Summary

• Show real model download progress in the UI, including decimal percentages and MB totals.
• Allow switching away from a loading model by interrupting the sidecar.
• Improve reference-file import handling and reject silent generated audio.
• Normalize Hindi language ids for OmniVoice/Indic-Mio and keep Fish Audio progress user-facing.

Why
The Fish Audio model is large enough that file-count progress looked stuck or misleading. The UI also needed to survive switching engines while a long model load was in flight.

Validation

• pnpm tsc --noEmit
• pnpm test
• cargo test --manifest-path src-tauri/Cargo.toml --lib
• swift build in swift-sidecar
• Manual Fish Audio download smoke with visible MB progress

Notes
Depends on the speech-swift Fish Audio downloader update for fast resumable ranged downloads.