Voice dictation drops speech on multichannel USB audio input (Steinberg UR22C); works on built-in mic

**Environment**
- Claude Code 2.1.240, VS Code extension, macOS (Darwin 25.6.0), 2019 16" MacBook Pro (Intel i9)
- Input device: Steinberg UR22C "Voice" virtual device (2 ch, 48 kHz, both channels identical) — also tried the "DAW" device (4 ch, 48 kHz)
- Mic: Shure SM7B with an inline booster; input level verified with ffmpeg on the same device: -6 dB peak, -23 to -27 dB RMS

**What happens**
- With the UR22C selected as the system default input, the level meter in the dictation box moves normally, but nothing is transcribed unless I begin with a loud burst ("check, check, check"). After a pause mid-sentence the transcription stops again while the meter keeps moving.
- After "Reload Window", one recording returned only the loud wake phrase and dropped everything spoken after it, on a verified-clean signal.
- With the built-in "MacBook Pro Microphone" (1 ch) selected, dictation works without the wake phrase, including whispered speech.
- Behaviour tracks channel count, not loudness: built-in (1 ch) works; Voice (2 ch) erratic; DAW (4 ch) captured a second then went silent.
- Network is clean during the failures (0% loss, ~40 ms to api.anthropic.com).
- Across a few minutes on the same device I see three different behaviours: (a) text appears live while speaking; (b) no live text, the whole message lands when the mic turns off; (c) nothing lands at all. Starting every recording with a loud "check, check, check" makes (a) far more likely, which suggests the live transcription session needs a moment to come up and speech before that is lost.

**What I expect**
Speech captured on a 2- or 4-channel 48 kHz input device to be downmixed/resampled and transcribed like the 1-channel built-in mic.

**Guess**
The capture path may be reading interleaved multichannel frames as mono (audio arrives slowed/smeared at the transcriber), and/or holding the device it opened first when the system default changes.

**Repro**
1. Plug in a USB audio interface exposing a 2-ch or 4-ch input at 48 kHz; set it as the macOS default input.
2. Press the dictation mic button, speak at normal level.
3. Observe the meter moving with no transcription, unless the recording starts with a loud burst.
4. Switch the default input to the built-in mic and repeat; transcription works.

Reproduce the failure through the dictation mic button with the Steinberg UR22C Voice or DAW device selected as the macOS default input, then compare it with the built-in MacBook Pro Microphone. Trace the dictation capture path used by the VS Code extension, focusing on multichannel 48 kHz input. Done means 2- and 4-channel speech is captured and transcribed reliably like the 1-channel input, including after pauses.

Written by the indexing model from the issue text.