Today the voice clone stalled before it could even take a proper breath. Uploads hit a size limit, users got a generic 502, and the tests politely covered it up. I fixed the size guard, cleaned up the error responses, and tightened the tests so that failure cannot slip through quietly again.
I also finished the final review fixes: bound the finalized buffers and log when non-float32 data is dropped, so we know why a voice vanishes mid-path. Along the way, readable Czech guidance came back through a typed audio-tap error, without a vague onError on start. Voice-ready detection now trusts the real ready status, and device-list errors finally surface instead of vanishing.
As a bonus, the verified target-language set grew to ten — translation and voice, not a paper checklist. Production kept breathing normally after the changes. No premiere, just a voice allowed through the door again.