refactor(i18n): fold the supplementary table into the dictionary literal

[PathGao]

i18n.ts merged a supplementary table into the exported translations with Object.assign at module load, so the source literal and the exported value were different objects.

One Object.assign, not several: a nested top-level loop merging interactiveLabelTranslations (26 locales, sections settings/toc/common). Added wholesale by #196 — appending a table rather than editing 26 places in a 6,700-line literal. No locale-override machinery, no merge-conflict residue, nothing load-bearing. Folded.

What the merge was actually doing: 259 keys added, 23 restated identically (dead rows), 20 locales got their entire toc section from the loop's else-branch, and 4 keys overwritten with a different value.

The four keys that depended on the overwrite

locale	key	value in the literal (lost)	value from the table (won)
zh-TW	settings.toolbarOnBar	列	工具列
ko	settings.toolbarPlacement	툴바 위치	도구 모음 위치
ko	settings.toolbarOnBar	표시줄	막대
ko	settings.resetToolbar	툴바 초기화	도구 모음 초기화

All four winners preserved. The three losing Korean values no longer appear in the source at all — which is the point: two rows claimed the same key and only one could ever win, and now only one is written down.

The proof is the deliverable, not the fold

The fold is mechanical; showing nothing moved is not. A snapshot of the fully-merged dictionary, flattened to 6,210 sorted dotted keys, taken before and after: byte-identical, sha256 cd46b9b1…. Independently re-taken on the reviewer's side: same 6,210 keys, same digest.

The literal is regenerated by a printer that was first proved to reproduce the pre-fold literal byte-for-byte, so only the merge's own effects appear in the semantic diff.

That forced two commits. The literal had drifted into three styles at once — 165 tab-indented lines, 26 double-quoted values, 25 trailing commas — which the printer would have silently normalised inside the interesting diff. Commit 1 is that normalisation alone, snapshot verified unchanged; commit 2 is the fold.

The two tests were not deleted, and the premise for deleting them was wrong

I asked for two i18nCoverage tests to be removed as workarounds for the merge. They are not workarounds.

dictionaries in i18nCoverage.test.ts is built from the imported runtime translations — already post-merge. It is not static analysis of the literal, so no test ever had to reach past the merge; there was no barrier to work around.

• "Simplified Chinese directly translates the window organization labels" checks menu.* / toast.* / home.*, sections the supplementary table never carried. Measured against the pre-merge dictionary: 0 of its 11 keys missing. No relationship to the merge at all.
• "interactive button labels are directly translated for every supported language" does read keys that lived in the table, but its purpose — documented above it — is per-locale enforcement: the suite's general rules only require a key in English, and per-locale completeness is reported, never asserted. It is the only test that fails when a specific locale silently falls back to English, and after the fold it is the only thing standing between those 26 locales and a deleted translation.

Both still pass unchanged.

Guardrail

A RULES row, the translation dictionary is one literal, allowed: [].

The marker pins unqualified Object.assign( rather than a pattern naming translations, because the loop reached the dictionary through a local alias (const currentSection = translations[language][section]) that a translations-anchored regex cannot see — which is precisely how the mutation would come back. src contains no other Object.assign, so the blanket ban costs nothing. A second alternative catches the direct translations[l][s] = … spelling of the else-branch.

✖ single implementation: the translation dictionary is one literal
  AssertionError: second implementation found — `translations` was a 6,700-line
  literal *plus* a second table…

Numbers

npm test 903 → 904 · vitest 357 · check 0 errors.

The node count rises by one — the new rule — rather than dropping by the two tests I expected to delete.

i18n.ts 7,326 → 7,114 lines, shorter despite the literal growing 299 lines (259 keys plus 40 lines of new toc braces): removing the 510-line table, its two now-unused types and the loop more than paid for it. Duplicate-key scan across the folded literal: 0.

🤖 Generated with Claude Code