162 `--mantine-color-*` CSS vars + full `--mantine-radius-{xs,sm,md,lg,xl}` + `--mantine-spacing-*` + `--mantine-font-family`, h1 34px/700 / h2 20px/700 sampled, GNB sticky 65px white, search input r=8px, primary green pill submit r=999px `#00C471`) + `/courses` (8 raw_samples, 38 `mantine-Card-root` tiles observed, pagination active page solid `#00C471`). Brand-owned context: `tech.inflab.com/20260305-new-header/` engineering blog post confirming GNB rebuild as shared MFE component. | ✗ negative — no `design.inflearn.com` DNS; Inflab GitHub org has no public DS / Storybook / component-library repo (verified `gh search repos inflab` / `inflearn design` — only course-content repos surface); `tech.inflab.com` has engineering posts (MFE / GNB / infra) but no canonical token table. The DS lives **in production code as a Mantine UI theme override** — accessible only by runtime inspect. Documented as authoritative negative in §4 footer + `_research.md`. | ✗ getdesign.md `inflearn` → "No designs found for 'inflearn'" (verified 2026-05-14); ✗ styles.refero.design `?q=inflearn` no result cards (verified 2026-05-14) | 20 raw_samples in `.live-inspect-proof.json` (well above ≥5 floor — 12 home + 8 courses) | 162 `--mantine-color-*` CSS vars + 5 `--mantine-radius-*` + 5 `--mantine-spacing-*` + `--mantine-font-family` exposed on `:root` — Inflearn theme-extends Mantine UI wholesale | **Framework signal authoritative**: Inflearn is a Mantine UI app; the "DS" is the Mantine theme primitive set with the `green` scale overridden to `#00C471`. Designers picking from Inflearn should treat the Mantine semantic palette (red/blue/yellow/cyan) as part of the system. **Two radius families** are the visual signature: `8px` for cards/inputs/pagination, `32px` for nav pills, `999px` for the green search submit + tag chips. **`#00C471` fails AA for body text** (2.45:1 on white) — Inflearn correctly restricts it to white-on-green CTAs (4.4:1, passes AA), pagination active fills, and >18pt accents; called out in §9. **IP guardrails**: tagline "라이프타임 커리어 플랫폼" quoted once as factual brand-narrative context (§11), not adopted as design copy; voice samples in §10 are fresh illustrative reconstructions in casual-polite Korean. **Inferred tokens** flagged `(Inferred)` in §4: labelled primary CTA (only icon-pill captured live), form focus, disabled, error-focus, tag chip alpha. Follow-up UPDATE pass recommended on a course-detail + payment surface. |
-| bunjang | Bunjang (번개장터) | KR | marketplace | https://m.bunjang.co.kr | ✓ CDP :9222 / getComputedStyle on `m.bunjang.co.kr` across **two viewports** — desktop 1280×713 dpr=2 + mobile-emulated 390×844 dpr=3 (iPhone UA). 211 DOM samples across header / nav / cards / buttons / footer / heart / price / meta. Websocket required `suppress_origin=True` to bypass `--remote-allow-origins` block. | ✗ **negative (documented)** — Bunjang ships a real internal DS branded **"Bun UI"** (proven by `--bun-ui-*` CSS-variable prefix + vanilla-extract class patterns `Box__7nn0kn17` / `Flex__wsrgth3` / `Typography_typography__1wr8iu13` + explicit z-index/radius/aspect-ratio token families), but **no public artifacts as of 2026-05-14**. Probed: `design.bunjang.co.kr` / `bun-ui.bunjang.co.kr` / `tech.bunjang.co.kr` / `brand.bunjang.co.kr` / `company.bunjang.co.kr` / `blog.bunjang.co.kr` all DNS no-resolve (000). `medium.com/bunjang-tech` → 404. `github.com/bunjang` org exists but **0 public repos / 0 members / 0 packages**. `npmjs.com/package/@bunjang/bun-ui` → 403 (does not exist). Authoritative negative documented in DESIGN.md verification footer. | ✗ getdesign.md `bunjang` → "No designs found for 'bunjang'"; ✗ styles.refero.design `?q=bunjang` no entry. Consistent with KR-10 systemic finding. | 9 raw_samples in `.live-inspect-proof.json` (≥5 floor) covering: primary CTA `#d80c18` bg, body color `#191919`, footer bg `#f6f6f6`, Pretendard Variable on body, secondary-button `1px solid #e5e5e5` + 6px radius, card price 16px/700/`#191919`, card title 14px/500/`#666`, card meta 12px/500/`#999`, circular icon button 999px radius | **73 CSS custom properties** on `:root` — full 11-stop gray ladder (`--color-gray-{10..900}`), 4-stop semantic ladders (`--color-{red,green,blue,yellow}-{50,100,400,500}`), **two specialty ladders distinctive to Bunjang**: `--color-care-*` (warm amber `#FDF4E2`→`#FFE1A6` for 감정완료 luxury-auth chips) and `--color-safe-*` (cool indigo `#EFF2FE`→`#5558A8` for 안전결제 escrow chips). 