Feat/french language support

[AliAhmed-004]

Add French language support (TTS + STT)

Adds French (fr) as a conversation-only, LTR language, following the same
config seams as the Indian-language suite (#75). French is exposed globally —
deployments opt in via SETTINGS_LANGUAGES / region_features.supported_languages;
no francophone country is hardcoded, so lesson-plan localization uses the existing
region-neutral fallback (French is spoken across France, Canada, Belgium, and
much of West/Central Africa).

What's included

Text + LLM

• supported-languages.js — fr (Français), LTR, not RTL
• language-prompts.js — natural French Rumi persona
• gamma-languages.config.js — French lesson-plan / presentation output
• language-config.js — French coaching reflective-questions
• system-messages.js — French for all 11 canned messages
• utils/language-detector.js — French /language override commands (Latin-script,
  so explicit commands + Soniox trust rather than a Unicode rule)

Voice — TTS

• tts-voices.js + constants.js — fr routes to ElevenLabs eleven_v3; defaults
  to the shared multilingual voice, override with ELEVENLABS_VOICE_ID_FR; OpenAI
  tts-1 remains the automatic fallback

Voice — STT

• audio.service.js — fr added to the Soniox multi-language auto-detect hints
  (consolidated the duplicated inline list into SONIOX_AUTODETECT_HINTS)
• language-detector.service.js — fr added to SONIOX_TRUSTED_LANGUAGES

DB + config

• 017_add_french_language.sql — no-op-safe preferred_language CHECK update
• .env.template — ELEVENLABS_VOICE_ID_FR + French picker examples

Bug fix (general, not French-only)

While wiring STT I found that when Soniox returns language only per-token (v3/v5,
top-level language is null), the code fell back to a rule-based text detector
that can't read Latin-script languages — it mislabeled French as en/es, which
would make the bot reply to a French voice note in the wrong language. Fixed at the
root: new AudioService._languageFromTokens() derives the language from Soniox's
own per-token IDs before the text fallback. This is strictly more reliable for
every language, not just French.

Testing

• tests/language/french-language.test.js — TTS/config surfaces (persona, gamma,
  system messages, voice routing, override commands, region-neutral localization)
• tests/language/french-stt.test.js — Soniox auto-detect hints, per-token
  language derivation (incl. region-strip + null fallback), and fr trust
• Full suite: 1491/1491 passing, 144 suites, all conformance guards green

Verified live end-to-end

• TTS — generated French audio via eleven_v3; transcription + back-translation
  by an independent model confirmed correct, intelligible French
• STT — fed the audio back through the real Soniox path: fr identified on
  3/3 clips, transcripts accurate
• Full voice loop — seed text → TTS → audio → STT → transcript → TTS → audio:
  French identified 3/3, transcripts round-trip to the seed (punctuation aside)

Not included (intentionally deferred, same as the Indian suite)

• Reading-assessment WCPM/LCPM norms (still en/ur only)
• A hardcoded francophone region/currency (left to per-deployment config)

[AliAhmed-004]

These are the audios generated by the flow:

fr_1_greeting.mp3
fr_2_lesson_ready.mp3
fr_3_encouragement.mp3