l10n audit findings — what past locale passes actually found

Evidence from locale passes already done. Nothing here is a rule. The rules are in CLAUDE.md (short) and docs/l10n-workflow.md (the runbook); this file exists so those two do not have to carry the reasoning behind them.

Read it when you are about to audit a locale and want to know where the defects usually are, or when a measurement is telling you something surprising and you want to know whether someone has already met that surprise.

Defect rates, and what predicts them

Locale	Defects	Of	Shape
is	235	1052 pre-existing (22%)	wrong gender, wrong case, malformed compounds, garbled words, foreign stems, wrong senses
ca	128	2052 (6.2%)	no dominant term; biggest class 35 keys, eleven classes of 2–11
cs	113	2052 (5.5%)	terminology drift and internal inconsistency; zero garbled words, zero foreign stems, zero agreement failures, zero bad arrays
lb	77	—	60 of them ONE orthographic rule (Eifeler Regel)
sk	57	2052 (2.8%)	37 of the 57 were a single term

The rate tracks how healthy the locale is upstream, not how long it went un-audited. cs predates all of this tooling and still came in at 5.5% against is's 22%.

Counting competing renderings per English term is the highest-yield first move — about 70 of cs's 113 and 37 of sk's 57, with no knowledge of the language required. Do that count before reading values. npm run l10n:termdrift -- <loc>.

A corrections count of 0 means "unverified", not "clean". sk had a measured register, a detector, a reviewed cognate set and an empty corrections, and the audit still found 57 real defects including two semantic reversals invisible to every gate. A 0 does not imply a quarter of the file is waiting either. Read it as "nobody looked".

Mechanical morphology checks almost never pay — 4 of 239 on is, ~0 of 113 on cs. The one exception is a deterministic sandhi rule; see below.

Collisions the user can see on one screen

ca's two worst defects were byte-identical renderings sitting next to each other: Settings/Configuration both rendering Configuració as sibling tab labels in three dialogs, and Logs/Registers both rendering Registres as the two tabs of one tab bar.

Hunt this first in any locale whose language collapses two of the app's nouns: grep the tabs: arrays and the paired empty states. A collision the user can see is a defect no amount of "both words are correct Catalan" excuses.

Still open, same class: es and pt render Logs and Registers identically (Registros / Registos), and those two blocks appear together in RegistersSideBar.

Sandhi: ask the question, expect a different answer each time

lb is the counter-example to "budget for terminology, not grammar". Its terminology was healthy and 60 of its 77 corrections were one orthographic rule — the Eifeler Regel, where word-final -n deletes before any consonant but n d t z h. Obligatory, fires several times per sentence, invisible to every gate, broken in both directions. It found 44 further violations in the half written during the pass, so split at HEAD and re-run it on your own work. Method: docs/l10n-workflow.md §8.11.

The same shape applies to French elision/liaison, Irish initial mutation, Italian lo/il.

But mk is where the question was asked and came out no. Macedonian obligatorily doubles a definite direct object with an accusative clitic (Избриши ГО објектот), which looks exactly like the Eifeler Regel and is not the same shape: the trigger is semantic (definiteness) rather than orthographic, and the clitic's position depends on clause type. The check scored 0 of 5. What paid on mk instead was capitalisation — 118 occurrences over 105 values, the dominant class of that pass.

Measuring a bundle's own practice

Aggregate by WORD CLASS, not only per lemma. On lb, believing otherwise cost a whole round: a lemma occurring in only one environment carries no information, so Lueden (always with -n) and Deele (always without) both read as "consistent", and 26 values were excused on a "16:0, no counter-example" count that was really one copy-pasted phrase. By word class the family does both, 118 to 108. A uniform lemma is evidence of one decision copied, not of a rule.

A capitalisation ratio conflates two populations, and the confound invents work. Measure mid-sentence casing over prose only (English key ≥6 words and not Title Case), and separately ask whether short labels mirror the key's Title Case. On sq the naive scan reported 25–35% against a family rate near zero — a tidy ~110-value defect class that does not exist: prose is 0-of-177, and every hit was a Title-Cased heading correctly following its source. If a defect class is large, uniform and concentrated in short values, you are measuring the source, not the translation.

