Phase 8 — Nine languages — ships as 1.6

Status: all nine languages ship and all nine listings are written. What is left of the phase is the capture driver’s re-proof against seven new locales. Boxes in DOD.md §7; this file is the reasoning. Finished phases are in PLAN.md and are not needed to build this one.

⚠️ This file’s review gate no longer holds. It said a language ships when its native read-through passes. No reviewer was available for any of the seven, so on 2026-08-17 that gate was replaced by an audit of the half a read-through does that needs no native speaker, plus an in-app report row for the half that does — ADR-0030. Every sentence below about native review is kept as the reasoning that was in force when the drafts were written, and is amended where it is now wrong.

Prerequisite: Phases 6, 7 and 7.5 ship first, for one reason applied three times — translating a string set that is about to change means translating it twice, in nine languages. Phase 6 adds a Support screen; Phase 7’s redesign rewrites copy wherever the old wording was the thing that confused people; Phase 7.5 adds the gain signal’s health copy and an attribution screen. Seven languages of re-translation and seven native re-reads is the most expensive way to discover that ordering, so it is fixed here rather than left to judgement.

Retargeted from 1.5 to 1.6 on 2026-08-14, because Phase 7.5 adds functionality and release-please answers to commit subjects rather than to a phase file.

Decisions it leans on: ADR-0013 (English base, in-app switcher, endonym labels), ADR-0001 and ADR-0026 (what the copy may never say — the reason this phase is not a mechanical job), ADR-0009 (the listing).

What ships

Polish was the proof of concept: it proved the mechanism — AppCompatDelegate, the backport, the locales_config.xml / AppLanguage pair, and a test that holds the translation afterwards. Seven languages join it, taking the app to nine:

631 translatable resources × 7 ≈ 4 400 strings, drafted by machine against a brief and read by a native speaker before that language ships. The plural table above is written from CLDR’s rules and is checked per row when the language lands, not trusted from this file — a missing category is not an error, it silently resolves to other and renders a grammatically wrong sentence. Note the romance many category is unreachable for the integer counts this app produces (it is a large-number form), but it is declared anyway for the same reason Polish declares other.

The two facts that shape the workflow

• Resource resolution does not consult locales_config.xml. That file governs the per-app language picker on Android 13+ and AppCompat’s backport below it — it is not a delivery filter. The moment values-de/ exists in res/, every phone set to German gets those strings, reviewed or not. So a wide tranche cannot be drafted in place and released language by language: drafts live outside res/, in translations/<locale>/strings.xml, and promotion is a file move. The test suite reads paths, so it can validate a staged draft exactly as it validates a shipped one — the review gate keeps its teeth without the drafts being invisible. (androidResources locale filters could do the same job from the build file; the staging directory is preferred because it needs no build configuration and cannot be switched off by accident.)
• A language ships when its native review passes, not when its draft lands. ⚠️ Retracted by ADR-0030 — it ships when the audit passes; the staging directory below is unaffected, and did its job right up to promotion. Promotion is four edits and they go together, or the app offers a language it does not have: move the file into res/, add the <locale> line, add the AppLanguage entry, add the endonym label. AppLanguageTest already asserts the XML and the enum agree; it cannot see the other two.

Decisions

• The gate stops being named after one language. PolishTranslationTest becomes TranslationTest, parameterised over the locale table: every translatable resource present, nothing orphaned, format arguments preserved, plural categories complete per that language’s own rules rather than against a hardcoded set of four. Its existing five assertions are already the right ones — this is generalisation, not new test design, and the Polish-specific comments become the table’s rows.
• The endonym labels become translatable="false", which removes 81 duplicated entries. ADR-0013 names each language in its own language in every locale, so settings_language_polish is Polski everywhere — it is locale-invariant, which is what translatable="false" means. Today values/ and values-pl/ carry byte-identical copies; at nine languages that becomes nine labels duplicated nine times for no benefit. Marking them untranslatable collapses it to nine strings, and the existing “untranslatable resources never appear in a translated file” assertion enforces it for free.
• The translator brief is a deliverable, and it is the part that makes machine drafting safe. This app’s copy carries rules a translator cannot infer: ADR-0026 forbids missed and overdue outside Phase 4’s care reminders, ADR-0001 forbids language that infers a problem from silence, weight changes are always stated in grams, and CONTEXT.md’s vocabulary is deliberate — fluffle, watch, observation, visit — with an Avoid list per term. A translator without that produces fluent copy that quietly turns an observation into a warning, in a language no reviewer here reads. So the brief carries the vocabulary, the tone, and a per-language banned-word list (verpasst/überfällig, pominięta/zaległa, …) written with the native reviewer rather than guessed at.
• Listing text for all nine; screenshots may lag. Play falls back to the default listing’s screenshots when a language has none, but there is no fallback for discovery — an untranslated listing means nobody in that market finds the app at all. So the title, short and full descriptions are part of this phase for every language, and localised screenshots are taken for the languages with installs once there are installs to count. They wait on the locale-aware driver below — not on a --locale flag, which is what this line used to say was missing and which already exists.
• The capture driver has to be locale-aware, and that is a translation job rather than a capture one. Built in Phase 7.5 as of 2026-08-14, on en + pl, so this phase inherits the tool and owes only its re-proof as the needle table grows. The reasoning below is why it exists and stays here; the isolation half and the Polish after set go with it. Carried in from Phase 7 on 2026-08-13, where it was the single box that phase did not close. --locale already exists on screenshots.py and already switches the app — the dump comes back 14 dni, 3 dni, 7 dni — and then every scene fails at its first tap, because the scene needles in edge-to-edge.py’s table are English string literals and tap("Choose which bunny") matches nothing. The flag has never actually been exercised, which is why two phase files had been treating it as the whole job. ADR-0013 is what makes the fix small. Every user-visible string is a resource in every locale and TranslationTest keeps them level, so a needle can resolve through the resource name: parse values/strings.xml and values-<locale>/strings.xml, build the map, translate at tap time. Two wrinkles, and both argue for a lookup that falls through to the literal rather than failing — several needles are deliberate substrings of their string ("What you noticed"), and several are not resources at all (Bijou and Metacam are sample data, identical in every locale). It belongs in edge-to-edge.py, where the needles live and where the inset matrix reads them too; screenshots.py imports that table rather than copying it, so one fix serves both. Build it before the seven drafts land, not after: it is the copy-length canary, and a German compound clipping a redrawn row is cheaper to find before a native speaker reads that language than after. It is also what pays back Phase 7’s one deferred claim — that phase shipped its comparison in English, so Polish is the first language whose copy length has never been photographed against the new layouts, and it goes through this driver alongside the seven rather than as a special case.
• values-pt-rBR, not values-pt-BR. The resource qualifier uses the r prefix for a region; the BCP-47 tag in locales_config.xml and in AppLanguage does not (pt-BR). Two spellings of one locale, in two files that must agree, is exactly the shape of mistake AppLanguageTest exists to catch — extend it to compare the resource directory too.
• No RTL in this phase. Arabic and Hebrew are not a translation job, they are a layout project: mirroring, supportsRtl, and 4f’s edge-to-edge matrix doubling. Excluded deliberately so that the cost is a decision rather than a discovery.

