Translation on Firefox Translations models; new logo

- Email translation moves from LibreTranslate to our own translator
  service (translator/): Mozilla's Firefox Translations models on the
  Bergamot engine, fastText language detection (176 languages). Polish,
  English, German, French, Spanish and Italian on first start (~350 MB
  instead of ~10 GB); Admin panel > Email translation installs and
  removes languages. Built and pushed as the `translator` image.
- Every message is translated paragraph by paragraph as plain text;
  names travel with their sentence and are kept only where the model
  changed them; a sentence start the model copied ("Enclosed") is tried
  again in lower case; list markers are not sent
- The Translate button shows only for installed languages
- New logo: favicon, app/maskable/iPhone icons, notification badge,
  email logo and calendar copy regenerated; BIMI logo rebuilt in SVG
  Tiny PS with its gradients; masks render in browsers
- Launch screen background is the app's dark grey

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

backend/app/api/admin.py

@@ -15,10 +15,12 @@ import re
 import uuid
 from typing import Annotated
 
-from fastapi import APIRouter, Depends, Query
+from fastapi import APIRouter, Depends, Path, Query
+from pydantic import BaseModel
 from sqlalchemy import func, select
 from sqlalchemy.ext.asyncio import AsyncSession
 
+from app import translation
 from app.api.errors import raise_api_error
 from app.config import get_settings
 from app.db import get_session
@@ -332,3 +334,61 @@ async def remove_user(user_id: uuid.UUID, admin: CurrentAdmin, db: DbSession) ->
     except MailServerError as exc:
         raise_api_error(502, "mail_server_unavailable", f"the mail server did not answer: {exc}")
     return {"status": "deleted"}
+
+
+# --- email translation languages (translator service) ------------------------------
+
+
+class InstallProgressOut(BaseModel):
+    done: int
+    total: int
+
+
+class AdminTranslationOut(BaseModel):
+    enabled: bool  # TRANSLATOR_URL is set
+    reachable: bool
+    installed: list[str] = []
+    installing: dict[str, InstallProgressOut] = {}
+    available: list[str] = []  # what Mozilla has models for (empty: offline)
+    errors: dict[str, str] = {}  # why a language's last install failed
+
+
+_LANG = Path(pattern=r"^[a-z]{2,3}(-[A-Za-z]{2,4})?$")
+
+
+@router.get("/translation", response_model=AdminTranslationOut)
+async def translation_status(admin: CurrentAdmin) -> AdminTranslationOut:
+    if not translation.enabled():
+        return AdminTranslationOut(enabled=False, reachable=False)
+    try:
+        data = await translation.status()
+    except translation.TranslationError:
+        return AdminTranslationOut(enabled=True, reachable=False)
+    return AdminTranslationOut(enabled=True, reachable=True, **data)
+
+
+@router.post(
+    "/translation/languages/{code}",
+    response_model=AdminTranslationOut,
+    dependencies=[Depends(require_csrf)],
+)
+async def install_language(admin: CurrentAdmin, code: str = _LANG) -> AdminTranslationOut:
+    """Download a language's models (to and from English) in the background."""
+    try:
+        await translation.install(code)
+    except translation.TranslationError as exc:
+        raise_api_error(502, "translator_failed", str(exc))
+    return await translation_status(admin)
+
+
+@router.delete(
+    "/translation/languages/{code}",
+    response_model=AdminTranslationOut,
+    dependencies=[Depends(require_csrf)],
+)
+async def remove_language(admin: CurrentAdmin, code: str = _LANG) -> AdminTranslationOut:
+    try:
+        await translation.remove(code)
+    except translation.TranslationError as exc:
+        raise_api_error(502, "translator_failed", str(exc))
+    return await translation_status(admin)

backend/app/api/translation.py

@@ -25,6 +25,6 @@ class TranslationInfoOut(BaseModel):
 async def translation_languages(user: CurrentUser) -> TranslationInfoOut:
     found = await translation.languages()
     return TranslationInfoOut(
-        available=bool(found),
+        available=len(found) > 1,  # English alone translates nothing
         languages=[LanguageOut(code=x.code, name=x.name) for x in found],
     )

backend/app/config.py

@@ -71,9 +71,9 @@ class Settings(BaseSettings):
     # Spelling and grammar checking in the compose window (LanguageTool server,
     # e.g. http://languagetool:8010). Empty: the check is off for everyone.
     languagetool_url: str = Field(default="", alias="LANGUAGETOOL_URL")
-    # Email translation: a LibreTranslate server (app/translation.py), e.g.
-    # http://libretranslate:5000. Empty: translation is off.
-    libretranslate_url: str = Field(default="", alias="LIBRETRANSLATE_URL")
+    # Email translation: the translator service (translator/, app/translation.py),
+    # e.g. http://translator:5000. Empty: translation is off.
+    translator_url: str = Field(default="", alias="TRANSLATOR_URL")
     # New passwords are checked against known data breaches (Have I Been Pwned's
     # k-anonymity range API: only 5 hex digits of a SHA-1 leave the server).
     password_breach_check: bool = Field(default=True, alias="PASSWORD_BREACH_CHECK")

backend/app/translation.py

@@ -1,19 +1,22 @@
-"""Email translation, on this server: LibreTranslate (Argos models) in its own
-container, reached only by the API. A message's text never leaves the server.
+"""Email translation, on this server: the translator service (translator/:
+Mozilla's Firefox Translations models and fastText language identification)
+in its own container, reached only by the API. A message's text never leaves
+the server.
 
