Custom HF Inference Endpoint handler for Meta Seamless-M4T v2 speech-to-text — the endpoint counterpart of src/transcribers/seamless.py (generate(generate_speech=False)).

The target language is set per-endpoint via SEAMLESS_TGT_LANG (default som; gaz=Oromo, amh=Amharic), since the request carries only audio. Model via SEAMLESS_MODEL (default facebook/seamless-m4t-v2-large); long audio is chunked into SEAMLESS_CHUNK_SEC (default 30 s) windows. Seamless does NOT support Afar/Tigrinya.

Protocol: POST {"inputs": <base64 audio>} → {"text": ...}. Deploy on a medium GPU, min replica = 0. See docs/asr-endpoints-deploy.md.