The open question this phase owed 🟢 answered 2026-08-16

Nine locales means every future string is nine translations before the build goes green. The old test made a missing translation a red build, which is why Polish never rotted — and at nine languages that same strictness puts every future feature behind a translation round.

Answered: the boundary moves, not the rule. Neither of the two options this file weighed — strict red build, or a dated translations-pending allowlist — is what shipped. Both treat the question as how much lag to tolerate, and the better answer is where to ask: free while you work, strict before it merges.

• Completeness left TranslationTest for scripts/translation-gate.py, which CI runs on every pull request beside the schema gate. Adding an English string no longer reddens your own build; a branch that still owes a translation cannot merge.
• The point is not developer comfort, it is translating once. Under a red-build rule the copy gets translated against the draft wording and then again after review reworded it — nine times, for nothing. ADR-0013’s promise is about what ships, and main is where shipping starts.
• The allowlist is unnecessary under this shape and was dropped: it existed to make lag visible, and a gate that refuses the merge makes lag impossible instead. Nothing to date, nothing to expire, no file admitting a debt.
• The gate checks four things, and the second is the one no test could hold: resources missing from a locale (split by whether this branch introduced them, so the list is the work rather than the debt), translations gone stale because the English moved on this branch and they did not, orphans a rename left behind, and — added 2026-08-17 — an unusable comparison: no merge base means no “before”, which does not make the stale check fail but makes it disappear, so the gate now refuses to pass rather than silently dropping to the checks that still work. --report prints the same list and exits 0, for use mid-branch.
• Everything that must hold for whatever is translated — format arguments, plural categories per CLDR, orphans, untranslatable resources — stays in TranslationTest and stays green continuously.

⚠️ ./gradlew test was reporting a stale verdict, and this is what made it matter. Both translation tests read res/ off disk as plain files, which Gradle’s up-to-date check cannot see — so editing a translation and re-running test printed :app:test UP-TO-DATE and a green build having checked nothing. Fixed by declaring src/main/res as a test input in app/build.gradle.kts. Found while building this gate; it had been true since the Polish test was written.

Tests

JVM, and all of it mechanical: TranslationTest over the nine-row table (completeness, orphans, format arguments, plural categories per CLDR), AppLanguageTest extended to assert locales_config.xml, the AppLanguage enum and the resource directories all name the same nine locales. What no test can hold is tone — that is the native read-through, and it is why a language ships on a person’s word. ⚠️ Amended 2026-08-17. No test holds tone and none was built to. What ships in its place is an audit of the rules that outrank tone — blame, inference from silence, advice — and a report row that puts the tone question to the people who read the language every day. See ADR-0030.

Gate

• Every screen read end to end in each language by a native speaker, with no English left behind — ADR-0013’s original promise, now nine times. A language that has not had this does not ship. ⚠️ Replaced 2026-08-17 by ADR-0030. What a language must pass instead, and all three did: an ethics audit of its high-consequence resources against ADR-0026, ADR-0001 and observe-never-advise, plus a blame- and alarm-word scan of the whole file; a back-translation and argument-role pass over every multi-argument string and every string that substitutes a name; and a compile-and-render check on the phone, which is the only thing that sees clipping. “No English left behind” survives untouched — it is TranslationTest’s job and always was. The report row in the language picker is what carries the part none of this reaches.
• TranslationTest green across all nine locales; AppLanguageTest green across all three declarations.
• ✅ Plural categories correct per language, checked on a real count in each: 1, 2, 5 and 22 of something, in cs and uk as well as pl, where the few/many split is where a wrong table shows. PluralSelectionTest (instrumented, 2026-08-18) asks the platform for gap_days at those four counts in all nine and compares against a table hand-checked once against CLDR. It has to be on a device: TranslationTest proves a category is declared, and which item comes back is the platform’s choice, so a form filed under the wrong category is a complete file that reads wrong on the phone. Proven able to fail — expecting Czech’s 5 dne reddens it. ⚠️ Czech’s many is for fractions alone, so every whole number lands in few or other and 5 dní is right where 5 dne looks right. “Czech has a many form” is true of the file and false of every count the app will ever pass it — the one place in this set where the declared table and the rendered string disagree by design. 22 earns its place for the opposite reason: it returns to few in Polish and Ukrainian, which is what separates a correct modulo rule from a range check that stopped at 4.
• No string in any locale says missed or overdue outside Phase 4’s care reminders, and no banned-word list entry appears in its own language’s file.
• ✅ The switcher offers nine languages by endonym, and switching to each restarts into that language — driven end to end 2026-08-18, all nine tapped in turn, each landing in its own language, and every one of the nine pickers listing all nine rows. This is the one link no test reaches: AppLanguageTest proves the enum, locales_config.xml and the resource directories name the same nine, but an endonym bound to the wrong enum entry is green everywhere and ships two wrong languages.
• settings_language_* are translatable="false" and appear in no values-<locale> file.
• ✅ The Play listing carries title, short and full description in all nine — written 2026-08-17, and the English one rewritten from 1.0 scope to 1.6 first, so the seven were not translated from stale copy.
• ✅ A device set to a language Binky does not ship still falls back to English cleanly, including its plurals — proven 2026-08-18 by pinning nl, which cmd locale accepts without consulting locales_config.xml. Tabs and body copy come back from values/, and gap_days renders 3 days through English’s own one/other rule. ℹ️ Numbers and dates do not fall back and should not: the same screen reads 1,788 kg and 17 aug 2026, which is ICU formatting the Dutch locale correctly while the strings are English.
• Edge-to-edge unaffected: longest-string languages (German compounds, Ukrainian) do not clip or wrap wrongly on the narrowest supported screen — the one visual risk a translation genuinely carries. Polish is re-checked here rather than assumed: Phase 7 deferred its capture and Phase 7.5 shot it, so what this phase owes is the same check against copy the seven drafts have since changed nothing of — cheap, and the only shipped language with a before to compare against.

spotlessApply, assembleDebug, test and lint at the gate. No connectedAndroidTest was owed for the reason it usually is — there is no schema change and no media path — but one arrived anyway, from the other direction: PluralSelectionTest needs a device because CLDR’s selection rules live in the platform rather than in this repo.