 - A message's language is detected the first time it is shown (from its own
   text: quoted replies and signatures left out) and kept on the message, so the
   reading pane can offer "Translate" when it differs from the interface language.
 - A translation is of the message HTML the user sees (already sanitized; the
   result is sanitized again), cached per message, target language and whether
-  remote images were shown.
+  remote images were shown. It is made paragraph by paragraph, as plain text
+  (translation_prep.split_pieces), each paragraph from its own language.
+- The languages on offer are those installed on the service (admin panel).
 
-Off when ``LIBRETRANSLATE_URL`` is empty.
+Off when ``TRANSLATOR_URL`` is empty.
 """
 
 from __future__ import annotations
 
-import asyncio
 import html as html_lib
 import logging
 import re
@@ -30,10 +33,10 @@ from app.config import get_settings
 logger = logging.getLogger(__name__)
 
 DETECT_CHARS = 1500  # of the message's own text: plenty to tell the language
-MIN_CONFIDENCE = 50.0  # below this, LibreTranslate is guessing
+MIN_CONFIDENCE = 50.0  # below this, the detector is guessing
 MAX_HTML_CHARS = 150_000  # a translation request bigger than this is refused
 _TIMEOUT = httpx.Timeout(120.0, connect=5.0)  # long messages take a while on a CPU
-_LANG_CACHE_S = 600
+_LANG_CACHE_S = 60
 
 
 class TranslationError(Exception):
@@ -47,65 +50,108 @@ class Language:
 
 
 def enabled() -> bool:
-    return bool(get_settings().libretranslate_url.strip())
+    return bool(get_settings().translator_url.strip())
 
 
 def _url(path: str) -> str:
-    return get_settings().libretranslate_url.rstrip("/") + path
+    return get_settings().translator_url.rstrip("/") + path
 
 
-async def _post(path: str, payload: dict[str, object]) -> object:
+async def _call(method: str, path: str, payload: dict[str, object] | None = None) -> object:
     if not enabled():
         raise TranslationError("translation is not set up on this server")
     try:
         async with httpx.AsyncClient(timeout=_TIMEOUT) as client:
-            resp = await client.post(_url(path), json=payload)
+            resp = await client.request(method, _url(path), json=payload)
     except httpx.HTTPError as exc:
         raise TranslationError(f"the translation service is unreachable: {exc}") from exc
-    if resp.status_code != 200:
+    if resp.status_code not in (200, 202):
         detail = resp.text[:200]
         raise TranslationError(f"the translation service refused: {resp.status_code} {detail}")
     return resp.json()
 
 
+async def _post(path: str, payload: dict[str, object]) -> object:
+    return await _call("POST", path, payload)
+
+
 _languages: tuple[float, list[Language]] = (0.0, [])
 
 
 async def languages() -> list[Language]:
-    """What the service can translate between (cached for a few minutes)."""
+    """The installed languages, English among them (cached for a minute). The
+    browser names them in the interface language; ``name`` is the code."""
     global _languages
+    if not enabled():
+        return []
     stamp, cached = _languages
     if cached and time.monotonic() - stamp < _LANG_CACHE_S:
         return cached
-    if not enabled():
-        return []
     try:
         async with httpx.AsyncClient(timeout=httpx.Timeout(10.0, connect=3.0)) as client:
             resp = await client.get(_url("/languages"))
         resp.raise_for_status()
-        rows = resp.json()
-    except (httpx.HTTPError, ValueError) as exc:
+        codes = resp.json()["installed"]
+    except (httpx.HTTPError, ValueError, KeyError, TypeError) as exc:
         logger.info("translation languages unavailable: %s", exc)
         return []
-    found = [Language(code=str(r["code"]), name=str(r["name"])) for r in rows if "code" in r]
+    found = [Language(code=str(c), name=str(c)) for c in codes]
     _languages = (time.monotonic(), found)
     return found
 
