Files
team-tryouts/docs/translations.md
GGThedandClaude Opus 5 37f70c89e3 feat(i18n): traduire les messages flash et de validation
198 appels flash dans les sept modules de routes, plus les 23 messages de
validation de app/validators.py. Le catalogue compte desormais 631 chaines,
aucune non traduite.

validators.py utilise lazy_gettext : les champs de schema sont construits a
l'import, donc avant qu'une requete existe. Un gettext ordinaire s'y
resoudrait une seule fois, dans la langue active au demarrage.

Un bug introduit par la conversion, puis corrige
  Le convertisseur automatique ne voyait que le premier litteral d'un appel
  flash, ce qui a casse deux chaines concatenees sur plusieurs lignes dans
  users.py -- le resultat n'etait meme pas du Python valide. Ma premiere
  verification ne l'a pas vu : elle enchainait py_compile sur head, or head
  reussit toujours, donc le "OK" s'affichait quoi qu'il arrive. Les deux
  appels sont reecrits et la verification refaite correctement.

Un bug plus interessant, revele par le test de fumee
  La langue choisie ne survivait pas a la connexion. login() et logout()
  appellent tous deux session.clear() -- l'un contre la fixation de session,
  l'autre pour terminer la session -- et le choix de langue partait avec le
  reste. Concretement : quelqu'un qui lisait la page de connexion en anglais
  se retrouvait en francais des qu'il se connectait.

  La langue est une preference d'affichage, pas un etat appartenant au
  compte. Les deux endroits la reportent maintenant explicitement, a cote du
  jeton CSRF. Quatre tests couvrent le cas, dont un qui verifie que corriger
  une cle preservee n'a pas fait tomber l'autre.

Detail de nommage : le convertisseur avait genere %(value)s pour une
expression conditionnelle, ce qui n'aide pas un traducteur. Renomme en
%(player)s.

Les 14 traductions ecrites avec une apostrophe droite sont normalisees en
apostrophe typographique. Sans consequence en HTML, ou ' s'affiche
correctement -- mais les blocs <script> ne decodent pas les entites, et
autant que le catalogue soit homogene.

200 tests.

Co-Authored-By: Claude Opus 5 <[email protected]>
2026-08-07 22:12:27 -04:00

6.1 KiB
Raw Permalink Blame History

Translations

French is the primary language of the site. English remains available through the switcher in the sidebar (and on the login page, for visitors who have not signed in yet).


1. How it works

Source strings stay in English and act as gettext message ids. The French wording lives in a catalogue.

app/i18n.py                                    locale selection
app/translations/fr/LC_MESSAGES/messages.po    French catalogue (edited)
app/translations/fr/LC_MESSAGES/messages.mo    compiled (read at runtime)
app/translations/en/LC_MESSAGES/messages.po    English, msgstr == msgid
babel.cfg                                      extraction rules

This keeps the codebase in one language — the same one as its comments and docstrings — while what a member sees defaults to French.

A string with no translation falls back to English, not to a raw identifier. That is why this migration can proceed template by template without ever leaving the site half broken: an untranslated page is a page in English, not a page full of nav.dashboard.label.

Which language a visitor gets

  1. An explicit choice made through the switcher, kept in the session.
  2. Failing that, the browser's Accept-Language, restricted to fr and en.
  3. Failing that, French.

An explicit choice always wins, including over an English browser.


2. Marking a string for translation

In a template

<span>{{ _('Dashboard') }}</span>
<p>{{ _('The page you are looking for does not exist.') }}</p>

With a value inside:

{{ _('Welcome back, %(username)s!', username=user.username) }}

For a longer block:

{% trans %}This tryout has ended and can no longer be modified.{% endtrans %}

In Python

from flask_babel import gettext as _

flash(_('This account has been deactivated.'), 'danger')
flash(_('Welcome back, %(username)s!', username=user.username), 'success')

Two things that do not work

Never build a sentence by concatenation. Word order differs between languages, and the translator sees fragments with no context.

flash(_('Player ') + name + _(' has been removed.'))   # no
flash(_('%(name)s has been removed.', name=name))      # yes

Never translate at import time. A module-level _() runs before any request exists, so it resolves once, in whatever locale happened to be active — usually the default. Use lazy_gettext when the string has to sit in a constant or a class attribute:

from flask_babel import lazy_gettext as _l

ROLE_LABELS = {'coach': _l('Coach'), 'player': _l('Player')}

Extraction picks up _l because babel.cfg is invoked with -k _l.


3. Updating the catalogues

After marking new strings:

# 1. Re-extract every marked string
pybabel extract -F babel.cfg -k _l -o messages.pot --project=team-tryouts .

# 2. Merge into the existing catalogues, keeping current translations
pybabel update -i messages.pot -d app/translations

# 3. Fill in the new French entries
#    edit app/translations/fr/LC_MESSAGES/messages.po

# 4. Compile
pybabel compile -d app/translations

messages.pot is regenerable and not tracked. The .po and .mo files are tracked: deployment is a plain file mirror with no build step, so an uncompiled catalogue would mean an English-only site in production.

Entries needing attention

pybabel update marks changed strings as #, fuzzy. A fuzzy entry is ignored at runtime — the string falls back to English. Review the guessed translation, then remove the #, fuzzy line.

Adding a language

pybabel init -i messages.pot -d app/translations -l es

Then add the code to SUPPORTED_LOCALES and LOCALE_NAMES in app/i18n.py. The switcher picks it up on its own.


4. Checks

tests/test_i18n.py fails the build when:

  • a compiled .mo is missing — otherwise the site silently serves English everywhere, with nothing in the logs;
  • a catalogue still contains an untranslated entry.

That second check is what keeps the migration honest: adding {{ _('...') }} to a template without translating it turns the suite red.


5. State of the migration

Done. 631 strings, all translated, in both catalogues. The suite fails if that stops being true.

Covered: navigation and shared layout, login, the five error pages, every page template, every flash message, and every validation message.

Flash messages are done too — 198 calls across every route module — along with the validation messages of app/validators.py, which use lazy_gettext because schema fields are built at import time.

What is not covered yet

Model constantsESPORT_GAMES, GAME_POSITIONS — are data, not interface. Game names stay as they are; positions such as "Support" or "Duelist" are the terms players use in English and are deliberately left alone.

Two traps met during this migration

HTML entities are not text. &times;, used as a close-button label, was marked by the automated pass. Jinja escaped it to &amp;times;, so the button would have shown the literal text &times; instead of ×. Anything that is markup rather than prose must stay out of _().

Strings inside <script> blocks. Eight remain, in match_form.html, one_on_one.html and coach_availability.html. They work, but they are fragile: Jinja escapes & < > " ' inside a script block, and those entities are not decoded there — a translation containing an apostrophe would land in the JavaScript string as &#39;. The French wording uses typographic apostrophes (), which are untouched, so the current strings are safe. Anything added there should be passed through a data- attribute or a <script type="application/json"> block instead.

The language choice must survive session.clear(). Both login and logout wipe the session — login to prevent session fixation, logout to end it. The locale is a display preference, not state belonging to the account, and was being discarded with the rest: someone who read the login page in English was dropped straight back into French on signing in. Both call sites now carry locale across explicitly, alongside csrf_token.