world-2026-hub/.agents/project-memory.md

# Project Memory — World Cup 2026 Hub

Persistent memory for this project. Read this before any significant change.

---

## Context

Static web app showing the FIFA World Cup 2026 (Mexico/USA/Canada, 48 teams) — schedule, group standings, interactive knockout bracket with user simulation, stadiums. Hosted on GitHub Pages, all content from `data/*.json`. Started 2026-06-11 from two spec documents; built step-by-step with user approval between steps.

### What this project is
- A **personal/portfolio** piece — visual polish (glassmorphism, animations) is a primary goal, not a nice-to-have.
- A **static SPA**: one `index.html`, ES-module vanilla JS, JSON as the only "database".
- Maintained by editing JSON only — code should never need touching to update scores/teams.

### What this project is **not**
- No backend, no database, no build step, no bundler, no CDN dependencies, no frameworks.
- No automated tests, no linter (explicit spec constraint).
- Not real-data-complete: ships with mock data (fictional teams) to be replaced later.

---

## Priority objectives

1. **Spec compliance** — both spec files define scope; `complement-spec-worldcup2026-en.md` wins on conflict.
2. **Visual quality** — FIFA/UCL/Apple-inspired, glassmorphism, smooth animations; portfolio-grade.
3. **Interactive bracket** — hover path highlight, zoom, drag, simulation mode; the centerpiece feature.
4. **Easy maintenance** — real data drop-in via JSON; `bracket-config.json` is the only file edited after group stage.
5. **Performance/accessibility** — Lighthouse > 90, first render < 2s, JS < 300KB, ARIA + keyboard nav.

---

## Technical decisions and rationale

### Stack
- **Vanilla HTML/CSS/JS ES2022+, ES Modules** — spec mandate; GitHub Pages serves static files only. Frameworks/bundlers explicitly forbidden.
- **EN/PT-BR UI toggle via `i18n.js`** (user decision 2026-06-11) — not in spec; spec UI examples are EN, user wants both. Tiny dict + `t(key)`, persisted in `wc2026_prefs.lang`. Alternative (EN only) rejected by user.
- **`storage.js` pulled forward to step 2** (spec places it in phase 12) — prefs (`lang`, `lastTab`) are needed from the base-layout step; building it late would mean refactoring.

### Data model
- **All match times in UTC** in `matches.json`; converted at render via `Intl.DateTimeFormat` (`formatMatchTime`). `.ics` export depends on this.
- **Knockout matches carry `bracketRef` instead of teams** — teams resolved at runtime from standings + `bracket-config.json`; rounds after R32 have no config, generated by sequential pairing of winners (indices 0-1 → 0, 2-3 → 1, …).
- **Simulation never mutates JSON** — overlay stored in `localStorage` `wc2026_simulation`, keyed by bracket match ids (`R32-1`, `QF-2`, `FINAL`, …).

### Mock data design (2026-06-11, step 1)
- **Real country names, fictional results** — generated deterministically (seed 2026), script deleted after run; data is now static JSON.
- **State crafted to test both bracket modes:** matchdays 1-2 all `finished`; matchday 3 `finished` for groups A–F, `scheduled` for G–L (match 61, Group G, is `live`). So R32 slots fed by A–F resolve to real teams; G–L and all `third` slots show placeholders.
- **Knockout results:** `R32-2` (match 74) finished 1-1 + penalties 4-3; `R32-4` (match 76) finished 2-0. Everything else `scheduled` with `null` scores — `results.json` has an entry for **all 104 matches**.
- **Image paths:** `flag`/`image` JSON values are relative to `assets/images/` (e.g. `flags/mex.svg`, `stadiums/azteca.svg`).
- **Opener** (match 1) is MEX at Estadio Azteca 2026-06-11; **final** (match 104) at MetLife 2026-07-19.

### Base layout decisions (2026-06-11, step 2)
- **Hero priority: live > next scheduled** — a live match replaces the countdown with the score + pulse badge. Mock data keeps match 61 permanently `live`, so the countdown only shows if that status is changed (verified working by temporarily setting it to `scheduled`).
- **i18n mechanics:** static HTML uses `data-i18n` / `data-i18n-aria` attributes re-applied by `applyI18n()`; dynamic renders call `t()` and listen for the `langchange` event on `document`. Phases translate via `translatePhase()` (PT: R32 = "16 avos de final").
- **Tab routing:** hash (`#matches`) + `wc2026_prefs.lastTab`, `history.replaceState` to avoid history spam; precedence on load: hash → lastTab → home.
- **Default language:** `navigator.language` startsWith `pt` → PT, else EN; only persisted when the user clicks the toggle.
- **Preview server:** `.claude/launch.json` at `R:\lucas-kalil\Projects\` defines `worldcup2026` (python http.server, port 8126) for the Claude Preview panel.

