Count with a plural rather than an apologetic "(s)"

[vmillet-dev]

Closes #185.

⚠️ This one carries a parser I wrote rather than a library, and that is the part worth your eyes. Everything else is mechanical.

The dependency was installed, and it cannot run here

You steered me to @jsverse/transloco-messageformat, and the reasoning held: 18 of the 28 strings carry more than one (s), four of them three independent counts in a single sentence with three separate agreements. A key per form would have meant eight variants of those alone.

It compiles each message into a function with new Function. tauri.conf.json is script-src 'self'. The application booted onto a CSP error banner, seeded no sample note, and 19 of 21 e2e spec files failed.

⚠️ The 1339 unit tests passed the whole time. Vitest runs under jsdom, where there is no policy to violate — this is a class of failure only the assembled application can show, and I would have opened this pull request without it had I not run the suite.

What replaces it

plural-transpiler.ts, ~60 lines, reading exactly the ICU syntax already written — so the 28 converted strings are kept as they are. No dependency, no eval, no CSP change.

Two choices that bound the risk of hand-rolling this:

• It knows plural, exact =N, the categories Intl.PluralRules answers, and #. It does not know select, selectordinal, or a plural inside a plural. Nothing here needs them, and a parser that guesses at a grammar it does not implement is worse than one that refuses it.
• Plain interpolation stays {{name}} — the convention of the other 75 strings — so DefaultTranspiler keeps handling it and this code never touches it. That includes the sibling-key resolution {{app}} depends on.

Eleven tests, the two that matter most being:

expect(fr('{count, plural, one {# note} other {# notes}}', { count: 0 })).toBe('0 note');
expect(en('{count, plural, one {# note} other {# notes}}', { count: 0 })).toBe('0 notes');
expect(new Intl.PluralRules('fr').select(0)).toBe('one');

Zero taking the singular in French is the thing no doubled key of mine got right, and the third line keeps the reason written down rather than in a commit message.

⚠️ One debt, deliberate and commented: TranslocoService asks for the transpiler in its own constructor, so taking it by injection would be a cycle. The service is resolved on the first render instead.

The rest

• 28 strings to ICU in both locales, 0 (s) left, 506 keys in parity. Two guards on the shipped files: no (s) may come back, and no brace may appear that is neither an interpolation nor a plural.
• shortcuts.unavailable is the only one that needed the call site changed: it agrees three times over and had no count at all, only a list.
• The trash off-by-one. ⚠️ Not the cause the ticket gave. ClockService ticks every 30 s, so the now a card renders against can be behind an instant Rust has just stamped; the gap reads as 30 days and change, and rounding up made the change a whole day. The ticket proposed floor, which says 29 for the same note — the same contradiction reversed. It rounds now, and the test reproduces the exact string from the screenshot.
• 34 tooltips, each from the same key as its aria-label. ⚠️ The ticket said 84; that counted every occurrence of the attribute rather than the buttons a mouse user cannot read.
• One empty state, not five. ⚠️ The ticket listed five as if each had an exit to offer. Only noResults is a dead end — clearFilters() was sitting there unoffered — and it is now conditioned on hasUserFilters, never isFiltering, which is true inside a folder.

🤖 Generated with Claude Code