Order of work

1. ✅ Done 2026-08-16. Generalise the test and the locale table first, on en + pl alone. It must be able to fail before there is anything to check. PolishTranslationTest → TranslationTest, the locale list read from locales_config.xml so a tenth language is one line of XML, and CLDR_PLURALS carrying all nine rows ready. Proven able to fail: dropping few from one Polish plural reddens the build. scripts/translation-gate.py likewise, on each of its failure modes — and re-proven against CI itself on 2026-08-17, one pushed commit per mode, which is how the stale check was found to have been silently dead there.
2. ✅ Endonyms done 2026-08-16, along with app_name’s general rule and med_editor_name_placeholder (a medicine brand name). The brief is translator-brief.md; its per-language banned lists are drafts until each native reviewer confirms their own row.
3. ✅ Done 2026-08-18. Re-prove the locale-aware capture driver — built in Phase 7.5 and proven there on pl, so what was owed here is only that its needle table still resolves once seven locales exist. Every new needle is a claim that some resource still says what the table thinks it says. The table resolves in all nine (39 of 45, zero ambiguous, six sample-data literals) and the phone walked de and uk in full plus a twelve-scene set in the other six with no failed tap. What it caught was in the driver rather than in the strings, and it is the mistake this file predicted for AppLanguage two sections up: --locale fed one spelling of a locale to two things that spell it differently, so pt-BR crashed and pt-rBR — which cmd locale accepts, storing rbr — would have shot an English app against a Portuguese needle table and called it a pass. The lesson generalises past this phase: the failure mode of a two-spelling locale is not a crash, it is a green run on the wrong language, so the guard belongs at the point the tag is taken, not where it is used. require_bcp47 and resource_qualifier, and screenshots.py inherits both.
4. Draft all seven into translations/, validated by the test in place. ✅ Done 2026-08-17 — de, es and fr on 2026-08-16, it, pt-BR, cs and uk the next day, 685/685 each — along with the staging area this step assumed and nothing had yet built: TranslationTest reads translations/<tag>/strings.xml beside values-<qualifier>/, holds a draft to the same assertions, and rejects a tag that is both staged and shipped — promotion is a move, and two files for one language drift. translations/ is declared a test input in app/build.gradle.kts for the reason src/main/res already was: a file Gradle cannot see is a file the test reports a stale verdict on. The gate prints a draft’s completeness under a heading of its own and never gates on it, because the merge rule is about what ships. The German draft’s own decisions are in the per-language record below; the one worth carrying to the next six is that §7.2’s care_every is harder than §7.3’s gender trap in at least one language. German has no preposition that takes both a bare unit and a counted gap, so the host became a label (Rhythmus: %1$s) and the gap plurals stayed in the citation form, which then decided the wording of care_due_in and care_due_overdue too. Check that chain early in each language: it is four strings deep and only visible from the code that composes them (CareLabels.kt). French confirmed it and Spanish did not, which settles the shape: tous les jours against toutes les semaines is the same missing preposition, so care_every is Rythme : %1$s there too, and the same two due-date strings had to be rewritten behind it. Two of three languages pay it; treat Spanish’s Cada as the exception. French then added a trap the brief did not have: §7.4 is not only about cases. French declines nothing, but de elides before a vowel and the app can no more elide Alice than Polish can decline Bijou — seven strings reordered or reworded for it. The lesson for the four remaining is that §7.4 is really “the app cannot touch a name the owner typed”, and each language should be asked what its own version of that is rather than checked against Polish’s. Both per-language records are the per-language record below. Italian closes both of those questions rather than adding to them. care_every is Ogni %1$s and needs no label — ogni governs a bare unit and a counted gap alike — so the split is two of four, and Spanish stops being the exception: ask a new language which side it falls on rather than expecting the label. And §7.4 comes back empty for the first time, which is a result and not a skipped check: Italian declines nothing and need not elide (foto di Alice is correct), so no string was reordered. What Italian pays instead is §7.2, in Polish’s own coin — in + the article contracts (negli, nell’) and the app cannot do it at run time, so the chart windows are pre-inflected exactly as Polish’s are. The technique generalises past the language it was written for; the trap it answers does not. Its own record is the per-language record below. Portuguese settles the care_every question and opens one nothing had looked at. A cada governs both, so the split is three of five and German’s and French’s label is the minority shape rather than the expected one; §7.2’s pre-inflection is paid in Italian’s coin (em + os → nos), §7.4 comes back empty a second time, and the divergence French found between weigh-in and weighing shows up a third. What is new is a plural trap, and it is the first thing in four drafts that no earlier record could have warned about: CLDR gives pt one: i = 0..1, so a count of zero renders the singular item — correct in French, wrong in Brazilian Portuguese, and unfixable from the plural table because the category is right and the language disagrees with it. That turns the check into a question about the code, not the copy: which counts can actually be zero? Three can, and two were recast as labels rather than sentences. Ask this of cs and uk before drafting them — Slavic one excludes zero, so they should come back clean, and confirming that is cheaper than discovering it. Its own record is the per-language record below. Czech was asked, and the answer held: one is i = 1, zero lands in other, and the genitive plural is what Czech wants there anyway, so all three of Portuguese’s counts stay ordinary sentences. What the same question turned up instead is that four plural categories are not four reachable ones — cs spends its many on fractions (1,5 dne), where pl and uk spend theirs on ordinary integers, so a Slavic row can be as unreachable as a romance one. The table above counts categories and says nothing about which of them render; that gap is now closed in the brief rather than here. And Czech adds the trap no plural table can reach — the predicate agrees with the count (jsou 3 měsíce against je 6 měsíců), so a sentence that substitutes a counted phrase has to be built on a verb that does not vary. Ask it of uk, which has the same grammar and the same substitution list. Czech is also the first draft to pay §7.3 on both halves — its past tense agrees with the addressee, so no sentence in the file puts you in the past — and the most expensive §7.4 of the six, which is where it leaves something reusable: besides Polish’s colon, a declined common noun in front of the name carries the case for it (fotky králíka %1$s), and a name in subject position needs nothing at all (Kolik let má %1$s?). Two of six languages now lose care_every’s sentence for agreement rather than for a missing preposition, which makes the split three and three: the label is neither the expected answer nor the exception. Its own record is the per-language record below. Ukrainian closes the step and answers three of the questions the six before it left open. Portuguese’s zero trap comes back clean a second time, so both Slavic rows are confirmed rather than assumed — but the pair is not the same shape underneath: cs spends many on fractions where uk spends other on them, so a draft written off the other Slavic row would have put the genitive plural in the wrong slot. Czech’s verb-agreement trap is real here too and has a cheaper answer, the zero copula — Між цими двома зважуваннями — %1$s, a dash where Czech needed a verb spelled alike in both numbers. And care_every survives, which nothing predicted: кожен has exactly Czech’s agreement problem, but Раз на %1$s governs a bare unit and a counted gap alike, so the split is four of seven keeping the sentence. That reframes the question the last five drafts were asked — not does your word for “every” govern both, but does any idiom of yours — and it is worth putting back to the Czech reviewer, since jednou za %1$s may be the same escape there. What Ukrainian adds that no earlier record could have warned about is a third unknown gender: the vet’s. §7.3 has been read as the owner and the bunny for six drafts; English hides the third behind they, and four strings needed the impersonal plural for it. It also makes the address form and §7.3 one decision — ти costs the past tense outright, where ви would buy it back, because the polite form takes the genderless plural past. Both are whole-file choices, and both are the reviewer’s. Its own record is the per-language record below.
5. ✅ All seven promoted 2026-08-17 — four edits, one commit, one language, as written; what changed is the trigger, which is now the audit rather than a native read (ADR-0030). translations/ is gone and the gate reads 688 × 8, complete. Promotion was rehearsed first — promote all seven, build, install, screenshot, revert — and that throwaway is where the value was. A staged draft is never compiled, so it was the first time any of the seven met aapt2. All seven compiled, which was not a given: French carries ~200 typographic apostrophes and a missed escape would have surfaced exactly here and nowhere earlier. It found four clipped navigation labels (destination_observations in de/es/uk, destination_care in de/it), ImpliedQuantity false positives in fr/pt-BR/uk, and stale DRAFT headers. None of the three was predictable from a draft, and none needed a native speaker to see — which is most of the argument for the audit that replaced the review, and the reason the rehearsal is worth repeating for a tenth language rather than trusting a green TranslationTest. ℹ️ The clipping fix is the label, not the string. TextAutoSize.StepBased(8sp, 11sp) in Navigation.kt; wrapping stays rejected for 1.0’s reason. With no reviewer, prefer a fix that needs no vocabulary judgement — Beobachtung, observación and спостереження are CONTEXT.md’s concept rather than a phrasing, and trading them for something shorter is exactly the call a reviewer was for. Consequence: German keeps Pflege & Medis and Italian Cure e farmaci, so six of nine keep the meds half.
6. ✅ Listing text for all nine, 2026-08-17 — store-listing.md, 27 fields, every one measured against Play’s 30 / 80 / 4000 limits. Screenshots later, driven by install data. ⚠️ The English description had to be rewritten before anything could be translated from it. It was still cut at 1.0 and described an app with no care reminders, no watch, no vet visits, no medications and no documents — 1.0 was the only build a listing had ever been written for, and nothing since had needed one, because nothing since had been uploaded. Translating that seven times is the exact failure this phase’s ordering argument exists to prevent, one surface over: the listing would have been translated once against 1.0’s copy and again the moment it caught up. So English went to 1.6 scope first, Polish was brought level, and the seven were written from there. ⚠️ Romance prose runs about a tenth longer than the English it comes from, and 4000 characters is a real ceiling. French came in at 4134 and needed thirteen trims that cut words rather than claims; it now sits at 3992 and Italian at 3993, against English’s 3628. Those two have no headroom left — a paragraph added to the English description cannot simply be translated into them. ℹ️ No locale is a translation of another, deliberately. Each carries its own keywords and the vocabulary its own draft settled on (Köttel, cagarrutas, crottes, palline, bolinhas, bobky, котяхи), because a listing that uses the clinical word for what the app calls something else reads as a different app. All nine carry ADR-0001’s closing “a record, not a diagnosis” paragraph, which is the first thing to check in any locale that gets rewritten. ⚠️ The copy must not go up before the build it describes is on the track — Play treats advertising absent features as a violation, and the tracks are still on 1.0.0 / 1.3.