 
+def forget_languages() -> None:
+    """After an install or removal: ask the service again."""
+    global _languages
+    _languages = (0.0, [])
+
+
+async def status() -> dict[str, object]:
+    """The service's own account (admin panel): installed, installing (bytes
+    done of total), available from Mozilla, and why an install failed."""
+    async with httpx.AsyncClient(timeout=httpx.Timeout(20.0, connect=3.0)) as client:
+        try:
+            resp = await client.get(_url("/languages"))
+            resp.raise_for_status()
+            data = resp.json()
+        except (httpx.HTTPError, ValueError) as exc:
+            raise TranslationError(f"the translation service is unreachable: {exc}") from exc
+    return data if isinstance(data, dict) else {}
+
+
+async def install(code: str) -> None:
+    await _call("POST", f"/languages/{code}")
+    forget_languages()
+
+
+async def remove(code: str) -> None:
+    await _call("DELETE", f"/languages/{code}")
+    forget_languages()
+
+
 _CJK = re.compile(r"[\u3040-\u30ff\u3400-\u9fff\uac00-\ud7af\uf900-\ufaff]")
 MIN_LETTERS = 25  # fewer, and the detector guesses ("Jacqueline Lian" → Esperanto)
 MAX_DETECT = 60  # pieces of one message asked about
 MIN_SHARE = 0.15  # of a message's letters, for its language to count
 
 
+async def _detect_many(texts: list[str]) -> list[str]:
+    """The language of each of ``texts`` ('' where unsure), in one request."""
+    if not texts:
+        return []
+    result = await _post("/detect", {"q": [t[:DETECT_CHARS] for t in texts]})
+    if not isinstance(result, list) or len(result) != len(texts):
+        return [""] * len(texts)
+    return [
+        str(r.get("language", ""))
+        if isinstance(r, dict) and float(r.get("confidence", 0)) >= MIN_CONFIDENCE
+        else ""
+        for r in result
+    ]
+
+
 async def _detect_one(text: str) -> str:
-    result = await _post("/detect", {"q": text[:DETECT_CHARS]})
-    if not isinstance(result, list) or not result:
-        return ""
-    best = result[0]
-    if not isinstance(best, dict) or float(best.get("confidence", 0)) < MIN_CONFIDENCE:
-        return ""
-    return str(best.get("language", ""))
+    return (await _detect_many([text]))[0]
 
 
 def _weight(text: str) -> int:
@@ -126,13 +172,8 @@ async def languages_of(parts: list[str]) -> list[str]:
         for i, part in enumerate(parts)
         if _CJK.search(part) or sum(c.isalpha() for c in part) >= MIN_LETTERS
     ][:MAX_DETECT]
-    gate = asyncio.Semaphore(4)
-
-    async def one(i: int) -> None:
-        async with gate:
-            told[i] = await _detect_one(parts[i])
-
-    await asyncio.gather(*(one(i) for i in asked))
+    for i, language in zip(asked, await _detect_many([parts[i] for i in asked]), strict=True):
+        told[i] = language
     out = list(told)
     following = ""
     for i in reversed(range(len(parts))):
@@ -178,47 +219,163 @@ def _base(code: str) -> str:
     return code.lower().split("-")[0]
 