### Schedule + performance decisions (2026-06-11, step 3)
- **`schedule.js` ⇄ `app.js` circular import is intentional** — `app.js` calls `initSchedule()`, `schedule.js` imports `getData`/`formatMatchTime`/`flagSrc` back. Safe in native ESM because all calls happen after both modules evaluate; keep this pattern for `groups.js`/`bracket.js`/`modal.js`.
- **Filter UX:** toolbar is rebuilt only on init/langchange/clear (state restored programmatically); list re-renders on every filter change. Filter state is in-memory only (not persisted) by design.
- **Knockout cards show "TBD"** until step 7 swaps `teamColumnHTML`'s lookup to `resolveBracketTeams()`.
- **Perf: no `backdrop-filter` on repeated cards** — `.match-card` overrides `.glass` blur (huge paint cost × 104 cards, invisible over the smooth gradient). Same rule applies to any future card grid (stadiums, bracket).
- **Perf: fixed gradient lives on `body::before` (position: fixed)**, not `background-attachment: fixed` (the latter repaints the whole background on scroll).

### Standings decisions (2026-06-11, step 4)
- **Only `status: "finished"` matches count toward standings** — live scores are ignored until full-time (keeps standings stable and bracket resolution deterministic).
- **Tiebreak order:** points → goal difference → goals for → team id alphabetical (stable fallback). Verified against an independent Python computation for groups A and G.
- **`computeStandings()` / `isGroupFinished()` exported from `groups.js`** — bracket.js (step 7) must import these instead of recomputing.

### Stadiums decisions (2026-06-11, step 5)
- **`stadiums.js` module added** — spec §4 has no module for the stadiums view; a dedicated view module keeps the per-view pattern (schedule/groups/bracket) instead of growing `app.js`.
- **Cross-link:** "View matches" on a stadium card calls `setStadiumFilter(name)` (exported by `schedule.js`, resets other filters) + `navigateTo('matches')` (exported by `app.js`).

### Modal decisions (2026-06-11, step 6)
- **Native `<dialog>` + `showModal()`** — focus trap, Esc-to-close and `::backdrop` come free; no custom trap code. Backdrop click detected via `event.target === dialog` (content is in a padded inner div).
- **Focus restore:** opener element saved in `openMatchModal()` and re-focused on `close`.
- **Card → modal wiring is event delegation on `#schedule-root`** (click + Enter/Space keydown), so it survives list re-renders. Cards got `tabindex="0" role="button" aria-label` already in this step.
- **`openMatchModal(matchId)` is the public API** — bracket (step 7) and any future view should call it rather than building their own.

### Bracket decisions (2026-06-11, step 7)
- **Tree model is language-neutral**: slots are `{ teamId }` or `{ ph: { kind, … } }`; placeholder text is produced at render time by `slotDisplay()` so language switches never invalidate the tree.
- **Tree is cached** in `bracket.js`; `invalidateBracket()` exists for the simulation overlay (step 9). Results are static per page load, so nothing else invalidates it.
- **`resolveBracketTeams(matchOrRef)`** accepts a match object (group or knockout) or a bracketRef string and always returns `{ home, away }` as `{ team: Team|null, label: string }` — schedule cards, modal, and search/team filters all consume this, so knockout matches become searchable/filterable once resolved.
- **Connector geometry**: all bracket columns share equal height with `flex: 1` slots, so pair children sit at 25%/75% and the next round's node at 50% — pure-CSS connectors (`::before`/`::after` stubs + pair vertical) meet exactly. Column gap 44px = 22px out-stub + 22px in-stub. Breaking the equal-height invariant breaks the lines.
- **Final column** holds champion box (top), FINAL (middle, aligns with SF pair), third-place block (bottom); its champion/third slots suppress the incoming connector stub.

