[BUG] Short Whisper transcripts are dropped by text-length confidence heuristic #675

Bug description

Live Whisper transcription can silently drop short but valid utterances such as acknowledgements, confirmations, and brief replies.

Current behavior

WhisperEngine::transcribe_audio_with_confidence derives a confidence-like value from the UTF-8 byte length of the decoded segment. The live transcription worker then rejects non-empty results below a 0.3 threshold.

Several common English replies fall below that threshold:

• OK receives approximately 0.12.
• Yes receives approximately 0.13.
• No receives approximately 0.12.
• Sure receives approximately 0.14.
• Got it receives approximately 0.16.
• Thanks receives approximately 0.16.

All of these can be skipped even though Whisper returned non-empty text. The derived value is not an acoustic or token probability.

Expected behavior

A non-empty transcript returned after VAD and Whisper's own no-speech handling should be emitted. The current confidence value may remain available as metadata, but it should not decide whether the transcript is discarded.

Proposed fix

• Base transcript emission on non-empty content rather than the uncalibrated confidence value.
• Keep empty and whitespace-only results filtered out.
• Add regression coverage for several short English utterances and a whitespace-only result.

Affected revision

Reproduced on devtest at 0bedd8b.

Checklist

• Searched for a similar issue
• Identified the root cause in the live transcription worker
• Added a failing regression test before the fix