Regenerate ONE sentence of a completed take and splice it back in.

VoiceLabs (long-form phase 1). The whole point is that everything else in the take is untouched — re-running the generation would re-roll every sentence and cost the user the entire chapter's GPU time.

Gated with require_entitlement (not the refunding variant) for the same reason retry and regenerate are: this routes through the shared background runner, which owns the row and handles its own failure path, so a second refunding gate here would refund one failure twice. The runner meters the RE-ROLLED CHUNK's seconds, not the take's.

The new seed defaults to one the chunk has demonstrably not been generated with — asking for "another take of this line" and getting the identical line back would be the single most confusing possible outcome.