### Bracket interaction decisions (2026-06-11, step 8)
- **"Full path" on hover/focus** = hovered node + its entire feeder subtree + its winner's route to the FINAL; THIRD-PLACE lights both SFs. Non-path nodes dim via `.has-path` on the canvas. Computed from ref arithmetic (`floor(i/2)` up, `2i/2i+1` down), no tree lookup.
- **Zoom = CSS `transform: scale()` on the canvas + a `#bracket-zoom` box sized to `natural × scale`** (transform doesn't affect layout, so the box gives the scroll container the right scrollable area). Pointer-anchored: scroll adjusted so the point under the cursor stays put. Clamped 0.4–2.
- **Natural canvas size is measured lazily** (`ensureMeasured()`) because the bracket panel can be `hidden` at render time (offsetWidth 0).
- **Pan + pinch via Pointer Events** with `touch-action: none` on the wrap (we own all gestures there; page scroll over the bracket is intentionally captured, like a map widget).
- **Drag–click conflict:** >5px movement sets a flag; a capture-phase click listener on the wrap swallows the click that ends a drag. The flag resets on the next `pointerdown`, so synthetic `el.click()` without pointerdown can be falsely suppressed in tests — dispatch pointerdown/up first.
- **Zoom level survives langchange re-renders** (module-level `view.scale`) but intentionally not reloads (not in prefs).

### Simulation decisions (2026-06-11, step 9)
- **Separation of concerns in the tree:** `decide()` applies only real finished results; `applySimulation()` overlays user picks afterwards and never overrides a real result. Stale entries (winner no longer among the resolved teams) are silently ignored — same validation the prediction import (step 12) will use.
- **Sim UX:** "Simulation" toggle in the bracket toolbar; eligible nodes (both teams resolved, real result still `scheduled`) get dashed blue borders; clicking one opens a small native `<dialog>` picker. An unequal score auto-selects the winner; a draw requires an explicit pick (penalties implied). Empty score defaults to 1-0 for the picked winner.
- **Storage format** is exactly the complement-spec shape: `wc2026_simulation = { "R32-6": { winner: "FRA", score: "2-1" } }` keyed by bracketRef, score oriented home-away.
- **`simchange` custom event** fires after any pick/reset; `schedule.js` listens and re-renders so simulated teams appear on knockout cards too (intentional leak — resolved tree is the single source).
- **Simulated nodes** show a blue "SIM" corner chip and blue scores; the modal shows none of this (it reads real results only).

### Responsive/a11y decisions (2026-06-11, step 10)
- **Breakpoints:** ≤767 (tight spacing, bracket `--node-w: 168px`/gap 36px — connector stub offsets must stay at gap/2, overridden in the same media query), 768–1439 (single-row header, centered menu), 1440+ (container widens to 1360px).
- **Tabs follow the WAI-ARIA pattern:** roving tabindex + ArrowLeft/Right/Home/End in `initTabs()`; focus follows activation.
- **Dialogs get `aria-label`** set at open time (match name + phase); schedule count is `aria-live="polite"`; countdown has `role="timer"` + label.
- **Entry animations:** every unhidden `.panel` fades in; card grids (`.match-grid/.groups-grid/.stadiums-grid > *`) slide up with a 45ms stagger on the first 6 children, 260ms for the rest. All killed by `prefers-reduced-motion`.

### Extra features decisions (2026-06-12, step 12 — done before step 11 at user request)
- **Favorites:** single global capture-phase click delegation in `app.js` handles every `.fav-btn` (schedule, groups, modal) and dispatches `favchange`; each view re-renders itself. Stars never trigger the card/modal click (guard via `closest('.fav-btn')`). Bracket shows highlight only (no stars — nodes too small). Favorite involvement = gold left border.
- **`getFavoriteMatches(matches, favorites)`** lives in `bracket.js` (needs `resolveBracketTeams`), imported by `schedule.js` for the "My matches" filter.
- **Time mode:** header `#time-toggle` flips `wc2026_prefs.timeMode` and dispatches `timemodechange`; `formatMatchTime()` already defaulted to the pref, so views just re-render.
- **Challenge:** sim entries for real-finished matches can't be created in the UI (locked) but old ones persist in storage — that's by design: predictions are made while matches are `scheduled`, then scored when results land. Card renders only when ≥1 finished knockout match exists.
- **Share/import:** `?prediction=` is stripped from the URL via `history.replaceState` whether applied or not (prevents re-prompt loops). Declining `confirm()` keeps local picks; unknown refs are rejected wholesale.
- **`.ics`:** RFC 5545 TEXT escaping (`\,` etc.) applied even though the spec template shows raw commas — RFC compliance wins; verified output imports with CRLF-only endings.
- **Custom events now in play:** `langchange`, `simchange`, `favchange`, `timemodechange` — all on `document`; views own their re-renders.