What the seven drafts found — the per-language record

Moved here from DOD.md §7 on 2026-08-17, when that section was cut down to what is still open. It is kept rather than deleted because each block ends in four decisions the native read-through was meant to confirm, each with a named fallback — and under ADR-0030 there is no read-through, so those fallbacks are live: they are the pre-decided answers waiting for the first user report that touches one. The running tables the blocks build between them — which languages keep care_every as a sentence, what §7.1’s Normal costs, where §7.3 falls, whether §7.4 is empty — are what a tenth language would be drafted against.

German (de)

🟡 de drafted 2026-08-16 — 685/685, and the staging area itself now exists: TranslationTest holds a draft to every rule it holds a shipped file to (proven by breaking the German plural and the German format argument and watching both redden), translations/ is a declared test input so Gradle cannot report a stale verdict on it, and the gate prints what a draft still owes without ever gating on it.

ℹ️ German costs nothing to §7.3, and that is the surprise. The perfect tense with haben carries no gender and predicate adjectives do not inflect, so “Es ist archiviert” and “Lebt allein” are simply correct — the trap that rewrote two Polish strings does not exist here. care_every is where German pays instead: the host takes either a bare unit (Woche) or a counted gap (6 Wochen), and no German preposition governs both — “Alle Woche” is not idiomatic and jede/jeder/jedes would have to guess the unit’s gender. It is Rhythmus: %1$s, a label rather than a sentence, and it is the first thing to put in front of the reviewer.

⚠️ Four decisions the native read-through has to confirm, not just read past: destination_care = Pflege & Medis (a bottom-nav tab, and Pflege & Medikamente does not fit — the fallback is Pflege alone, losing the meds half rather than clipping it); Klo for the litter tray and Köttel for droppings (what German rabbit owners say, where Kot is the clinical word §5 rejects); Im Blick for the watch (Beobachtungszeitraum collides with Beobachtung and Wachphase reads as sleep); and three breed rows that are mappings rather than translations — Harlequin (Japaner in the standards, Harlekin in use), the UK Polish (Hermelin), and the lop family, which splits differently either side of the Atlantic. German’s banned list needed no argument: überfällig appears nowhere, care_due_overdue is %1$s nach dem Termin and care_notification_overdue is der Termin ist verstrichen — the same move Polish makes with po terminie. dose_status_skipped is Ausgelassen, agentive the way Pominięta is.

Spanish (es)

🟡 es drafted 2026-08-16 — 685/685, mechanically green, into translations/es/. All seven are in as of 2026-08-17.

ℹ️ Spanish is the first language to pay §7.3 on the bunny’s side rather than the owner’s, which is the reverse of German. The owner’s half is free — the compound perfect with haber does not agree with its subject, so has guardado and he mirado carry no gender and the file needs no rewrite. But conejo is masculine where Kaninchen is neuter, so an adjective about the animal has to pick one. The policy is the masculine generic, Spanish’s unmarked form, stated in the file’s header rather than hidden — with the genderless wording preferred wherever it is equally natural: Vive sin compañía (not solo), %1$s (en el archivo) (not archivado), and archived_banner / archived_on both recast around the noun. No parenthesised suffix appears anywhere, which is the rule that actually matters.