 
-async def _translate_paragraphs(html: str, target: str) -> str:
-    """``html`` in more than one language, in ``target``: each paragraph from
-    its own language; paragraphs already in ``target`` are left as they are."""
+async def _texts(texts: list[str], source: str, target: str) -> list[str]:
+    """``texts`` translated, in batches."""
+    out: list[str] = []
+    for start in range(0, len(texts), 100):
+        batch = texts[start : start + 100]
+        result = await _post("/translate", {"q": batch, "source": source, "target": target})
+        got = result.get("translatedText") if isinstance(result, dict) else None
+        if not isinstance(got, list) or len(got) != len(batch):
+            raise TranslationError("the translation service sent no translation")
+        out += [str(x) for x in got]
+    return out
+
+
+def _kept(name: str, translated: str) -> bool:
+    """Whether ``name`` came through a translation: each of its words still
+    there, give or take an ending ("Lukas" → "Lukasie": yes; "Milan" →
+    "Mediolan": no)."""
+    return all(
+        word[: max(3, len(word) - 2)] in translated for word in name.split() if len(word) >= 2
+    )
+
+
+# A capitalised word that starts a sentence ("Enclosed is a summary…").
+_SENTENCE_START = re.compile(r"(?:^|(?<=[.!?…]\s))(?P<word>[^\W\d_][^\W\d_'’-]{2,})\b")
+
+
+def _copied_starts(core: str, text: str, names: list[str]) -> list[str]:
+    """Words that start a sentence of ``core`` and came through ``text``
+    untranslated, and are no known name: the models take a capitalised
+    sentence start for a name ("Enclosed is…" → "Enclosed jest…")."""
+    found: list[str] = []
+    for m in _SENTENCE_START.finditer(core):
+        word = m.group("word")
+        if not word[0].isupper() or word.isupper() or any(word in n for n in names):
+            continue
+        if re.search(rf"(?<!\w){re.escape(word)}(?!\w)", text):
+            found.append(word)
+    return found
+
+
+def _lowered(core: str, words: list[str]) -> str:
+    """``core`` with ``words`` lower-cased where they start a sentence."""
+    return _SENTENCE_START.sub(
+        lambda m: m.group(0).lower() if m.group("word") in words else m.group(0), core
+    )
+
+
+def _capitalised(text: str) -> str:
+    """Sentence starts capitalised again (after ``_lowered``): the first
+    letter, and one after a sentence's end — a full stop after a word of four
+    letters or more, so not after "r.", "np.", "tzw." (abbreviations)."""
+    text = re.sub(r"^(\W*)([^\W\d_])", lambda m: m.group(1) + m.group(2).upper(), text)
+    return re.sub(
+        r"((?:[^\W\d_]{4,}[.…]|[!?])\s+)([^\W\d_])",
+        lambda m: m.group(1) + m.group(2).upper(),
+        text,
+    )
+
+
+def _as_before(text: str, first: str) -> str:
+    """A word after a comma capitalised in ``text`` but not in ``first`` (the
+    first try) is lower-cased: lower-cased input makes the models capitalise
+    elsewhere ("…2026, od 13 lipca" → "…2026 r., Od 13 lipca")."""
+
+    def repl(m: re.Match[str]) -> str:
+        word = m.group(2)
+        lower = word[0].lower() + word[1:]
+        return m.group(1) + (lower if f", {lower}" in first else word)
+
+    return re.sub(r"(,\s+)([^\W\d_][\w'’-]*)", repl, text)
+
+
+async def _with_names_kept(core: str, names: list[str], source: str, target: str) -> str:
+    """``core`` translated around ``names``, which are left as they are."""
+    pattern = re.compile("(" + "|".join(re.escape(n) for n in names) + ")")
+    parts = pattern.split(core)
+    todo = [
+        (i, m.groups())
+        for i, part in enumerate(parts)
+        if i % 2 == 0 and (m := _EDGE_SPACE.match(part)) and any(c.isalnum() for c in part)
+    ]
+    done = await _texts([groups[1] for _, groups in todo], source, target)
+    for (i, (lead, _, trail)), text in zip(todo, done, strict=True):
+        parts[i] = f"{lead}{text}{trail}"
+    return "".join(parts)
+
+
+async def _translate_paragraphs(
+    html: str, target: str, sources: list[str], names: Iterable[str] = ()
+) -> str:
+    """``html`` in ``target``, as plain text paragraph by paragraph: from the
+    one language in ``sources``, or (several, or none known) each paragraph
+    from its own. Paragraphs already in ``target``, or in a language that is
+    not installed, are left as they are.
+
+    ``names`` go along with their sentence, which then translates as a whole
+    (kept apart, "I came across <Imaginary Pixels>" lost its verb: "Przyszedłem
+    Imaginary Pixels"). Only a piece where one came out changed ("Milan" →
+    "Mediolan") is translated again with it left as it is."""
+    watch = sorted({n.strip() for n in names if len(n.strip()) >= 2}, key=len, reverse=True)
     pieces, paragraphs = prep.split_pieces(html)
-    found = await languages_of([p.text for p in paragraphs])
+    if len(sources) == 1:
+        found = [sources[0]] * len(paragraphs)
+    else:
+        found = await languages_of([p.text for p in paragraphs])
+    installed = {x.code for x in await languages()}
     # Chinese and Japanese go without spaces: two pieces of one sentence
     # ("…奎琳，</span>我从…") need one between them once translated.
     followed = {i for p in paragraphs for i in p.pieces[:-1]}
     by_language: dict[str, list[int]] = {}
     for paragraph, language in zip(paragraphs, found, strict=True):
-        if language and _base(language) != _base(target):
+        if language in installed and _base(language) != _base(target):
             by_language.setdefault(language, []).extend(paragraph.pieces)
     for language, where in by_language.items():
         todo: list[tuple[int, str, str, str]] = []
         for i in where:
             m = _EDGE_SPACE.match(html_lib.unescape(pieces[i]))
             lead, core, trail = m.groups() if m else ("", "", "")
+            # A list marker stays out: the models repeat it ("1. Thursday" →
+            # "1. 1. Czwartek").
+            marker = _LIST_MARKER.match(core)
+            if marker:
+                lead, core = lead + marker.group(0), core[marker.end() :]
             if not trail and i in followed and _base(language) in _NO_SPACES:
                 trail = " "
             if any(c.isalnum() for c in core):
                 todo.append((i, lead, core, trail))
-        for start in range(0, len(todo), 100):
-            batch = todo[start : start + 100]
-            result = await _post(
-                "/translate",
-                {"q": [core for _, _, core, _ in batch], "source": language, "target": target},
+        done = await _texts([core for _, _, core, _ in todo], language, target)
+        # Sentence starts that came through untranslated: once more, lower-case.
+        again = [
+            (k, words)
+            for k, ((_, _, core, _), text) in enumerate(zip(todo, done, strict=True))
+            if (words := _copied_starts(core, text, watch))
+        ]
+        if again:
+            retried = await _texts(
+                [_lowered(todo[k][2], words) for k, words in again], language, target
             )
-            out = result.get("translatedText") if isinstance(result, dict) else None
-            if not isinstance(out, list) or len(out) != len(batch):
-                raise TranslationError("the translation service sent no translation")
-            for (i, lead, _, trail), text in zip(batch, out, strict=True):
-                pieces[i] = html_lib.escape(f"{lead}{text}{trail}", quote=False)
+            for (k, words), text in zip(again, retried, strict=True):
+                # Still there: a name after all, and the first try stands.
+                if len(_copied_starts(todo[k][2], text, watch)) < len(words):
+                    done[k] = _as_before(_capitalised(text), done[k])
+        for (i, lead, core, trail), text in zip(todo, done, strict=True):
+            lost = [n for n in watch if n in core and not _kept(n, text)]
+            if lost:
+                text = await _with_names_kept(core, lost, language, target)
+            pieces[i] = html_lib.escape(f"{lead}{text}{trail}", quote=False)
     return "".join(pieces)
 
