Run extraction off the event loop; harden startup and ingest

Record the serving tradeoff in the deployed source

Return a readable message when audio cannot be decoded

Fix residue: align estimate to mixture before subtracting

Deploy fully self-trained Vanta (separator + speaker encoder)

Deploy from-scratch trained separator (val SI-SDR +7.93 dB)

Switch to SepFormer backend for +9.7 dB SI-SDR

Initial deploy: Vanta TSE backend (FastAPI + trained checkpoint)