8 social-partner brand colors. Layout max-widths (480/640/1240). Drawer 480px. Aspect-ratio tokens `--bun-ui-aspect-ratio-{vertical: 81/100, square: 1/1}`. Radius token `--bun-ui-radius-pill: 999px`. Full z-index scale (header 200 / drawer 1000 / popup 1200 / snackbar 1500). iOS safe-area vars `--bun-ui-{sat,sab,sal,sar}`. | **One red, one job**: `--color-primary: #d80c18` appears on exactly one CTA variant (`_variant-primary`) — red as scarcity asset, cleanest example in the Korean-marketplace cohort. **Zero box-shadow** on any sampled production element — depth is borders + tints exclusively. **81:100 portrait thumbnail** is brand-defining (phones/sneakers/K-pop goods photograph tall). **Two specialty color tiers map to revenue services** (care → 감정/luxury-auth, safe → 안전결제/escrow) — rare case where color IS product taxonomy. **Pretendard Variable exclusive** (100% of 211 samples). **Weight jump 500→700 without 600** accelerates price emphasis. IP guardrail: logo SVG archived for verification only; no verbatim taglines or product copy reproduced; service-feature names (안전결제 / 감정완료 / 내폰시세) used descriptively. Bottom tab bar / sell flow / chat / product-detail page not inspected this pass — flagged for UPDATE. |
-| channeltalk | Channel Talk (채널톡) | KR | saas | https://channel.io/en | ✓ CDP :9222 / getComputedStyle on `channel.io/en` — 118 element samples, 559 CSS custom properties extracted from `:root`, 10 raw_samples retained | ✓ **positive — canonical Bezier Design System** at `github.com/channel-io/bezier-react/packages/bezier-tokens/src/{global,semantic}` (MIT-licensed). Fetched via `gh api`: color.json (6.7KB, full 12-hue × 3-weight × 4-alpha ladder), font.json (3.0KB, Inter+Noto stacks KR/JP), typography.json (2.7KB, 11-36px size scale), radius.json (463B, 2-44px+42% ladder), transition.json (945B, 150/300/450ms × `cubic-bezier(0.3,0,0,1)`), elevation.json (2.3KB), input.json (1.5KB). **Gold-standard Korean SaaS DS publishing — first Tier-1-positive in this audit file.** | not consulted — Bezier is canonical Tier-1, third-party indexes redundant; flagged in DESIGN.md footer for follow-up UPDATE | 10 raw_samples in `.live-inspect-proof.json` (≥5 floor; from 118 total live captures) | 559 `--color-*` / `--inverted-*` / `--bg-*` / `--bdr-*` / `--bgtxt-*` CSS custom properties on `:root` (full alpha + inverted-theme + semantic-role naming) | **Cobalt-only brand accent** `#329BE7` — semantic hue palette reserved for status. **Translucent off-black** `rgba(0,0,0,0.85)` body text (141/118 sampled elements). **Marketing-vs-product type cliff**: 64px hero vs 36px product cap; weights 500/600 marketing-only. **AI manifesto plum canvas** `rgba(25,3,49,0.898)` is homepage-only rhetorical device, not a Bezier token. §11 founder list / founding year / total funding / HQ city partially unverified (Crunchbase 403'd on WebFetch; only current CEO Choi Si-won confirmed Channel-side) — flagged in DESIGN.md footer. IP guardrail: §10 voice samples explicitly marked tone-shape-only with 3 fresh derivation alternatives provided inline ("Routine, automated. Strategy, yours." / "AI for the queue. You for the moves that matter." / "Every channel. One inbox."). |
-
----
-
-## Artifacts written
-
-- `references/kbank/DESIGN.md` — frontmatter `id=kbank, country=KR, category=finance, verified=2026-05-14`
-- `references/kbank/assets/_reference/tokens.json` — normalized brand/neutral/semantic/extended palettes + typography + radius + spacing
-- `references/kbank/assets/_reference/fonts.json` — Pretendard K Edition + fallback stack
-- `references/kbank/assets/_reference/structure.json` — IA tree, heading outline, layout observations
-- `references/kbank/assets/_reference/.live-inspect-proof.json` — 6 raw_samples (≥5 required) + 43 CSS variables + top observed colors
-- `references/kbank/assets/_reference/.live-inspect-raw.json` — full CDP capture (provenance)
-- `references/lunit/DESIGN.md` — frontmatter `id=lunit, country=KR, category=healthcare, verified=2026-05-14`
-- `references/lunit/assets/_reference/tokens.json` — color (ink/surface/accent) + typography scale + binary radius rule + zero-elevation + AOS motion
-- `references/lunit/assets/_reference/fonts.json` — ClashGrotesk (display) + Lexend (text), both LIGHT-weight defaults
-- `references/lunit/assets/_reference/structure.json` — 11-band page anatomy (sticky header → hero → stats → tech overview → solutions grid → partnering → testimonials → scale → investor → news → footer)
