f1ef566af9
Six user-facing strings in SemanticTreeLegend were hardcoded English, in the
one surface that generates a document translated into all eleven locales. A
reader in Arabic opened the export menu and met "Poster" and "Packet" beside a
packet whose every page is Arabic — two honest surfaces disagreeing.
legend.export.poster menu item
legend.export.packet menu item
legend.exportButton the trigger's visible label
legend.exportSelection the trigger's title AND aria-label
legend.noDataAtThisLevel the empty state
aria.expandLegend reused for the collapsed-legend reopen button,
which had a title and NO aria-label at all, so a
screen reader announced an unlabelled button
WHY A GATE NEVER CAUGHT THESE. Both menu items only exist after a click opens
the menu, so a check that renders a leaf and scans visible text cannot see
them. The trigger's accessible name is worse: it is never visible text at all.
KEY NAMES CARRY THE FALLBACK, so they are not free. This component has no
i18n on a bare host (useI18n() is guarded) and falls back to
humanizeKey(lastSegment). I first named the keys after their ROLE — .label,
.selection, .empty — and measured the result: the export button rendered
"Label". The last segment must BE the English word, which is why
legend.columns.name -> "Name" has always read correctly. Verified all six:
Poster / Packet / Export / Export Selection / No Data At This Level / Expand
legend. Never blank, never a raw key.
TRANSLATED, NOT COINED. check-i18n-never-translated warns against inventing
terminology because a wrong term looks authoritative to a reader who cannot
check it. That warning was written about "UUPGs", a term of art. These are
ordinary print-shop words with undisputed equivalents, and the test is whether
a bilingual reader would dispute the word: not for "poster", yes for an
unreached-people-group acronym. Leaving them English was the defect.
es/pt legitimately share "Exportar"; the gate fails only on all-eleven.
Verified: check:i18n-parity PASS (11 locales complete, 0 debt),
check:i18n-never-translated PASS (my six keys added nothing to the allowlist),
BUNDLE=my-upg-100-list vite build 318 modules clean, and a static sweep of the
template now finds zero hardcoded English in text nodes or in title/aria-label.
|
||
|---|---|---|
| .. | ||
| api | ||
| colors | ||
| components | ||
| composables | ||
| constants | ||
| data | ||
| i18n | ||
| layouts | ||
| maps | ||
| packet | ||
| scripts | ||
| stores | ||
| styles | ||
| utils | ||
| index.js | ||
| ProfileLoader.vue | ||
| README.md | ||
library/ (@map) — the shared map template
The other half of the architecture (see ../plugin-designs/README.md):
every map in plugin-designs/<your-map>/ composes the reusable, parameterized pieces that live here.
Fix something here and every map benefits on the next build; break something here and every map
breaks — read components/WARNING.md before touching shared pieces.
| Folder | What lives there |
|---|---|
ProfileLoader.vue |
The root of every bundle: parses profile-config, loads the profile, provides the stores + config |
components/ |
Shared UI — map, legends, controls, popup (PeopleGroupDetail), tables, charts, hints (README) |
composables/ |
Shared logic — map engine (useMapInstance), provider switch, shadow-DOM styles, theming |
stores/ |
Per-instance Pinia stores (ui / map / data) — provided by ProfileLoader, injected by children |
api/ |
Data-source registry (sources.json) + fetch layer — the JAMstack runtime wiring |
colors/ |
Color sets maps opt into via profile-config.colorSet |
i18n/ |
The shared string catalogs (all maps speak the same languages) |
data/ |
Small shared datasets baked at build time (big served data lives in ../data/) |
constants/ |
Shared constants |
scripts/ |
Maintainer pipelines (posters, research) — not part of any bundle |
The rule of thumb (same as WARNING.md): custom code in your app-profile, reusable code here — and things only move from the first to the second once they've earned it.