fix(voice): stop AudioInput.to_base64() from mutating caller's buffer

[adityasingh2400]

Summary

AudioInput.to_base64() currently rebinds self.buffer to an int16 array when the input is float32:

def to_base64(self) -> str:
    if self.buffer.dtype == np.float32:
        # convert to int16
        self.buffer = np.clip(self.buffer, -1.0, 1.0)
        self.buffer = (self.buffer * 32767).astype(np.int16)
    ...

This silently corrupts any reference the caller still holds to the original float32 buffer. A user who creates an AudioInput, calls to_base64() (e.g. for tracing), and then inspects audio_input.buffer again will find an int16 array where they put a float32 one — different dtype, different values.

The companion _buffer_to_audio_file() already does the conversion via a local rebinding (the parameter buffer is shadowed inside the function), so to_audio_file() does not have this bug. This PR aligns to_base64() with that pattern: do the int16 conversion into a local variable and leave self.buffer untouched.

Changes

• src/agents/voice/input.py: rewrite to_base64() to convert into a local int16_buffer instead of mutating self.buffer.
• tests/voice/test_input.py: add a regression test asserting audio_input.buffer.dtype and contents are unchanged after to_base64(), and that calling it twice produces the same encoding.

Diff is 25 lines total.

Test plan

• pytest tests/voice/ — all 31 tests pass
• Regression test fails on main, passes with this change
• ruff check clean

[seratch]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Nice work!

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".