### Build complete (2026-06-12, step 11 — all 12 steps done)
- README is the user-facing manual (run, deploy, JSON maintenance, localStorage keys, acceptance checklist). Keep it in sync when data formats change.
- **Verified at completion:** spec §18 criteria all pass; JS = 74 KB total (budget 300 KB); no root-absolute paths (GitHub Pages safe). **Not yet verified:** Lighthouse > 90 (needs a deployed URL or local Lighthouse run); actual GitHub Pages deploy.
- No commits made yet — repo initialized but empty; commit when the user asks.

### Workflow
- **12-step build plan with approval gates** — user approves each step before the next starts; summary after each step. Plan: `C:\Users\Lucas\.claude\plans\read-r-lucas-kalil-projects-web-worldcup-goofy-meerkat.md`.
- **Git repo, commits only when the user asks.**

---

## Known gotchas

### 1. `fetch()` of JSON fails on `file://`
**Where:** any local testing of `index.html`
**Why:** browsers block `fetch` of local files (CORS/origin rules)
**Symptom if forgotten:** blank app, console CORS errors, wasted debugging
**Solution:** always serve via `python -m http.server` (or any static server) from the project root

### 2. `.ics` requires CRLF line endings
**Where:** `assets/js/calendar.js`
**Why:** RFC 5545 mandates `\r\n` between lines; some calendar apps reject `\n`
**Symptom if forgotten:** exported event silently fails to import in Outlook/Apple Calendar
**Solution:** join VCALENDAR lines with `\r\n` explicitly

### 3. Third-place slots are `null` until defined
**Where:** `data/bracket-config.json` → `thirdPlaceAssignment`
**Why:** the 8 best third-place teams are only known after the group stage
**Symptom if forgotten:** crash or "undefined" team names in R32 rendering
**Solution:** `resolveBracketTeams()` must return placeholder labels ("Best 3rd #1", "Group A Winner") whenever a slot is `null` or the group isn't finished

### 4. Claude Preview screenshots can hang (tooling, not app)
**Where:** Claude Preview panel during verification
**Why:** the preview window's screenshot pipeline occasionally gets stuck; `preview_eval` keeps working
**Symptom if forgotten:** wasted debugging hunting a nonexistent app freeze
**Solution:** `preview_stop` + `preview_start` recovers it; verify state via `preview_eval` first before suspecting the app

### 5. Stale JS modules in the dev browser
**Where:** any JS edit while previewing via `python -m http.server`
**Why:** the server sends no cache headers, so browsers heuristically cache ES modules; a normal reload can keep serving old code
**Symptom if forgotten:** "module does not provide an export" errors or old behavior despite correct code on disk
**Solution:** `Promise.all(files.map(f => fetch(f, { cache: 'reload' })))` then `location.reload()`, or DevTools hard reload

### 6. `setPointerCapture` on pointerdown kills element clicks
**Where:** `assets/js/bracket.js` drag/pan handling on `#bracket-wrap`
**Why:** capturing a pointer retargets the eventual `click` event to the capture element, so delegation via `event.target.closest('.bracket-match')` never matches — modal and simulation clicks silently die
**Symptom if forgotten:** bracket nodes unclickable with real input while synthetic `el.click()` tests still pass
**Solution:** capture only after the drag threshold (>5px) is exceeded, inside `pointermove`, wrapped in try/catch. **Always verify click flows with `preview_click` (trusted input), not `element.click()`.**

### 7. GitHub Pages serves under a subpath
**Where:** all asset/data URLs in `index.html` and JS `fetch` calls
**Why:** project pages live at `https://<user>.github.io/<repo>/`, so root-absolute paths (`/data/...`) break
**Symptom if forgotten:** works locally, 404s on GitHub Pages
**Solution:** use relative paths (`data/matches.json`, `assets/...`) everywhere

### 8. Claude Preview: resize beyond the native window breaks clicks/screenshots
**Where:** Claude Preview panel, `preview_resize` wider than the native window (~791 CSS px)
**Why:** viewport emulation desyncs the capture/click surface from the page (screenshots show the page squeezed in the left half; `preview_click` lands on wrong coordinates and silently does nothing)
**Symptom if forgotten:** "clicks don't work" / half-black screenshots at desktop widths, wasted debugging — the app itself is fine (`preview_eval` geometry confirms)
**Solution:** at emulated widths > native, navigate via `preview_eval` + exported `navigateTo()` and verify geometry via eval/inspect; trust screenshots only at widths ≤ native. `preview_resize preset: desktop` resets to native and fixes clicks.