The conditioned measurement can still come back real, and on mk it did. Restricted to prose keys the bundle capitalised 41 against 146 where the sibling frontends run 1:273 and core 2:~480, and the same rule was broken again in headings (65:507 against core's 7:122). What tells that apart from sq's phantom: it does not evaporate under the restriction, and the bundle is internally inconsistent — објект is 0:26 lowercase while Датотека is 20:0 capitalised. So there was no convention to follow, unlike sr, which capitalised its six first-class concepts one-sidedly across the whole bundle.

Three ways a casing measurement misleads, all met on mk:

• A stem alternation written without the i flag matches only the lowercase form, so every capitalised occurrence — the entire thing being measured — is invisible, and every term reports a clean 0-up.
• (?<=.) excludes the value's first word but not a later sentence's, so … пребарување. Објектите се … scores as a mid-sentence capital. Three first-cut hits were that, and each would have been "corrected" into an error.
• An opening parenthesis, a leading emoji and deliberate all-caps each license a capital the same way a sentence start does — 14 of mk's 132 raw hits.

Review the word list before applying any casing fix.

Detectors

An ACRONYM can be a homograph of a register marker, and case is the only thing separating them. ВИ is Macedonian for AI (вештачка интелигенција) and this bundle uses it (ВИ-агент, ВИ-функции), while ви is the formal dative clitic. A detector that folds case scores the app's AI vocabulary as deferential address. detectors/mk.js therefore consumes the all-caps form in fold() before lowercasing. The hyphen goes in the left guard only: Macedonian attaches acronyms to the following noun with a hyphen, so the acronym must still match with a hyphen after it.

That is the lb dir/Dir situation from the other end — lb had to preserve case throughout, mk needs it for one token and can spend it up front.

A marker guard must know where the target puts a MORPHEME boundary, not just what is a letter. (?<!\p{L}) is wrong for Albanian, which attaches the definite ending to acronyms after a hyphen (UUID-je, Token-i, PHP-ja), so it matched an inflectional ending as the 2sg copula. detectors/sq.js uses (?<![\p{L}-]) — but must NOT guard the apostrophe, since t'ju is formal and s'ke informal. Catalan needed the interpunct inside the token class: same question, opposite answer.

Thin core coverage is more dangerous than none. rm and mt had zero catalogues, so the scan threw and the step could not run by accident. lb had one catalogue with 72 values, so it succeeded — and would have reported a register verdict computed from 0 markers, since those 72 values contain no address form. Measure the marker count, not the catalogue count. bs (55 values) is the same trap.

A functional field described in prose is not a functional field. The mk pass wrote the whole pluralBoundary rationale into pluralNote and never set pluralBoundary itself. Third instance of this after pluralOrder and spellAllow, but the first to fail loudly, because runtime-check treats an unacknowledged boundary disagreement as fatal. The two earlier ones failed in the safe direction and went unnoticed for passes. Set the field, then re-run the check that reads it.

Button conventions

sq's LENGTH-GRADED convention replicated on mk, which is what makes it a shape worth checking rather than one locale's quirk — and the crossover landed in the same place, ~40 characters. Core mk runs 128:1 for the 2sg imperative at ≤14 characters and 6:12 the other way at 80+.

The Select/Choose/Enter prompt override sits on top of the gradient in both locales, and it is lexically bounded rather than "any prompt": Search goes the other way and is not close (core mk 61:1 for 2sg). A dropdown you pick from and a field you type into address the user; a toolbar button is something you press.

Before "fixing" an outlier, check core AND the call site

Core cs overturned four candidate corrections, one of which was the only value in its family that actually matched core. On sk the two together overturned ten — more than any single class that pass corrected. Core sk itself collapses Refresh/Restore onto Obnoviť, ships Prvé/Posledné/Predchádzajúce and bare Hľadať verbatim, and uses Riešiteľ for Handler because the field holds a person. Full list in docs/l10n-workflow.md §6.9.

The reports have their own blind spots. l10n:spell was splitting every Catalan l·l word in half and reporting the halves as misspellings — a locale's orthography can defeat the tooling silently. Read a report for tooling failure before you read it for defects.