 
 _NO_SPACES = {"zh", "ja", "th"}
+_LIST_MARKER = re.compile(r"(?:\d{1,3}[.)]|[a-zA-Z][.)]|[-•*–])\s+(?=\S)")
 _EDGE_SPACE = re.compile(r"^(\s*)(.*?)(\s*)$", re.DOTALL)
 
 
 # Web and mail addresses in the text: the models break them up ("https: / /
-# vimeo.com"), so they are marked translate="no", which LibreTranslate keeps.
+# vimeo.com"), so they are marked translate="no" and not sent.
 _ADDRESS = re.compile(r"(?:https?://|www\.)[^\s<>\"']+|[\w.+-]+@[\w-]+(?:\.[\w-]+)+", re.IGNORECASE)
 _TAG = re.compile(r"(<[^>]*>)")
 
@@ -243,36 +400,9 @@ def protect_addresses(html: str) -> str:
     return "".join(out)
 
 
-_PUNCT = re.compile(r"[\W_]+")
-MAX_NAMES = 200  # names checked for one message
-
-
-async def names_to_keep(names: Iterable[str], source: str, target: str) -> set[str]:
-    """Of ``names``, those the model changes when it translates them alone
-    ("Milan" → "Mediolan", "Imaginary Pixels" → "Imaginanal Pixels"); one
-    request for them all. If the check fails, all of them."""
-    wanted = sorted({n.strip() for n in names if len(n.strip()) >= 2})[:MAX_NAMES]
-    if not wanted:
-        return set()
-    try:
-        result = await _post(
-            "/translate", {"q": wanted, "source": source or "auto", "target": target}
-        )
-    except TranslationError:
-        return set(wanted)
-    out = result.get("translatedText") if isinstance(result, dict) else None
-    if not isinstance(out, list) or len(out) != len(wanted):
-        return set(wanted)
-
-    def same(a: str, b: str) -> bool:
-        return _PUNCT.sub("", a).casefold() == _PUNCT.sub("", b).casefold()
-
-    return {name for name, got in zip(wanted, out, strict=True) if not same(name, str(got))}
-
-
 # Bumped when what goes to the service, or what is done with the result,
 # changes: translations kept from before are made again.
-PIPELINE_VERSION = 5
+PIPELINE_VERSION = 8
 
 
 async def translate_html(
@@ -288,16 +418,6 @@ async def translate_html(
     prepared = prep.join_wrapped_lines(html)
     for language in sources:
         prepared = prep.translate_phrases(prepared, language, target)
-    single = sources[0] if len(sources) == 1 else "auto"
-    candidates = {*names, *prep.find_names(prepared, single if sources else "")}
-    terms = {*keep, *await names_to_keep(candidates, single, target)}
-    prepared = prep.keep_symbols(protect_addresses(prep.keep_terms(prepared, terms)))
-    if len(sources) > 1:
-        return prep.tidy(await _translate_paragraphs(prepared, target))
-    result = await _post(
-        "/translate",
-        {"q": prepared, "source": source or "auto", "target": target, "format": "html"},
-    )
-    if not isinstance(result, dict) or "translatedText" not in result:
-        raise TranslationError("the translation service sent no translation")
-    return prep.tidy(str(result["translatedText"]))
+    candidates = {*names, *prep.find_names(prepared, sources[0] if sources else "")}
+    prepared = prep.keep_symbols(protect_addresses(prep.keep_terms(prepared, keep)))
+    return prep.tidy(await _translate_paragraphs(prepared, target, sources, candidates))

backend/app/translation_prep.py

@@ -1,25 +1,25 @@
 """Getting a message ready for translation, and tidying what comes back.
 
