Add fail-closed stable speaker inference

Fix Space dependency resolution

Pin verified model and speaker encoder revisions

Stabilize onset pace and endpoint stopping

Use stable production inference profile

Voice clone: single reference-clip mode only (drop centroid/ECAPA stable mode + speechbrain)

Update for checkpoint step_0006000

Voice cloning: extract ECAPA speaker centroid from upload (no transcript); add speechbrain

Recreate Space (zero-a10g) after region corruption; streaming included