feat(stt): serve /tasks/audio/transcriptions from the 51-language ASR model

[PatrickCmd]

Summary

Replaces the Modal/RunPod dispatch behind POST /tasks/audio/transcriptions with Sunbird's faster-whisper ASR deployment (Sunbird/faster-whisper-51-african-languages, a Whisper large-v3 fine-tune), taking the endpoint from 10 to 51 African languages.

The 51-code list was taken from language_map.json in the converted repo rather than the prose model card. All ten previously supported languages (ach teo eng lug lgg nyn swa kin xog myx) are still served — no language regresses.

The deployment is OpenAI-compatible and maps Sunbird ISO 639-3 codes to Whisper's language-token slots server-side, so codes are forwarded verbatim. Because the fine-tune reuses those slots, auto-detection is meaningless and language stays required.

⚠️ Breaking change

The endpoint no longer accepts platform, adapter, whisper, recognise_speakers, org, or gcs_blob_name — these now return 422. Speaker diarization and the organization workflow remain available on the deprecated /tasks/stt* routes.

New contract:

POST /tasks/audio/transcriptions
  audio      : file            required
  language   : <one of 51>     required
  timestamps : bool = false    -> populates `segments`

⚠️ Deploy requirement

ASR_BASE_URL must be set in Cloud Run before this ships. It has no working default — the fallback is a literal <ENDPOINT_ID> placeholder, so an unconfigured deploy fails at request time, not at boot. It must include the trailing /v1. Auth reuses the existing RUNPOD_API_KEY.

ASR_MODEL_NAME (sunbird-asr-51) and ASR_TIMEOUT_SECONDS (600, generous for cold starts) are optional.

Changes

File	Change
app/services/asr_service.py	New. ASRService — httpx client for the deployment; maps HTTP/timeout/connection/parse failures to ExternalServiceError.
app/core/config.py	Adds asr_base_url, asr_model_name, asr_timeout_seconds; promotes RUNPOD_API_KEY to a real Settings field.
app/schemas/stt.py	Adds ASRLanguage (51 codes), ASR_LANGUAGE_NAMES, TranscriptionSegment. STTTranscript gains segments + duration_seconds. SttbLanguage untouched for the legacy routes.
app/routers/audio.py	Endpoint rewritten to the contract above.
app/deps.py	TranscriptionServiceDep → ASRServiceDep.
app/services/transcription_service.py	Deleted — existed only to dispatch across the removed platforms.
.env.example, app/docs.py, docs/tutorial.md, frontend/src/pages/Tutorial.tsx	Full 51-language list, timestamps section, breaking-change notice. Endpoint matrix grew from 44 to 68 rows.

Two behavioural notes: audio_url is a gs:// URI (matching what the legacy RunPod path returned, not a signed HTTPS URL), and the GCS upload is best-effort — if it fails, transcription still succeeds with audio_url/audio_transcription_id as null.

Test plan

• pytest app/tests/ — 1189 passed. 6 failures in test_config.py / test_inference_service.py reproduce on a clean git stash: pre-existing, caused by a local .env leaking into Settings.
• New test_asr_service.py (12 tests, mocked httpx transport: request shape, verbose_json/segments, language fallback, HTTP/timeout/connection/malformed-body error mapping, health check).
• test_audio_transcriptions.py rewritten — new contract, removed-param rejection, segments on/off, 51-code enum, upstream-error → 502.
• make lint-check — black/isort clean; only pre-existing node_modules/flatted E722 remains.
• npm run build (incl. tsc) passes.
• Not run: npm run lint — pre-existing breakage, ESLint 10.5.0 installed against an .eslintrc.* config (v9+ requires eslint.config.js). Unrelated to this PR; no JS dependencies added.
• Not verified: live transcription against the deployed endpoint — needs ASR_BASE_URL pointed at a running worker.

npm audit reports 8 pre-existing vulnerabilities (3 high, vite/launch-editor); the fix is a breaking vite@8 major, left for a separate PR.

🤖 Generated with Claude Code