fix(#687): triage the I3 literal tail — 21 sites, 7 files on the ratchet

[Weegy]

Follow-on to #747, which made the I3 category measurable. This spends that measurement: translate goes from 40 → 19, and the per-file ratchet grows from 1 → 7 files.

Two scanner rules, learned from the triage

Neither was guessed — both come from looking at what the first pass actually returned.

1. Version / phase / build stamps are diagnostic, not prose.

site	literal
store/page.tsx:229	omadia · v1
store/[id]/page.tsx:464	omadia · v1 · Slice 1.1
store/builder/page.tsx:181	B.0 Draft-Store
Workspace.tsx:2070	Phase B.5 (Workspace-UI)

They sit in product footers and read like text, but localising them changes an identifier an operator quotes back in a bug report. One of them is telling: store/builder/page.tsx already renders {t('footer.phaseLabel')} and then the hardcoded value right beside it — the label was translated, the value was not.

Checked in the dangerous direction: the diagnostic bucket contains exactly those four and nothing legitimate.

Whether internal roadmap vocabulary (Slice 1.1, Phase B.5) belongs in a shipped footer at all is a product question. Flagged, not answered.

2. Every placeholder= is a judgement call, not only the one-word ones.

#747 forced single-token placeholders to placeholder and left multi-word ones in translate. The triage showed that was the wrong cut — of six hits, five were example values that must not be translated (Release sign-off, Release approver, id-1, id-2, …, email | uri | date-time | uuid) and one was a real instruction (Reason (optional)). Nothing in the syntax separates an example value from an instruction, so multi-word placeholders now land in review — the bucket that means "a human decides", which is exactly true here.

The translations

27 keys across six files, all of which reach zero actionable hits and join SWEPT_COMPONENTS.

The judgement calls worth naming, since that is what review exists to force:

• block / boost (kg-priorities action select) → German. GLOSSARY.md puts descriptions of what is happening in German and keeps product nouns English; these are actions an operator picks, so blockieren / hervorheben. The value attributes stay block / boost — the wire contract is untouched.
• Vault, Bucket → stay English (product noun, established loanword) but are now catalogue keys with explicit brand / loanword entries in i18n-identical-allowlist.json, rather than implicit exemptions.
• Status → identical in both locales, allowlisted as loanword.
• Resolution → Behebung. Not Auflösung, which in German software reads as screen resolution.

On that allowlist point: these five would have passed silently either way. i18n-validate exempts short capitalised nouns from the identical-value gate by design — a gap its own comment names as the follow-up. Listing them anyway puts the decision where a reader looks instead of relying on a rule that happens not to fire.

Verification

• npm run i18n:check — OK, 3760 keys
• tsc --noEmit clean; eslint clean on every changed file
• npm test — 742/742 across 88 files
• Mutation check: restoring one literal (label={t('prefix')} → label="Prefix") turns exactly the new per-file ratchet red; reverting returns 15/15.

What is left, and why

19 translate hits remain. Most are sentences split across JSX elements — the scanner sees rendered markdown (, Stream live ·, · last activity, s tool-time, User · Tenant — where the fix is an ICU message with placeholders, not a 1:1 key. Restructuring those changes markup rather than just extracting strings, so it belongs in its own pass rather than being smuggled into a translation PR.

The rest are taglines (an Agentic OS) and format names (OpenAPI 3 (JSON)) that need a brand decision before a key.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[Weegy]

Pushed a second commit that widens the scan — and it found three shipped German strings that the first version structurally could not see.

Three blind spots, each a construct the parser never visited

1. Template literals with substitutions. stringValueOf handled plain string literals only, so this was invisible:

title={`Triage-Klassifizierer: ${routing.classifierModel} → ${routing.bucket} → ${routing.model}`}

A hardcoded German tooltip in chat/page.tsx. Worth noting it also slips past the repo's own self-check in web-ui/CLAUDE.md — git diff | grep '[äöüÄÖÜß]' finds nothing, because "Triage-Klassifizierer" has no umlaut.

2. Strings in plain objects rendered later.

const verdict = { simple: { label: 'einfach' }, complex: { label: 'komplex' }, … };
…
<span>{v.label}</span>

Neither JSX text nor a JSX prop. Two more German labels, same file. Restricted to the same closed prop list, so the only false positive it can produce is a label on an API payload — 24 hits repo-wide, 10 of them actionable, no explosion.

3. My own over-exemption in the first commit. It skipped all of chat/page.tsx because web-ui/CLAUDE.md exempts MOCK_KG_WALK. But the docs exempt a symbol, not a file — the fixture is a const at line 79, and the German tooltip sat at 1251. Exemptions are now scoped to the declaration. An exemption must be no wider than the thing it excuses, and that is the same over-broad-blessing failure this whole issue is about.

Two corrections that fell out

• Rendering a template as {} per substitution collided with the placeholder rule (\{[^}]*\}), which would have marked every interpolated string exempt and silently re-hidden the class the change exists to reveal. Classification now runs on the literal parts; only the display form carries {}.
• Text with umlauts or ß is translate outright — German in a component is the severest class in the repo rules, so it outranks every exemption below. Precision over recall on purpose: it cannot false-positive, and it cannot catch umlaut-free German, which still lands in review. That is the honest limit; no cheap test separates the two languages, and guessing would put real English labels in the wrong bucket.

The ratchet earned its keep

Widening the scan immediately turned ListView.tsx red — a title={User: ${user}} that was invisible when the file was declared swept in #747. That is the point of making the guard content-based rather than a list of strings: widening the question re-asks it of every swept file, instead of leaving old files blessed under the old, narrower question.

Verification

i18n:check OK (3766 keys) · tsc --noEmit clean · eslint clean · 742/742 vitest. Mutation check: restoring label: 'einfach' makes the object-label detector report it again; reverting returns 0.