Document the nine languages, and what Beta means on them

The translations shipped today and the docs still said "no languages but
English". They also need to say the harder thing, which is that eight of
the nine have never been read by anybody who speaks them.

- README gains the language list, with the Beta caveat in the same line
  rather than a footnote.
- FEATURES.md gains an Interface language section: the list, why it is a
  separate setting from the date locale, and the two design properties
  that follow — a missing entry renders English, and plurals are asked of
  Intl.PluralRules rather than assumed, which is why Russian carries
  three forms and Japanese one.
- ROADMAP.md no longer lists translations as "not yet". What replaces it
  is the half that is genuinely not done: a translation anybody has
  checked. RTL is split out as its own entry, because holding Arabic,
  Hebrew and Persian back is a layout decision and not a queue position.
- KNOWN-ISSUES.md gains two entries. The unread catalogues, which is the
  one thing on that page that cannot be closed by testing. And the
  coverage number that read 100% while two hundred strings rendered
  English in every language — recorded as a general lesson rather than an
  i18n one, since a coverage number measures what it can see and the rest
  is exactly what nobody is checking.

KNOWN-ISSUES.md

@@ -27,6 +27,10 @@ works the same way — and dropped where 0.15 was the whole subject. Support for
 0.15 was removed on 2026-08-26; the last release that runs on it is tagged
 [`stalwart-0.15-support`](https://github.com/Coffey-Labs/ihasmail/releases/tag/stalwart-0.15-support).
 
+- **Eight of the nine interface languages have never been read by anybody who speaks them.** They were produced by AI against standard dictionaries on 2026-08-31 — German, Spanish, French, Dutch, Portuguese (Brazil), Russian, Ukrainian, Simplified Chinese and Japanese — and every one is marked **Beta** in the picker, with that stated in Settings beside a link for reporting anything that reads wrongly. This is the entry that matters most on this page, because it is the one thing here that cannot be closed by testing: a translation can be complete, consistent, pass every check, and still read like a machine wrote it, and nobody on this project can tell which. What *is* verified is the machinery around them. A missing key renders its English source, so a bad line can simply be deleted; a stale key — one whose English no longer exists — is caught by `npm run i18n:check` rather than sitting in the file looking correct and never being looked up. Plurals are asked of `Intl.PluralRules` rather than assumed, which is why Russian and Ukrainian carry three forms and Japanese and Chinese carry one; supplying `one` for Japanese would have been filling in a distinction the language does not draw. Confirmed live on the deployed instance (2026-08-31) against a 6,289-message mailbox: role folders localise and the ~20 custom folders keep the names their owner gave them, dates and the calendar follow the language, and 6,289 renders as *6289 листувань* — the genitive plural a number ending in nine takes, which is the first time the plural machinery ran on anything but a hand-picked value.
+
+- **`npm run i18n:coverage` reported 100% while about two hundred strings rendered English in every language.** It reads JSX text, and it was not wrong about what it measured — none of them were JSX text. They were `toast.error(...)` arguments, `confirmDialog({ title, confirmLabel })` props, `title=` and `aria-label=` attributes, and template literals: every one built from an expression a codemod cannot read. The calendar's own view switcher was the clearest case, spelling its labels `v[0].toUpperCase() + v.slice(1)` — correct English, untranslatable anywhere else, and galling because **Day**, **Week**, **Month** and **Agenda** were already in all nine catalogues and the buttons simply never asked for them. Reported from production, where the switcher stayed English in a Japanese interface. All of them are now wrapped, and `npm run i18n:check` grew a second half (`scripts/i18n-literals.mjs`) that accepts a string wrapped where it is written *or* present as a catalogue key — the constant-table convention, where `SECTIONS` holds `label: "About"` and the render site calls `t(s.label)` — and refuses one that is neither, because that is a string no catalogue can translate however many languages ship. It found twenty more than a hand sweep had. Worth recording as a general lesson rather than an i18n one: a coverage number measures the thing it can see, and the strings it cannot see are exactly the ones nobody is checking.
+
 - **A compressing hop in front of Stalwart truncated every blob download, and nothing said so.** Node decompresses a gzip response before the code ever sees the body, but leaves the `content-length` header describing the *compressed* bytes. The blob proxy copied that header onto the longer body it forwarded, so the browser stopped reading exactly that many bytes in and called the download complete. Reported on [#76](https://github.com/Coffey-Labs/ihasmail/issues/76) against a Coolify deployment, where Traefik's compress middleware only engages above 1 KiB: filter rules one and two were fine and the third pushed the script past the threshold, after which it came back cut off mid-rule — 384 bytes of a 1.3 KB script. The size threshold is what made it look like a race. This is the *second* cause behind that issue, and the first fix did not touch it: a truncated script is neither unknown nor empty, so the "refuse to save from a baseline we could not read" guard never fired — the script parsed, just with rules missing, and the next save wrote the short version back over the real one. Every blob download shared the fault, not just Sieve: message source, vCards, signature HTML, attachments being forwarded, and the `settings.json` sync. Settings degraded honestly by luck rather than design — a truncated file fails `JSON.parse`, which is caught and leaves the local cache in charge — so it stopped syncing between devices instead of being overwritten. The proxy now asks upstream for `identity` and, for a hop that compresses anyway, forwards no length at all rather than one describing different bytes. The image proxy is unaffected: it uses `node:http` directly, sends no `accept-encoding`, and never decompresses. The save path no longer trusts the transport either: a script is now checked for completeness against the shape the generator emits — every `# rule:` comment parses, every enabled rule has an `if` and a closed body below it, every block ends with a blank line — and saving refuses on anything short, as does the rule editor, which reports the script as unreadable rather than showing the rules that happened to parse. The check is structural rather than a re-serialize-and-compare, so a script written by an older version with a different serializer is still editable; refusing over a changed byte would be the worse bug. It catches a cut at every offset except the end of a complete rule block, which is a legitimately shorter script and indistinguishable from one in the bytes alone — that residual is what the proxy fix covers.
 
 - **Delete all spam destroys, and does not pass through Deleted Items** — this is the point of the feature and the thing worth checking on a real server, since a folder that empties into another folder has solved nothing. `Email/set destroy`, walked a page at a time so it survives `maxObjectsInSet` the way emptying Deleted Items already had to. **Confirmed live on 0.16.19 (2026-08-26)**: Junk Mail emptied and Deleted Items stayed empty afterwards. There is no undo, which is why all three entry points share one dialog that says so. Only Deleted Items and Junk Mail can be emptied this way, enforced in the store rather than only hidden in the menus.