Phase 6 — Seven more languages

The phase PLAN.md promised would exist the day a third language was added. Its exact words were “Languages are not a phase of their own… Adding a third is an opt-in, and the day it happens it gets a phase file.” This is that day, and it arrives with seven at once rather than one: Czech, German, Spanish, French, Italian, Brazilian Portuguese and Ukrainian — the app in nine languages, and the Play listing in nine as well.

Sequence lives in PLAN.md; the live worklist in DOD.md; the decision this phase makes is ADR-0014, and it retracts a promise ADR-0004 made rather than adding a new rule. The words the copy has to use are CONTEXT.md’s, in nine languages now instead of two, which is what §5 of translator-brief.md is for.

What this phase is actually about is not the translating. Ninety-nine strings times seven is a long afternoon, and nothing about it is interesting. The phase exists because of the sentence standing in front of it: a language ships on a native speaker’s read-through. Seven reviewers were not available, and the honest options were to wait indefinitely or to work out which half of a read-through needed a person. §2 is that argument; the rest is what it cost to make it true.

1. What the gate demanded, in its own words

Nothing here was guessed at. The two scripts named exactly what nine shipped languages owe, and both were run before a word was written:

$ python3 scripts/notes-gate.py
0.8.0 has notes for 2 locales but the app ships 9.
Missing: cs-CZ, de-DE, es-ES, fr-FR, it-IT, pt-BR, uk.

$ python3 scripts/play-metadata.py --out … --strict
FAILED: cs-CZ is shipped in the app but has no listing
… seven times …

The toolchain was already built for nine. play-whatsnew.py’s LOCALES and play-metadata.py’s APP_TAG have carried all nine language names and Play codes since before this phase — so adding the locales to locales_config.xml turned two passing gates red without a line of script changing. That is the pipeline working as designed: the app’s locales_config.xml is the single declaration of what ships, and the listing gates read it rather than a list of their own.

2. The read-through, and why it is gone

ADR-0014 carries the decision and the alternatives. The part worth repeating here is the split, because it is what made the phase finishable: a read-through was doing two jobs, and only one of them needed a native speaker.

• Fluency — register, idiom, a compound broken the wrong way. Needs a speaker. Costs embarrassment. Cheap for any user of that language to report, and a one-line fix.
• The three rules that outrank fluency — the escape hatch stays findable and exclusive, no health claims, and never state what the app cannot know. Needs the brief and the string, not fluency. Costs, if broken: a person under a dark overlay who has been told there is another way out when there is not.

So the second half was done properly and the first half got a channel instead of a gate. translator-brief.md §2 is the checklist, written in this phase because it was a placeholder before it and an audit against a placeholder is theatre.

3. The audit, and what it found

R1 — forbidden-claim scan, all 99 × 8. Health and filter vocabulary in each target language rather than in English, matched as substrings, case-insensitively.

Eighteen candidates. All eighteen morphological false positives. Polish oczy inside Samoczynne, Czech oči inside Samočinné, Ukrainian сон inside сонця (sun) and очі inside щоночі (nightly), Polish oko inside blokady and odblokowuje. Zero rule violations.

The negative result is worth as much as the method, and it is why the method is written down: a substring scan over an inflecting language is mostly noise, so a clean read is only believable alongside the noise it produced. A scan that had returned zero candidates would have meant the pattern list was wrong.

R2 — escape-hatch audit, all nine locales. Every locale’s dim_notification_warning_body keeps the exclusive only and names the current screen; every locale’s warning title names the notification’s button; and the word it uses is the same word the notification’s own action uses (Zastavit, Stopp, Detener, Arrêter, Ferma, Parar, Зупинити, Zatrzymaj, Stop).

R3 — the invariant from R2, turned into a test. That pairing is exactly the failure no build could see: both halves fluent, naming two different things. TranslationTest now asserts that a string which quotes another by name still contains it, over two pairs — shade_notification_stop inside dim_notification_warning_title, and settings_restart inside schedule_rom_note. It was confirmed to fail when deliberately mispaired, because a test whose body silently skips is worse than no test, and both of these guard resources that all nine locales have.

R4 — the report row. “Something read wrong?” under the language picker, one tap to the existing support hand-off, with a Gloam #language subject of its own so translation reports sort apart from bugs. The resolved locale is now stamped into every report’s fact block — read off resources.configuration, not from currentAppLanguage(), which is null for “follow the phone” and would have told a translation report nothing.

4. Device readings

Taken on the Xiaomi over USB, debug build, cmd locale set-app-locales per locale and uiautomator dump rather than screenshots: the dump carries rendered text and bounds, which is what actually answers “is anything clipped”, where a screenshot needs a person to squint at it.