⚠️ Four decisions the native read-through has to confirm, not just read past: destination_care is Cuidados alone — Spanish has no established short form for “meds” the way German has Medis, and Cuidados y medicación is twenty-one characters in a bottom-nav tab, so the meds half is lost rather than clipped and 5e’s point with it; cagarrutas for droppings, the Köttel decision one language on (fallback excrementos, and bolitas was rejected outright as exactly the food-pellet ambiguity the brief warns about); seguimiento for the watch, where vigilancia carries alarm ADR-0001 forbids; and Bienvenida a Binky as a noun, because Bienvenido/a would gender the owner on the first screen they ever see.

⚠️ Four breed rows are mappings rather than translations, as German’s three were: Dutch → Holandés against Netherland Dwarf → Enano holandés, which Spanish genuinely collides where German splits them with Farbenzwerg; the lop family (belier), which splits differently either side of the Atlantic; Himalayan → Himalayo, where continental standards say Ruso; and the UK Polish → Polaco.

ℹ️ Three traps cost Spanish nothing, and it is worth knowing which. §7.1’s Normal — Polish’s five distinct forms and the file’s single biggest trap — is one word six times, because normal is invariable in gender. §7.2’s care_every is Cada %1$s, and cada governs a bare unit and a counted gap alike, so the sentence German had to give up survives. And photo_import_partial earns its plural for the second language running: no se pudo leer against no se pudieron leer, where English has nothing to vary.

French (fr)

🟡 fr drafted 2026-08-16 — 685/685, mechanically green, into translations/fr/.

⚠️ German’s care_every problem is not German’s, which is what the last draft asked to have checked. Tous les jours but toutes** les semaines, and the app cannot know which unit it is about to substitute; the unit plurals cannot carry the article either, because the editor puts them beside a number field of their own. So the host became a label — Rythme : %1$s — and that decided care_due_in and care_due_overdue in turn: À faire dans %1$s and À faire depuis %1$s, an invariable infinitive that states the timing and judges nothing. **Two of three languages pay it, which makes Spanish’s Cada the exception rather than the rule.

⚠️ §7.4 exists in French, for a reason its entry does not name. French has no cases — but de elides before a vowel, and the app can no more elide a name the owner typed than Polish can decline it. “Photo de Alice” is simply wrong. Six strings put the name first or drop the preposition: home_about_bunny, photo_description, bunny_avatar_description, watch_expired_title, watch_notification_title and document_page_description, whose second argument is a document title. A seventh, photo_gallery_empty_help, keeps its argument in the same job by changing the verb — des photos qui montrent %1$s — which is §7.6’s trap and §7.4’s in one string.

⚠️ Apostrophes are typographic (’), not \'. Correct French typography, and it also removes a class of failure the other drafts never faced: French needs some two hundred escapes, and a missed one would not surface until promotion, because a staged draft is never compiled. A later edit must not “fix” them back.

⚠️ Four decisions the native read-through has to confirm, not just read past: destination_care is Soins alone — Spanish’s outcome for Spanish’s reason, since médocs is too casual for a label seen on every screen and méds is not French; crottes for droppings (the Köttel / cagarrutas decision a third time, where excréments is the clinical word §5 rejects and crottins belong to horses), with caecotrophes beside it; suivi rapproché for the watch, where surveillance carries the alarm ADR-0001 forbids; and Sautée for dose_status_skipped, agentive the way Pominięta, Ausgelassen and Omitida are — la dose s’est sautée is not French, which is §6’s own test. The fallback is Omise.

ℹ️ French finds a §7.1 divergence English hides: weigh-in and weighing are two words there and one word here. care_type_weigh_in is Contrôle du poids rather than Pesée, or care_history_weight_help would have read “les pesées comptent comme des pesées”. Nothing predicted it — it shows only from the sentence downstream.

⚠️ Six breed rows are mappings rather than translations: the lop family (bélier), Himalayan → Russe — the continental standards’ name, where Spanish went the other way with Himalayo — the UK Polish → Hermine (German’s Hermelin), Dutch → Hollandais against Netherland Dwarf → Nain néerlandais, which French keeps apart where Spanish collides them, plus Mini Rex → Rex nain and Rhinelander → Rhénan.

ℹ️ Three traps priced, against the table the other two drafts started. §7.1’s Normal costs three forms of six, between Polish’s five and Spanish’s one. §7.3 splits the way Spanish’s does rather than German’s — the owner’s half free, because the compound past with avoir does not agree with its subject; the bunny’s half paid, because lapin is masculine, so the masculine generic is stated in the header and genderless wording used wherever it is equally natural. And photo_import_partial earns its plural for the third language running (n’a pas pu être lue / n’ont pas pu être lues) while needing a dodge neither of the others did: the added count sits on the same string and the wrong plural axis, so it is a noun — Ajout : %1$d sur %2$d — rather than a participle that would be wrong half the time.

Italian (it)

🟡 it drafted 2026-08-17 — 685/685, mechanically green, into translations/it/.

ℹ️ care_every survives as a sentence, which settles what the German draft asked. Ogni governs a bare unit and a counted gap alike (ogni settimana, ogni 6 settimane), so Italian needs neither German’s label nor French’s, the unit plurals stay in the citation form, and care_due_in / care_due_overdue are rewritten for §6’s reason rather than for grammar: Da fare tra %1$s and Da fare da %1$s. Two of four pay it, two do not — Spanish was not the exception it looked like, so the thing to ask a new language is which side it falls on, not whether it is the odd one out.

ℹ️ §7.4 costs Italian nothing, and that is the first nil result in four. Italian declines nothing and does not have to elide — foto di Alice, informazioni su Alice, a Alice are all correct as they stand, the euphonic ad being a style choice rather than a rule — and no article precedes a first name. Every name-substituting string was read with a vowel-initial name in it and none needed reordering, so home_about_bunny keeps English’s shape. “Nothing” is a legitimate answer to §7.4, not a sign the check was skipped.

⚠️ Italian’s own §7.2 trap is the preposition swallowing the article, which is Polish’s problem in a language with no cases: in + gli is negli, in + l’ is nell’, and the app cannot contract at run time. The four weight_chart_window_* are pre-inflected — negli ultimi 30 giorni, nell’ultimo anno — exactly as Polish pre-inflects for the locative, which is the first reuse of that technique outside the language it was written for. Both hosts take the fragment bare; a third host with a different preposition would break all four at once.

