Summary

OpenAISTTTranscriptionSession._stream_audio base64-encoded the raw audio
buffer and sent it in input_audio_buffer.append. StreamedAudioInput
explicitly accepts np.int16 | np.float32, but the session configures the
realtime transcription input as int16 PCM ({"type": "audio/pcm", "rate": 24000}). So for float32 input we shipped raw 32-bit IEEE-754 float bytes
(4 bytes/sample) to an endpoint expecting PCM16 (2 bytes/sample), producing
garbage/failed transcriptions. The batch/tracing path (_audio_to_base64)
already converted float32 → int16 correctly, so the streaming and tracing
paths disagreed.

This extracts the conversion into a shared _audio_buffer_to_base64 helper
(clip to [-1.0, 1.0], scale by 32767, cast to int16 — matching the
existing convention in voice/input.py) and uses it in both _stream_audio
and _audio_to_base64. The conversion is non-mutating (np.clip returns a
new array), so the buffer retained in _turn_audio_buffer for tracing is
unchanged and int16 input is unaffected.

Test plan

• Parametrized coverage verifies:
  • int16 pass-through,
  • float32 clipping and PCM16 conversion,
  • exact transmitted bytes, and
  • input-buffer immutability.
• code-change-verification: format, lint, typecheck, and full tests passed.

Checks

• I've added new tests, if relevant
• I've run .agents/skills/code-change-verification/scripts/run.sh
• I've confirmed all verification steps pass
• If using Codex, I've run /review before submitting this PR