The detection backend was reworked so inference is no longer tied to a single provider. Detection jobs can now run across providers rather than depending on one.

A dedicated Sidekiq worker now handles detect inference, and there is a polling fallback for cases where a Replicate prediction is still processing when the job checks on it.

The point of the refactor is reliability under load. Spreading inference across providers and adding the fallback keeps detections completing when any one path is slow or busy.