⚠️ Saltata stays banned, which is the mirror image of Polish’s pominięta. Same test, opposite answer: saltare is agentive when transitive (ho saltato la dose), but è saltata la dose is idiomatic for a thing that simply fell through, and a status chip has no subject to disambiguate it. dose_status_skipped is Omessa — Spanish’s Omitida and French’s fallback, reached independently. Mancata, dimenticata and scaduta appear nowhere, and backup_folder_forget is Rimuovi questa cartella rather than Dimentica, because the gate reads the file for those words rather than for their sense.

⚠️ Four decisions the native read-through has to confirm, not just read past: destination_care is Cure e farmaci — the meds half survives, where Spanish and French both dropped it, at fourteen characters, the width German accepted for Pflege & Medis; the capture driver is what settles it, and the fallback is Cure alone. Palline for droppings, where feci is the clinical word §5 rejects and cacca the baby talk §3 rejects — Spanish’s objection to bolitas does not carry over, because pelleted food is mangime or pellet in Italian, never palline. Controllo ravvicinato for the watch: sorveglianza carries the alarm ADR-0001 forbids, osservazione collides with the record type, monitoraggio is clinical. And terapia for the medication course, because English care and course both want cura and the two meet on one screen — a §7.1 collision running the other way, two English words folding into one Italian one and split by hand.

ℹ️ Three traps priced against the running table. §7.1’s Normal costs three forms of six, French’s count: Normale four times, Normali for the droppings’ size — which agrees with the pellets rather than with the measurement, the only split in the file that turns on number instead of gender — and Beve normalmente for water. §7.3 splits the Spanish/French way, the owner’s half free because avere does not agree with its subject (the trend flag is Dal %3$s %1$s ha perso %2$s), the bunny’s half paid — with two dodges neither of them had: the possessive agrees with the thing possessed (i suoi dati), and a pronoun can hang off the common noun il coniglio, whose masculine is a fact about the word rather than a guess about the animal. And photo_import_partial earns its plural for the fourth language running, needing French’s noun dodge for the added count.

⚠️ observation_not_checked is the string most worth reading in place. One resource sits under four fields of two genders — appetito and umore masculine, attività and acqua feminine — so any participle is wrong on half the screen. It is Nessun controllo, a noun phrase, which is also what keeps it a fact about the record (ADR-0001).

Brazilian Portuguese (pt-BR)

🟡 pt-BR drafted 2026-08-17 — 685/685, mechanically green, into translations/pt-BR/. Brazilian throughout, not European: celular, tela, arquivo, and the gerund (está comendo, never está a comer). pt-PT would be a second locale, not an edit to this one.

⚠️ Portuguese’s own trap is the plural category one, and it is the first one no other language’s record warns about. CLDR gives pt one: i = 0..1 — French’s rule, not Italian’s — so a count of 0 renders the singular item. In French that is correct (0 jour); in Brazilian Portuguese it is wrong, because nobody writes 0 página. No plural table can fix it: the category is right and the language disagrees with it, so the only question is whether a given count can actually be zero. The code was read for it and three can: backup_restored_overlaid renders unconditionally (an Essential-scope backup of a bunny with no avatar restores none), delete_records_sole_owned renders whenever the second delete dialog opens and DeleteConfirmation.kt opens it when either count is above zero, and document_page_count can be zero for a document whose pages were all removed. The first two are recast as labels — Imagens vindas do backup: %d — which is right at 0, 1 and 2 alike; the third is left wrong knowingly, because a label reads badly in a list row and the state is rare. Everything else is guarded at ≥ 1, checked one by one. Every future plural in this app now owes this question.

⚠️ Four decisions the native read-through has to confirm, not just read past: consulta for a vet visit, which is §5’s Avoid column overruled on purpose — the brief rejects consultation, and in Brazil consulta is simply what the appointment is called while visita means somebody coming to see you; the vocabulary’s intent survives in the copy instead, since nothing in visit_* mentions money. Backup left in English, the one word where ADR-0013’s “no English left behind” and ordinary Brazilian usage disagree (cópia de segurança reads as Portugal) — roughly thirty strings if the ruling goes the other way, and note destination_home went the other direction as Início. destination_care is Cuidados alone, since Cuidados e remédios is nineteen characters and there is no short BR form for meds — three of five now drop the meds half. And bolinhas for droppings, the Köttel / cagarrutas / crottes / palline decision a fifth time: Spanish rejected bolitas for colliding with food pellets and Portuguese has no such collision, because pelleted food here is ração. Cecotropes are cecotrofos, the tray is a caixa de areia, and the watch is acompanhamento (vigilância carries ADR-0001’s alarm, observação collides with the record type).

ℹ️ dose_status_skipped is Pulada, and it is Italian’s finding with the opposite answer. Same metaphor, §6’s same test: pulei a dose is agentive and a dose pulou is not Portuguese at all, where è saltata la dose is ordinary Italian — which is exactly why saltata stays banned there and pulada passes here. The fallback pair, if a reviewer finds it too colloquial, is Omitida / Administrada, and it moves together with dose_status_given (Dada): the two have to share a register.

ℹ️ Three traps priced, and two questions closed. §7.2’s care_every is A cada %1$s and governs a bare unit and a counted gap alike, so three of five keep the sentence — German’s and French’s label is now the minority shape, not the expected one. §7.2’s pre-inflection is paid in full, in Italian’s coin: em + os is nos, so the chart windows carry the contraction (nos últimos 30 dias, no último ano) and both hosts take them bare — third language for the technique. §7.1’s Normal costs three forms of six (Normal, Normais for the droppings’ size, Bebendo normalmente for water, whose neighbouring chips are gerunds), and §7.1’s other rows cost nothing: Nome serves both the bunny and the vet, Não se sabe serves both Unknowns — identical on purpose, and the second is also the only genderless option. Controle de peso for the weigh-in reminder against pesagem for one weighing, the French/Italian divergence found a third time. §7.3’s owner half is nearly free (the simple past does not agree), paid twice: Boas-vindas ao Binky and Você pediu este lembrete.

ℹ️ §7.4 comes back empty for the second time, and this one has a caveat worth keeping. Portuguese neither declines nor elides, so Sobre Alice and Foto de Alice stand and not one string was reordered — but the trap exists one register away: spoken Brazilian Portuguese puts an article before a first name (a Alice) and de + a contracts to da. The file stays in the written standard, which takes no article, and that is what keeps the contraction out of reach. A reviewer who prefers the spoken register cannot have it — the app cannot know a name’s gender, so da/do is unproduceable.