R5 — nine locales, three screens each, 27 dumps. Zero clipped strings. No ellipsis anywhere, and every bottom-bar label renders in full. The widest are Polish Przyciemnianie at 267 px and Ukrainian Налаштування at 269 px, against a 610 px half-screen — so the tab bar was never close, and German Einstellungen (235 px) was not the risk it looked like.

R6 — and the reading that justified the whole exercise: six of the nine languages were unreachable.

LanguageRow was a plain Row. At two chips that is correct and looks correct; at ten the sweep found only System, English, Polski and Čeština laid out, and the remaining six simply absent — not off-screen-but-present, not laid out at all. Six languages translated, declared, gated, complete, and not selectable from the switcher that exists to select them.

It is a FlowRow now, with the comment DimControls already carried for the auto-off chips — five chips do not fit one line on a narrow screen in every language, and a clipped safety control is worse than a wrapped one. The language row simply never got the same treatment, because it had never needed it.

What no test could have caught it. AppLanguageTest compares the enum to locales_config.xml and passes; TranslationTest compares resources and passes; the translation gate counts strings and passes. Not one of them knows how wide a chip is. This is the compile-and-render check earning its place in translator-brief.md §8 — a staged draft has never met a layout, and a shipped language that cannot be chosen is a shipped language nobody reads.

R7 — the fix, confirmed. All nine endonyms render across three rows, none past the screen edge, with the report row and its hint below them:

y=829   System  English  Polski  Čeština
y=997   Deutsch  Español  Français  Italiano
y=1165  Português (Brasil)  Українська
y=1303  Liest sich etwas falsch?

The phone was put back as it was found: app locale override cleared, SYSTEM_ALERT_WINDOW re-granted by uid after the reinstall revoked it, dumps deleted from /sdcard. The shade was never started, so the dim level and the schedule are untouched.

5. A latent bug inherited and not shipped

currentAppLanguage() matched the platform’s locale against each entry’s whole tag. That was sound while every entry named a bare language, and pt-BR is the first that does not: Locale.getLanguage() reports plain "pt", "pt" != "pt-BR", and the chip the user had just tapped would never have looked selected.

It matches on the language subtag of both sides now, which also covers the original case the old comment described (the platform handing back "en-GB" for an entry that says "en"). The match is sound only while no two entries share a subtag, so AppLanguageTest asserts that rather than trusting it — a second Spanish or a second Portuguese would make it ambiguous instead of merely wrong, and firstOrNull would silently pick whichever was declared first.

The reference repository next door has the same bug, unfixed, with the same nine-language enum. It was found here by reading the matcher while adding the entry that breaks it, not by a test.

6. The listing, in nine languages

The app’s languages and the store’s listing are different surfaces, and this phase moved both — because --strict leaves no third option. A shipped language with no listing section is refused before supply starts, and a shipped language with a blank section is what Play itself refused in Phase 5 R1. So nine app languages means nine listings, in full.

The tight constraint is the full description, and it is tighter than it looks. Play’s limit is 4000 characters and the Romance languages run 12–20% longer than the English:

Five of seven drafts were over the limit and needed cutting. Where the cuts came from is the part worth recording: enumerations collapsed to lists (“over other apps, over the launcher, over your browser” → “over everything: other apps, launcher, browser”), and restatements compressed. What was never cut: the touch-through paragraph, the backlight-floor limits, the secure-screen explanation, and the closing health line. Those four are what the description exists to be honest about; the rest is arrangement.

Four languages now sit within five characters of the ceiling. That is a standing cost this phase hands forward, and it is recorded at the top of store-listing.md: an English sentence added to the full description is a cut in four other languages. The counts in the headings are computed, not estimated — they are the only warning before the Console refuses a paste.

Release notes: nine locales, longest 437/500. 0.8.0’s note was written in two locales in Phase 5 and is now written in nine, for a release that has not been promoted yet — which is the correct order, since Play does not allow release notes to be edited on a live release.

7. Known gaps, deliberately

• The screenshots are English-only — three shots, one locale, and Play will show them under all nine listings. A German listing with an English screenshot is a quality gap rather than a falsehood, and scripts/screenshots.py already carries a locale tag in its filenames (1_dim-en.png), so the route to fixing it exists and costs a device run per language. Not taken here: the listing copy is what a reader searches and reads, and eight more device runs buy less than the afternoon they cost.
• No native read-through, for seven of nine. That is the decision, not the gap — but it is named here too, because a reader of this file should not have to find ADR-0014 to learn it.
• Argument roles are still unchecked. TranslationTest holds that the same format arguments appear in every locale; nothing holds that an argument still refers to the same thing. That was the read-through’s half, it has no replacement, and the honest statement of it now sits in the test’s own comment rather than in a promise.