---

## Patterns for future changes

### Match stats no modal (2026-06-14)
- **Campo opcional `stats` em `results.json`** por jogo: `{ possession: {home,away}, shots: {home,away}, cards: {home,away} }` (`home`/`away` seguem `homeTeam`/`awayTeam` de `matches.json`; valores = posse em %, finalizações totais, cartões amarelos). Primeiro preenchido no jogo 6 (BRA 1–1 MAR) como teste do deploy: posse 51/49, finalizações 12/14, cartões 2/0 (fontes ESPN/Opta/Sofascore).
- **`modal.js` renderiza stats reais quando `result.stats` existe**; senão mantém o placeholder com `—` + nota `modal.statsSoon` (backward-compatível, verificado nos jogos 6 com stats e 4 sem). `statRow(home, label, away)` agora recebe valores; antes só o label.
- Reusa as chaves i18n já existentes: `modal.possession` / `modal.shots` / `modal.cards`. Adicionar stats a mais jogos = só editar `results.json`, nenhum código muda.

### How to update real-world data (scores, schedule)
Follow `how-refresh-data.md` (project root). In short:
1. Edit `data/results.json` (scores/status) or `data/matches.json` (schedule, rare).
2. Once group stage ends: fill `data/bracket-config.json` → `thirdPlaceAssignment` (slot → group letter). Nothing else changes.

### Real-data migration (2026-06-12)
- `how-update.md` (project root) is the full runbook for replacing mock `data/*.json` with real World Cup 2026 data: file-by-file schemas, order of operations (stadiums → teams → groups → bracket-config.round32 → matches → results → thirdPlaceAssignment), and a cross-file integrity checklist (group membership, id ranges, bracketRef uniqueness, stadium name/city matches).
- Flags one open decision: `stadiums.json` has 30 entries (original bid shortlist) vs. the 16 venues actually used by the real tournament — confirm with user whether to trim before/while editing `matches.json`.

### Real-data migration DONE (2026-06-12)
- **All 6 `data/*.json` files now hold real WC2026 data.** Sources: Wikipedia per-group + knockout-stage articles (primary; local kickoff times with explicit UTC offsets), cross-checked vs ESPN/NBC/FOX/olympics.com/lumenfield.com. Full smoke test + desktop/mobile layout pass, console clean.
- **Stadiums trimmed 30 → 16** (user decision). Cities use FIFA host-city names ("New York/New Jersey", "San Francisco Bay Area", "Boston"), not suburb names (East Rutherford, Santa Clara, Foxborough) — `matches.json` and `stadiums.json` must keep matching exactly.
- **Match ids:** group matches 1–72 are *chronological by UTC kickoff* (≠ 6-per-group blocks); knockout ids 73–104 are FIFA's official match numbers.
- **bracket-config app-order ↔ FIFA mapping:** R32-1..16 = FIFA matches 74, 77, 73, 75, 83, 84, 81, 82, 76, 78, 79, 80, 86, 88, 85, 87 (ordering chosen so the app's sequential pairing reproduces the official R16/QF/SF progression).
- **Third-place slots → allowed groups** (for filling `thirdPlaceAssignment` after the group stage, per official draw): slot 1 (M74) A/B/C/D/F · slot 2 (M77) C/D/F/G/H · slot 3 (M81) B/E/F/I/J · slot 4 (M82) A/E/H/I/J · slot 5 (M79) C/E/F/H/I · slot 6 (M80) E/H/I/J/K · slot 7 (M85) E/F/G/I/J · slot 8 (M87) D/E/I/J/L. Each group letter may appear in only one slot.
- **Results as of 2026-06-12:** ids 1–3 finished (MEX 2–0 RSA, KOR 2–1 CZE, CAN 1–1 BIH); USA–PAR (id 4) kicked off 01:00 UTC Jun 13 — first thing to update next session.