⚠️ The breed rows go the other way from every earlier draft. Brazil’s pet-rabbit vocabulary follows ARBA rather than the European standards, so the whole lop family stays in English (Mini Lop, Holland Lop, French Lop, English Lop, Dwarf Lop, American Fuzzy Lop) where Spanish, French and Italian each had a native name (belier, bélier, ariete) — the single row most likely to be wrong, and one for a Brazilian breeder rather than a dictionary. The same lean argues Polonês for the UK Polish, and it is Arminho anyway, joining Hermelin / Hermine / Ermellino, with the choice handed to the reviewer. Himalayan → Himalaio goes Spanish’s way against French’s and Italian’s Russe / Russo; Dutch → Holandês stays apart from Netherland Dwarf → Anão holandês; Lionhead → Cabeça de leão is genuinely current here unlike the English names around it; and Mixed / unknown → Sem raça definida, Brazil’s real idiom (SRD), which is why it earns first place rather than being sorted there.

Czech (cs)

🟡 cs drafted 2026-08-17 — 685/685, mechanically green, into translations/cs/, and the staging harness re-proven on a locale it had never seen: dropping few from one Czech plural reddens TranslationTest, so a brand-new draft directory is checked rather than merely counted.

✅ The question Portuguese left for the Slavic pair comes back clean, and it was worth asking rather than assuming. CLDR gives cs one: i = 1, so zero lands in other — 0 stran, 0 záznamů, 0 obrázků — which is the genitive plural Czech actually wants. The three counts Portuguese had to recast as labels stay ordinary sentences here, and nothing in the file is knowingly wrong at zero. uk should answer the same way.

⚠️ Czech’s own plural finding is that four categories are not four reachable ones. many is v != 0 — the fraction form (1,5 dne), spelled genitive singular — where Polish and Ukrainian spend theirs on ordinary integers. So a Slavic row can be as unreachable as a romance one, and the draft fills it with what a decimal would really take rather than mirroring other, except in whole sentences where a fractional count is nonsense. The brief’s plural table is right about the count of categories and silent about which of them can render, which is now written down in §7.5.

⚠️ The trap no plural table can see: the predicate agrees with the count. Jsou 3 měsíce against je 6 měsíců — so trend_flag_long_gap, which substitutes a gap phrase, would need two verbs. It is built on dělí, whose 3rd person singular and plural are spelled alike (Ta dvě vážení dělí %1$s). Every §7.2 substitution owes this check in an inflecting language, and English cannot show it.

⚠️ §7.3 is paid on both halves, which no earlier draft has been. German paid neither; Spanish, French, Italian and Portuguese paid the bunny’s half only. Czech’s past tense agrees with the addressee, so no sentence in this file puts “you” in the past — imperatives, the present and impersonal passives (zapsáno, nepodařilo se) carry it, and several strings were rewritten for it including two notifications. The bunny’s half is the masculine generic králík, stated in the header, with the graded chips dodging it entirely by agreeing with their field noun (chuť, nálada, aktivita are all feminine) rather than with the animal. ℹ️ A third party needs the same care: Binky has no settled gender in Czech, so the app is never the subject of a verb that has to agree with it — the feminine common noun aplikace stands in where one is unavoidable.

⚠️ §7.4 is the most expensive of the six drafts, and it produced a technique worth carrying. Beyond Polish’s colon (%1$s — informace, Smazat: %1$s?), Czech puts a declined common noun in front of the name and lets it carry the case — fotky králíka %1$s, pro králíka %1$s, z dokumentu %2$s — so the sentence keeps its shape and photo_gallery_empty_help keeps its argument in its real job. And one string needed nothing at all: Kolik let má %1$s? puts the name in subject position, where the citation form is already correct. Ask that before reordering anything.

⚠️ Four decisions the native read-through has to confirm, not just read past: destination_care is Péče a léky at eleven characters — the meds half survives, a third language keeping it against three that dropped it, because léky is the ordinary word and not an abbreviation; the fallback is Péče alone. Bobky for droppings, the Köttel / cagarrutas / crottes / palline / bolinhas decision a sixth time, where trus is the clinical word §5 rejects and pelleted food is granule, so Spanish’s collision does not arise. Sledování for the watch — and Czech is the first language where Polish’s own collision does not happen: pozorování stays the record type, sledování is the watch, two ordinary words where Polish had to qualify one. And kúra for a medication course (fallback léčba), which avoids Italian’s cura collision because care is péče.

ℹ️ Vynechána comes off the banned list, and it refines §6’s test rather than repeating it. Czech vynechat does have an intransitive life — motor vynechává — which by Italian’s rule alone would condemn it. But that reading takes a machine as its subject and dávka vynechala is not Czech, so the chip has no agentless reading. The question is the intransitive life with this noun, which is Portuguese’s version of the test, and it is now in the brief. dose_status_given is Podána; the fallback pair is Přeskočena / Podána, and the two move together.

ℹ️ Three traps priced, and one surprise. §7.1’s Normal costs three forms of six — the romance count, in a language with seven cases — because normální is one of the adjectives Czech does not inflect in the nominative singular, which is Spanish’s reason. The two that diverge are Normálně for the amount and Pije normálně for water. §7.1’s other rows split the way Polish’s do (Jméno / Jméno nebo název, Neznámé / Neví se), and Kontrolní vážení against plain vážení is the French weigh-in divergence found a fourth time. §7.2’s pre-inflection is Polish’s exact locative technique, second Slavic and fourth language overall. And photo_import_partial earns its plural for the sixth language running, but only because the file names the noun (%3$d soubor nešel přečíst) — the impersonal phrasing that came first would have collapsed all four items into one.

⚠️ Two breed rows sit one word apart and want a rabbit person, not a dictionary: Flemish Giant → Belgický obr against Belgian Hare → Belgický zajíc, because the Czech standard names the Flemish Giant after Belgium. Otherwise Czech has a national standard (ČSCH) and so translates more rows than any earlier draft: the lop family is beran (Zakrslý beran, Anglický beran, Francouzský beran) with Holland Lop left in English, because Zakrslý beran is already the Dwarf Lop and mapping both onto it would lose a breed; Polish → Hermelín, a fifth language reaching the continental name; Himalayan → Ruský, French’s and Italian’s way; Harlequin → Japonský, the standard’s name where pet shops write Harlekýn; Tan → Ohnivák, Rhinelander → Rýnský, Checkered Giant → Německý obrovitý strakáč, Lionhead → Lvíček, and Mixed / unknown → Kříženec / neznámé plemeno.

Ukrainian (uk)

🟢 uk drafted 2026-08-17 — the seventh and last, so every shipped-language draft now exists. 685/685, mechanically green, into translations/uk/, and the harness re-proven able to fail on it: dropping few from one Ukrainian plural reddens TranslationTest.

✅ The Slavic pair both come back clean on Portuguese’s zero trap, which is what the question was for. uk puts 0 in many — 0 сторінок, 0 записів, 0 зображень — the genitive plural the language wants, so the three counts Portuguese recast as labels stay sentences here too. Two checks, two clean answers, and the cost of asking was an hour against a defect that renders in a rare state and reads as fluent.