-LibreTranslate translates each piece of text between tags on its own, and its
-small models stumble on things a reader would not: a sentence wrapped onto two
-lines becomes two sentences ("ask you about" / "Plan for next month"),
-"Best," becomes "Najlepsze," and a company name gets translated word by word.
-So before a message goes out:
+A message is translated as plain text, piece by piece (the text between
+tags), and the models stumble on things a reader would not: a sentence
+wrapped onto two lines becomes two sentences ("ask you about" / "Plan for next
+month"), "Best," becomes "Najlepsze," and a company name gets translated word
+by word. So before a message goes out:
 
 - lines that carry on a sentence (no end punctuation, next starts lowercase)
   are joined, in HTML (<br>, one <div> per line) and in wrapped plain text;
 - greetings and sign-offs that make a whole line are put in the target
   language from a small phrase table (English, Polish, German);
-- names are marked translate="no", which LibreTranslate keeps — the people
-  on the message, the user's own names and signature names, and capitalised
-  multi-word names mid-sentence — but only the names the model would change
-  ("Milan" → "Mediolan"): a kept name splits its sentence, and LibreTranslate
-  translates each piece alone ("Dear <kept>Lukas Remis</kept>," →
-  "Kochanie."), so a name the model leaves alone stays in the sentence;
+- names are marked translate="no" and not sent — the people on the message,
+  the user's own names and signature names, and capitalised multi-word names
+  mid-sentence — but only the names the model would change ("Milan" →
+  "Mediolan"): a kept name splits its sentence into pieces translated alone
+  ("Dear <kept>Lukas Remis</kept>," → "Kochanie."), so a name the model
+  leaves alone stays in the sentence;
 - so are emoji, file names ("A_0023C017_p1CAC.mov") and text without a letter;
-- a message in more than one language is translated paragraph by paragraph,
-  each from its own language (split_pieces); those in the target stay;
-- and afterwards, the space LibreTranslate drops after a kept word comes back.
+- the message is translated paragraph by paragraph (split_pieces), each from
+  its own language when there are several; those in the target stay;
+- and afterwards, a kept word gets back the spaces around it.
 """
 
 from __future__ import annotations
@@ -315,7 +315,7 @@ def split_pieces(html: str) -> tuple[list[str], list[Paragraph]]:
 
 # --- after translation -------------------------------------------------------------
 
-# LibreTranslate drops the space after a kept span ("Bruton</span>i twoje").
+# A kept span can lose the space beside it ("Bruton</span>i twoje").
 # Only kept spans: they still carry translate="no" here (the sanitizer that
 # runs next takes it off); other spans are the sender's and stay as they are.
 _KEPT = re.compile(r'(<span translate="no">[^<]*</span>)')

backend/tests/api/test_admin_translation.py

@@ -0,0 +1,81 @@
+"""Email translation's languages in the admin panel (app/api/admin.py), with
+the translator service replaced by a stand-in."""
+
+from __future__ import annotations
+
+import pytest
+
+from app import translation
+from tests.api.test_admin_users import World, world  # noqa: F401 — the fixture
+
+
+@pytest.fixture
+def service(monkeypatch: pytest.MonkeyPatch) -> dict[str, list[str]]:
+    state: dict[str, list[str]] = {"installed": ["en", "pl"], "calls": []}
+
+    async def status() -> dict[str, object]:
+        return {
+            "installed": list(state["installed"]),
+            "installing": {"de": {"done": 10, "total": 40}} if "de" in state["calls"] else {},
+            "available": ["en", "pl", "de", "hu"],
+            "errors": {},
+        }
+
+    async def install(code: str) -> None:
+        state["calls"].append(code)
+
+    async def remove(code: str) -> None:
+        state["installed"].remove(code)
+
+    monkeypatch.setattr(translation, "enabled", lambda: True)
+    monkeypatch.setattr(translation, "status", status)
+    monkeypatch.setattr(translation, "install", install)
+    monkeypatch.setattr(translation, "remove", remove)
+    return state
+
+
+async def test_admin_installs_and_removes_languages(
+    world: World,  # noqa: F811
+    service: dict[str, list[str]],
+) -> None:
+    async with await world.client(await world.user(admin=True)) as c:
+        got = (await c.get("/api/admin/translation")).json()
+        assert got["reachable"] and got["installed"] == ["en", "pl"]
+        assert got["available"] == ["en", "pl", "de", "hu"]
+
+        got = (await c.post("/api/admin/translation/languages/de")).json()
+        assert service["calls"] == ["de"]
+        assert got["installing"] == {"de": {"done": 10, "total": 40}}
+
+        got = (await c.delete("/api/admin/translation/languages/pl")).json()
+        assert got["installed"] == ["en"]
+
+        assert (await c.post("/api/admin/translation/languages/x;rm")).status_code == 422
+
+
+async def test_only_admins(world: World, service: dict[str, list[str]]) -> None:  # noqa: F811
+    async with await world.client(await world.user()) as c:
+        assert (await c.get("/api/admin/translation")).status_code == 403
+        assert (await c.post("/api/admin/translation/languages/de")).status_code == 403
+    assert service["calls"] == []
+
+
+async def test_unreachable_service_is_reported(
+    world: World,  # noqa: F811
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    async def down() -> dict[str, object]:
+        raise translation.TranslationError("down")
+
+    monkeypatch.setattr(translation, "enabled", lambda: True)
+    monkeypatch.setattr(translation, "status", down)
+    async with await world.client(await world.user(admin=True)) as c:
+        got = (await c.get("/api/admin/translation")).json()
+    assert got == {
+        "enabled": True,
+        "reachable": False,
+        "installed": [],
+        "installing": {},
+        "available": [],
+        "errors": {},
+    }