### Daily refresh (2026-06-13)
- **Results updated through match id 6** (matchday 1 complete):
  - id 4: USA 4–1 PAR (Group D) — confirmed FIFA + Yahoo/ESPN
  - id 5: QAT 1–1 SUI (Group B) — confirmed FOX Sports + ESPN (Khoukhi 94th-min header)
  - id 6: BRA 1–1 MAR (Group C) — confirmed FOX Sports + NBC Sports (Saibari for MAR, Vinicius Jr. for BRA)
- Verified standings in Groups view: Group B shows Qatar/Switzerland each 1 pt; Group C shows Brazil/Morocco each 1 pt; Group D confirms USA 3 pts (W 4-1), Paraguay 0 pts.
- **Single-source caveat:** R16 match 94 (Jul 6, Lumen Field) time 17:00 PDT per Wikipedia; one ESPN summary implied 14:00 PDT. Re-verify when R16 nears.
- Next: matches 7–8 scheduled Jun 14 (HAI–SCO, AUS–TUR, both Group stage). Continue daily routine per `how-refresh-data.md`.

### Daily refresh (2026-06-14)
- **Results updated through match id 7** (HAI–SCO):
  - id 7: HAI 0–1 SCO (Group C) — confirmed Outlook India + VAVEL USA (McGinn 28')
- Verified standings in Groups view: Group C now shows Scotland 1W 3pts, Brazil/Morocco 1D 1pt each, Haiti 1L 0pts.
- Next: matches 8+ scheduled Jun 14 (AUS–TUR onwards). Continue daily routine.

### Daily refresh (2026-06-14 — stats backfill)
- **Stats added to matches 1–7:**
  - id 1 (MEX 2–0 RSA): possession 60/40, shots 16/2, cards 1/4 — sources: Yahoo Sports box score, ESPN
  - id 2 (KOR 2–1 CZE): possession 62/38, shots 16/4, cards 1/0 — sources: ESPN, Opta Analyst
  - id 3 (CAN 1–1 BIH): possession 61/39, shots 13/8, cards 1/3 — sources: ESPN, VAVEL USA
  - id 4 (USA 4–1 PAR): possession 65/35, shots 16/9, cards 1/5 — sources: ESPN, Opta Analyst
  - id 5 (QAT 1–1 SUI): possession 44/56, shots 8/11, cards 2/1 — sources: ESPN, Opta Analyst (possession adjusted to 100%)
  - id 6 (BRA 1–1 MAR): already had stats from prior session (possession 51/49, shots 12/14, cards 2/0)
  - id 7 (HAI 0–1 SCO): possession 48/52, shots 1/2, cards 1/0 — sources: ESPN, Sofascore
- Verified in preview: all seven matches now display stats in modal (confirmed via preview_eval). Possession, shots, and card counts render correctly in both EN and PT.
- No new match results to add (matches 8+ still scheduled).
- Next: continue daily routine when matches 8+ complete.

### Daily refresh runbook (2026-06-12)
- **`how-refresh-data.md` (project root) is the runbook for all updates during the tournament** — read it before touching any `data/*.json` from now on. It defines: daily `results.json` routine (scores/status, two-source rule, penalties only on ids 73–104), the one-time `thirdPlaceAssignment` fill (~Jun 27–28, slot → allowed-groups table), and the frozen files (stadiums/teams/groups/round32/assets/code — never edit).
- `how-update.md` stays as the schema reference for the (completed) mock → real migration; `how-refresh-data.md` supersedes it for day-to-day work.

### CI/CD — deploy automático para Hostinger via FTP (2026-06-14)
- **GitHub Actions** em `.github/workflows/deploy.yml`: a cada `push` em `master` (ou `workflow_dispatch` manual) envia o site pra Hostinger usando `SamKirkland/FTP-Deploy-Action@v4.3.5`.
- **Remote GitHub:** `origin` = `https://github.com/LucasKalil-Programador/world-2026-hub.git` (branch `master`). Push via credential manager do Windows (gh CLI NÃO está instalado nesta máquina).
- **Secrets necessários no repo** (Settings → Secrets and variables → Actions): `FTP_SERVER`, `FTP_USERNAME`, `FTP_PASSWORD` — vêm do hPanel da Hostinger (Files → FTP Accounts). Sem eles o workflow falha.
- **Config do workflow:** `protocol: ftps`, `port: 21`, `local-dir: ./`, `server-dir: worldcup2026/`. **Gotcha confirmado 2026-06-14:** a conta FTP da Hostinger faz login JÁ DENTRO de `public_html`, então o `server-dir` é relativo a ele — NÃO prefixar `public_html/` (causa `public_html/public_html/worldcup2026`). O destino final no disco é `public_html/worldcup2026/`. `exclude` remove do deploy: `.git*`, `.github/`, `.agents/`, `README.md`, `how-*.md`, specs `*-en.md` — só `index.html` + `assets/` + `data/` chegam ao site.
- **Sync incremental:** a action mantém `.ftp-deploy-sync-state.json` no servidor; só reenvia arquivos alterados. Não comitar esse arquivo (vive só no servidor).
- **Gotcha:** se a Hostinger não aceitar FTPS explícito, trocar `protocol` para `ftp`. Se o site ficar em subpasta, lembrar do gotcha #7 (paths relativos) — já está OK no projeto.