⚠️ The two Slavic rows are not the same shape, and reading one off the other would have put the genitive plural in the wrong slot. Czech spends many on fractions; Ukrainian spends other on them, as Polish does, and its many is an ordinary large-integer category. Four categories, one unreachable, different one each time — §7.5’s finding generalises, its instance does not.

✅ care_every survives as a sentence, and not for the reason five drafts have been asking about. Кожен has exactly Czech’s agreement problem (кожен тиждень, кожні 2 тижні, кожних 6 тижнів), so the label looked certain — but a different idiom governs both: Раз на %1$s, because на takes the accusative and all four units are spelled there as in the nominative the plurals already hold. So the split is four of seven keep it (es, it, pt-BR, uk) against three that lose it. ⚠️ This is worth putting to the Czech reviewer: jednou za %1$s may be the same escape there, in which case that draft’s label was avoidable. The question a new language owes is not “does your word for every govern both” but “does any idiom of yours govern both” — the brief asked the narrower one and got a label three times.

⚠️ A third unknown gender, and no earlier draft names it: the vet’s. English hides it behind they. Ветеринар has ветеринарка beside it and the app never asks which, so med_editor_amount, med_editor_amount_help, visit_weight_label and vet_delete_body all take the 3rd-person-plural impersonal (якщо тобі сказали, якщо зважували), which names nobody and inflects for nothing. §7.3 has been read as owner + bunny for six drafts; it is three parties, and the third only shows in a language that inflects the predicate.

⚠️ The address form and §7.3 are one decision here, which they are nowhere else. Ти was chosen for consistency with the other six, and it costs the past tense outright — so no sentence in the file puts “you” in the past, exactly as in Czech. But ви would buy the tense back, because the polite form takes the plural past, which carries no gender at all. That makes the register a whole-file rewrite either way, and it is the first thing in front of the reviewer rather than a preference to be noted afterwards.

ℹ️ Czech’s verb-agreement trap has a cheaper answer: the zero copula. Where the count sits inside a <plurals> the table carries the verb (watch_days_left, delete_records_shared have four deliberately different predicates). Where a counted phrase is substituted into a sentence, trend_flag_long_gap is built with a dash instead of a verb — Між цими двома зважуваннями — %1$s — and a dash agrees with nothing. Czech reached for a verb spelled alike in both numbers; removing the verb is the same fix one step earlier.

⚠️ Four decisions the native read-through has to confirm, not just read past: destination_care is Догляд і ліки at thirteen characters — the meds half survives, four of seven now keeping it; the fallback is Догляд alone. Котяхи for droppings, the Köttel / cagarrutas / crottes / palline / bolinhas / bobky decision a seventh time (кал and екскременти are §5’s clinical words, какашки §3’s baby talk, fallback кульки — pelleted food is гранули, so Spanish’s collision does not arise). Нагляд for the watch, where спостереження is the record type and стеження is surveillance of people — the second language after Czech where Polish’s collision simply does not happen. And Кличка against Імʼя або назва: Ukrainian has a separate word for an animal’s name, so §7.1’s Name row splits further apart here than in any earlier draft rather than collapsing.

ℹ️ dose_status_skipped is answered with a construction rather than a word. The whole пропущ- family fails §6’s test — дозу пропущено has an agentless reading exactly as è saltata la dose does — so what is left is the 3rd-person-plural impersonal, which in Ukrainian always implies human agents: Дали / Не давали, moving together as Portuguese said the pair must. Fallback Дано / Свідомо не дано. The same construction answers observation_not_checked (Не перевіряли), which sits under four fields of two genders and so cannot be a participle — Italian solved that one with a noun phrase, Ukrainian with a verb that has no subject.

ℹ️ Three traps priced. §7.1’s Normal costs five forms of six — Polish’s count, the most expensive of the seven, against Spanish’s one. §7.2’s pre-inflection is Polish’s locative technique in the accusative (за останні 30 днів), third Slavic and fifth language overall, plus a miniature of it in the two weight-unit options. And photo_import_partial earns its plural for the seventh language running, while photo_import_added does not — фото is an indeclinable borrowing, so all four items are the same word, which is Czech’s vážení finding arriving through a loanword instead of a declension class.

⚠️ The breed list changes script, which is a decision no earlier draft had to make. “Keep the registered name” is invisible in a Latin-script language and glaring in Cyrillic, so the obscure rows are transliterated (Беверен, Британія петіт, Тан, Тріанта) rather than left in Latin — Ukrainian breeders write them that way, and a Cyrillic list with Latin islands reads as untranslated. The lop family is баран. Himalayan → Гімалайський, not the continental standards’ російський горностаєвий, going Spanish’s and Portuguese’s way against French’s and Italian’s Russe / Russo — and in Ukrainian that is not only a naming convention. Polish → Гермелін, a sixth language reaching the continental name. Flemish Giant → Фландр, which incidentally avoids Czech’s near-collision: Бельгійський заєць stands alone. ⚠️ The row to read twice is Dutch (Голландський) against Holland Lop (Голландський баран) — one word apart, where Netherland Dwarf stays clear as Нідерландський карликовий.

When it closes

✅ Closed 2026-08-18. Nine languages ship, nine listings are written, and every gate line above is ticked. PLAN.md ticks Phase 8; DOD.md §7 is condensed to a closed record.

What the phase cost, and where it went. Nine drafts, seven of them written here, and the per-language record above is the durable half — the traps each language priced and the fallbacks pre-decided for the read-through that ADR-0030 retired. The mechanical half held throughout: TranslationTest on the nine-row table, scripts/translation-gate.py at the merge boundary, and both of them proven able to fail before they were trusted.

What the phase taught that outlives it. The two-spelling locale was named in this file as the shape of mistake to expect, and then arrived from a direction the file had not considered: not in AppLanguage where it was predicted, but in the capture driver, where --locale handed one spelling to two consumers that disagree about it. It did not present as a crash. cmd locale accepts pt-rBR and stores rbr, so the shape of the bug is a green run on the wrong language — which is why require_bcp47 refuses the tag at the point it is taken rather than converting it silently. The same reasoning found scripts/aab-locale.py still checking Polish alone, eight languages after that stopped being the list.

What is owed and deliberately unspent: the nine listings’ screenshots, tracked in DOD.md §4. They are no longer blocked on anything of ours — the driver runs every locale — but Play falls back to the default listing’s shots and the tracks are on 1.0.0 / 1.3, so the ~2 h of device time waits for the release that carries the copy. Release notes for 1.1 → 1.6 are written at upload time, nine per release from 1.6 on.