backend/tests/test_translation_text.py

@@ -5,7 +5,7 @@ from __future__ import annotations
 import pytest
 
 from app import translation
-from app.translation import detect, languages_of, names_to_keep, protect_addresses, translate_html
+from app.translation import detect, languages_of, protect_addresses, translate_html
 
 
 def test_addresses_in_text_are_kept_from_translation() -> None:
@@ -25,45 +25,55 @@ def test_style_blocks_are_left_alone() -> None:
     assert protect_addresses(html) == html
 
 
-async def test_only_names_the_model_changes_are_kept(monkeypatch: pytest.MonkeyPatch) -> None:
-    sent: list[object] = []
+async def test_names_go_with_their_sentence_unless_the_model_changes_them(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    sent: list[list[str]] = []
+    model = {
+        "I came across Imaginary Pixels today.": "Natknąłem się dziś na Imaginary Pixels.",
+        "Dear Lukas,": "Drogi Lukasie,",  # an ending: still the name
+        "Regards, Milan": "Pozdrawiam, Mediolan",  # changed: done again around it
+        "Regards,": "Pozdrawiam,",
+    }
 
     async def post(path: str, payload: dict[str, object]) -> object:
-        sent.append(payload["q"])
-        alone = {
-            "Lukas Remis": "Lukas Remis.",
-            "Milan": "Mediolan",
-            "Imaginary Pixels": "Imaginanal",
-        }
-        return {"translatedText": [alone[q] for q in payload["q"]]}  # type: ignore[attr-defined]
+        q = payload["q"]
+        assert isinstance(q, list)
+        sent.append(list(q))
+        return {"translatedText": [model.get(x, x) for x in q]}
+
+    async def installed() -> list[translation.Language]:
+        return [translation.Language(c, c) for c in ("en", "pl")]
 
     monkeypatch.setattr(translation, "_post", post)
-    kept = await names_to_keep(
-        ["Milan", "Lukas Remis", "Imaginary Pixels", "Milan", "x"], "en", "pl"
+    monkeypatch.setattr(translation, "languages", installed)
+    html = "<p>I came across Imaginary Pixels today.</p><p>Dear Lukas,</p><p>Regards, Milan</p>"
+    out = await translate_html(html, "en", "pl", names=["Imaginary Pixels", "Lukas", "Milan"])
+    assert out == (
+        "<p>Natknąłem się dziś na Imaginary Pixels.</p><p>Drogi Lukasie,</p>"
+        "<p>Pozdrawiam, Milan</p>"
     )
-    assert kept == {"Milan", "Imaginary Pixels"}  # "Lukas Remis" stays in its sentence
-    assert len(sent) == 1  # one request for all
-
-
-async def test_all_names_kept_when_the_check_fails(monkeypatch: pytest.MonkeyPatch) -> None:
-    async def post(path: str, payload: dict[str, object]) -> object:
-        raise translation.TranslationError("down")
-
-    monkeypatch.setattr(translation, "_post", post)
-    assert await names_to_keep(["Milan", "Lukas Remis"], "en", "pl") == {"Milan", "Lukas Remis"}
+    assert sent[1:] == [["Regards,"]]  # only the piece that lost a name went again
 
 
 def _fake_detector(monkeypatch: pytest.MonkeyPatch) -> list[str]:
     """A detector that knows Chinese signs, Polish letters, else English."""
     asked: list[str] = []
 
-    async def detect_one(text: str) -> str:
+    def one(text: str) -> str:
         asked.append(text)
         if any("\u4e00" <= c <= "\u9fff" for c in text):
             return "zh-Hans"
         return "pl" if any(c in "ąćęłńóśźż" for c in text.lower()) else "en"
 
-    monkeypatch.setattr(translation, "_detect_one", detect_one)
+    async def detect_many(texts: list[str]) -> list[str]:
+        return [one(t) for t in texts]
+
+    async def installed() -> list[translation.Language]:
+        return [translation.Language(c, c) for c in ("en", "pl", "zh-Hans")]
+
+    monkeypatch.setattr(translation, "_detect_many", detect_many)
+    monkeypatch.setattr(translation, "languages", installed)
     return asked
 
