[watzon]

The Swift engine now lists detected_language in hello and puts language in final when the request has language: auto.

• Whisper: the language of WhisperKit. When there are several results, the language of the most audio (by segment time) wins.
• SenseVoice: the language comes from the first tag of the model (<|en|>), as zh, en, yue, ja, or ko. For long audio, the language of the most audio wins over the pieces.
• No language with a named language, for other models, or with empty text.
• detects_language already passes through: the engine sends the catalog entries as they are.
• New unit tests: DetectedLanguageTests.

Tested on this Mac: swift build -c release, swift test (66 tests), cargo test -p sayso-core engine_catalog, and the conformance ready level with whisper-base (German WAV), whisper-tiny (English WAV), and sensevoice-small (English WAV): 31 checks each, 0 failed.

Raw transcribe with language: auto:

whisper-base, en.wav:  {"language":"en","text":"Hello! How are you doing today? ...","type":"final"}
whisper-base, de.wav:  {"language":"de","text":"Guten Tag, wie geht es Ihnen heute, ...","type":"final"}
whisper-base, de.wav, language "de": no language field
sensevoice-small, en.wav: {"language":"en", ...}
sensevoice-small, zh.wav: {"language":"zh", ...}
sensevoice-small, ja.wav: {"language":"ja", ...}
sensevoice-small, en.wav, language "en": no language field