[codex] Improve voice formatting quality

[IchenDEV]

Summary

• Strengthen ASR transcript sanitization for explicit no-speech artifacts, special tokens, repeated transcripts, and weak-audio template tails without dropping real short utterances.
• Add shared ASR quality prompt rules for hallucination tails, spoken formatting controls, technical term casing, and context fact boundaries.
• Reduce WhisperKit temperature fallback count to keep short or weak recordings more deterministic.

Validation

• swift test — 507 tests passed, 3 environment-gated streaming integration tests skipped.
• swift build — passed.