 
@@ -102,7 +112,7 @@ async def test_each_paragraph_from_its_own_language(monkeypatch: pytest.MonkeyPa
         q = payload["q"]
         if path == "/translate" and isinstance(q, list) and payload.get("source") != "auto":
             calls.append((str(payload["source"]), list(q)))
-            return {"translatedText": [f"PL({x})" for x in q]}
+            return {"translatedText": [f"PL({x.upper()})" for x in q]}
         return {"translatedText": list(q) if isinstance(q, list) else q}  # names: unchanged
 
     monkeypatch.setattr(translation, "_post", post)
@@ -117,7 +127,84 @@ async def test_each_paragraph_from_its_own_language(monkeypatch: pytest.MonkeyPa
     assert "<div>Widzę, że złożyłeś wniosek o otwarcie sklepu &amp; więcej.</div>" in out
     # Two pieces of one Chinese sentence get a space between them.
     assert "PL(我是速卖通的官方客户经理奎琳，) </span>PL(" in out
-    assert "PL(I noticed that you applied to join the platform & more.)" in out.replace(
+    assert "PL(I NOTICED THAT YOU APPLIED TO JOIN THE PLATFORM & MORE.)" in out.replace(
         "&amp;", "&"
     )
     assert '<span translate="no">+48 780 549 877</span>' in out
+
+
+async def test_one_language_needs_no_detector_and_skips_what_is_missing(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    asked = _fake_detector(monkeypatch)
+    calls: list[tuple[str, list[str]]] = []
+
+    async def post(path: str, payload: dict[str, object]) -> object:
+        q = payload["q"]
+        if path == "/translate" and isinstance(q, list) and payload.get("source") == "en":
+            calls.append(("en", list(q)))
+            return {"translatedText": [f"PL({x.upper()})" for x in q]}
+        return {"translatedText": list(q) if isinstance(q, list) else q}
+
+    monkeypatch.setattr(translation, "_post", post)
+    html = "<p>Hello there, how are you?</p><p>Short one.</p><p><b>1. Thursday 2nd July</b></p>"
+    out = await translate_html(html, "en", "pl")
+    assert out == (
+        "<p>PL(HELLO THERE, HOW ARE YOU?)</p><p>PL(SHORT ONE.)</p>"
+        "<p><b>1. PL(THURSDAY 2ND JULY)</b></p>"  # the list marker is not sent
+    )
+    assert asked == []  # the message's one language is known: no detection
+    # Hungarian is not installed: left as it is, nothing sent.
+    calls.clear()
+    assert await translate_html("<p>Jó napot kívánok</p>", "hu", "pl") == "<p>Jó napot kívánok</p>"
+    assert calls == []
+
+
+async def test_a_sentence_start_the_model_copied_is_tried_lower_case(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    model = {
+        "Enclosed is a summary. Natalia wrote it.": (
+            "Enclosed jest podsumowaniem. Natalia to napisała."
+        ),
+        "enclosed is a summary. natalia wrote it.": (
+            "załączone jest podsumowanie. Natalia to napisała."
+        ),
+        "Natalia wrote it.": "Natalia to napisała.",
+        "natalia wrote it.": "Natalia to napisała.",
+    }
+    sent: list[str] = []
+
+    async def post(path: str, payload: dict[str, object]) -> object:
+        q = payload["q"]
+        assert isinstance(q, list)
+        sent.extend(q)
+        return {"translatedText": [model[x] for x in q]}
+
+    async def installed() -> list[translation.Language]:
+        return [translation.Language(c, c) for c in ("en", "pl")]
+
+    monkeypatch.setattr(translation, "_post", post)
+    monkeypatch.setattr(translation, "languages", installed)
+    out = await translate_html("<p>Enclosed is a summary. Natalia wrote it.</p>", "en", "pl")
+    assert out == "<p>Załączone jest podsumowanie. Natalia to napisała.</p>"
+    # A name comes back the same lower-case too: the first try stands.
+    sent.clear()
+    assert await translate_html("<p>Natalia wrote it.</p>", "en", "pl") == (
+        "<p>Natalia to napisała.</p>"
+    )
+    assert sent == ["Natalia wrote it.", "natalia wrote it."]
+
+
+def test_capitalised_again_but_not_after_abbreviations() -> None:
+    assert translation._capitalised("załączony na lipiec 2026 r., od 13 lipca. dalej np. tak") == (
+        "Załączony na lipiec 2026 r., od 13 lipca. Dalej np. tak"
+    )
+
+
+def test_a_comma_word_keeps_the_first_try_case() -> None:
+    first = "Enclosed jest podsumowaniem na lipiec 2026, od 13 lipca."
+    again = "Załączony jest podsumowanie na lipiec 2026 r., Od 13 lipca, Anna wie."
+    assert translation._as_before(again, first) == (
+        "Załączony jest podsumowanie na lipiec 2026 r., od 13 lipca, Anna wie."
+    )