-- `references/lunit/assets/_reference/.live-inspect-proof.json` — 8 raw_samples (≥5 required), CDP method + browser version + page URL
-- `references/upbit/DESIGN.md` — frontmatter `id=upbit, country=KR, category=finance, omd=0.1, captured_at=2026-05-14`; 9 product sections + §10-15 philosophy (toss style-ref per skill rule P-1)
-- `references/upbit/_research.md` — Tier-by-Tier source map + confidence per § + IP guardrail log
-- `references/upbit/assets/_reference/tokens.json` — color (brand triple + rise/fall semantic + text + surface) + radius + typography + control sizing
-- `references/upbit/assets/_reference/fonts.json` — Roboto + Noto Sans KR primary stack, legacy fallback chain documented
-- `references/upbit/assets/_reference/structure.json` — home + exchange structural counts + 5 layout notes
-- `references/upbit/assets/_reference/.live-inspect-proof.json` — 12 raw_samples (≥5 required), 2 tab ids, 2 surface URLs, WS endpoint, total 121 samples
-- `references/upbit/assets/_reference/raw-capture.json` — home surface (41 samples)
-- `references/upbit/assets/_reference/raw-capture-exchange.json` — exchange surface (80 samples)
-- `references/flex/DESIGN.md` — frontmatter `id=flex, country=KR, category=saas, omd=0.1, verified=2026-05-14`; 9 product sections + §10-15 philosophy (toss style-ref per skill rule P-1)
-- `references/flex/_research.md` — Tier 1 (live CDP + thevc.kr) and Tier 2 (getdesign/refero both empty) source map
-- `references/flex/assets/_reference/tokens.json` — single-ink alpha-stepped color, Pretendard Variable scale, ring shadows, section paddings
-- `references/flex/assets/_reference/fonts.json` — Pretendard Variable + full system fallback chain
-- `references/flex/assets/_reference/structure.json` — header / hero / manifesto / service filter pill anatomy + tone_observed
-- `references/flex/assets/_reference/.live-inspect-proof.json` — 9 raw_samples (≥5 required) + CDP method + browser version
-- `references/flex/assets/_reference/.raw-inspect.json` — full 40-sample CDP dump (provenance)
-- `references/wadiz/DESIGN.md` — frontmatter `id=wadiz, country=KR, category=commerce, verified=2026-05-14`; 15 sections (§1 visual theme + 9 product sections + §10–15 philosophy/personas/anti-patterns/footer per toss style-ref); confidence **High**
-- `references/wadiz/assets/_reference/tokens.json` — color (brand mint family + Toss-family ink scale + 4-hue label badge system + surface + overlay) + radius scale (3/4/8/16/50%/100px with frequencies) + shadow tiers + typography sizes + weight frequency
-- `references/wadiz/assets/_reference/fonts.json` — Pretendard exclusive (no display accent), cdnjs 1.3.9 dynamic-subset, base body 15px, weights 400/500/700
-- `references/wadiz/assets/_reference/structure.json` — 3 surfaces, CSS Modules naming pattern, 5-axis button system, 4-color × 3-shape LabelBadge system, card anatomy, 5-tier elevation
-- `references/wadiz/assets/_reference/.live-inspect-proof.json` — 8 raw_samples (≥5 floor), 5 surfaces, full token-frequency table, CDP fallback rationale documented
-
-## IP guardrails confirmed
-
-- Brand assets reference-only (logo/name/navy `#0114A7` for inspiration; not redistributed).
-- §9 Voice is fresh analyst paraphrase — no verbatim taglines or marketing copy lifted.
-- Token values reproduced as factual CSS custom property values under analytical fair-use.
-
-## Flags
-
-1. **Tier 1 official DS = negative**. K bank has chosen not to externalise a public design system (unlike Toss/Kakao). The production CSS `:root` token set is the closest public artifact and was captured directly.
-2. **Tier 2 indexes empty** — consistent with the systemic Korean-coverage gap logged in `2026-05-13-kr10.md`.
-3. **A11y debt observed**: no landmark elements, image-replacement H1 logo, fixed 1280px canvas. Surfaced in §8 and §10 with explicit "don't steal" guidance.
-4. **Upbit locale rule (critical)**: Korean finance UI convention — red `#DD3C44` = up / 매수, blue `#1375EC` = down / 매도. Inversion of US convention. Surfaced in §3 and §9 of `references/upbit/DESIGN.md` with explicit porting guidance (must invert for US/EU markets).
-5. **Upbit IP guardrails**: title-tag positioning phrase ("가장 신뢰받는 디지털 자산 거래소") referenced only as positioning evidence in §11 narrative — not reproduced as headline copy; voice samples in §10 marked illustrative; persona block carries public-surveys-only disclaimer.