### Stadium SVG cleanup (2026-06-14)
- **All 16 `assets/images/stadiums/*.svg` stripped of the card chrome**: removed the `.card` background rect, top/bottom accent bars, `name`/`city`/`cap` `<text>` elements, divider `<line>`s, and the now-unused `.card`/`.name`/`.city`/`.cap`/`.div` style rules — `stadiums.js` already renders name/city/capacity as HTML, so the SVG no longer duplicates them.
- **`viewBox` cropped to just the `<g>` illustration** (~10px padding) per file, and the fixed `width="300" height="400"` attrs removed so the SVG's intrinsic aspect ratio matches its `viewBox` — needed for `.stadium-img { aspect-ratio: ...; object-fit: cover }` to crop correctly instead of showing a letterboxed 3:4 image.
- **White rectangular shapes still visible at the top of some cards (BBVA, Lumen, BMO) are intentional** — they're each stadium's press-box/scoreboard tower (`class="void"`, connected to the bowl by thin lines), not leftover text artifacts. Don't remove them.
- If adding a new stadium SVG, follow this trimmed structure: `<svg viewBox="...">` (no width/height) → `<defs><style>` with only `struct/thin/hair/concrete/stands/canopy/void/pitch/pline/acc/accs/green/ribs/louver` classes + `frit` pattern → single `<g>` with the illustration, cropped tightly. Aim for a viewBox aspect ratio near 4:3 (~1.2-1.3), to match `.stadium-img`'s `aspect-ratio: 4/3`.

### Stadium card image aspect ratio (2026-06-14)
- **`.stadium-img` (in `assets/css/style.css`) uses `aspect-ratio: 4/3`**, not `16/9`. The 16 cropped stadium SVG viewBoxes have natural ratios ~1.07-1.32 (avg ~1.25), much closer to 4:3 (1.333) than 16:9 (1.778) — `16/9` forced `object-fit: cover` to crop ~28% of the image height, slicing through rounded-rect/curved illustration paths (visible as a "diagonal" cut on NRG). `4/3` shows each illustration nearly whole across all 16 cards, including the BBVA/Lumen/BMO press-box towers and the Mercedes-Benz pinwheel.

### Data cache-busting via DATA_VERSION (2026-06-14)
- **`assets/js/app.js` `loadData()` appends `?v=${DATA_VERSION}`** to every `data/*.json` fetch (`DATA_VERSION` constant near the top of the file, currently `'2026-06-13-rev1'`). Fixes production browsers/Hostinger caching stale `results.json` after a daily refresh — `cache: 'reload'` only helps the developer's own browser, not real visitors.
- **Must be bumped on every data refresh** — added as step 4 of the daily routine in `how-refresh-data.md`. Format `YYYY-MM-DD-revN`; increment `revN` for same-day re-edits.

### How to add a UI label
1. Add the key to both `en` and `pt` dicts in `assets/js/i18n.js`.
2. Use `t("key")` at the render site — never hardcode the string.

### How to add a new localStorage preference
1. Extend the `wc2026_prefs` object shape (document the new field here).
2. Read/write only via `storage.js` `get`/`set`.

### How to add a step summary after finishing a build step
1. Mark the step `[x] ~~...~~` in `.agents/TODO.md`.
2. Append any new decisions/gotchas here (never rewrite existing entries).
3. Rewrite `project-map.md` if structure/functions changed.
4. Stop and wait for user approval before the next step.

---

## Success metrics

- Lighthouse > 90; first render < 2s; total JS < 300KB.
- Spec §18 acceptance criteria all checked (tracked in README checklist, step 11).

---

## Communication

- User communicates in English/Portuguese mix; docs in English per conventions.
- **Ask before each build step** — never chain into the next step without explicit go-ahead.