fix(wake): stabilize exclusive and duplex audio handoffs

[seppegadeyne]

What does this PR do?

Stabilizes the wake-to-voice audio handoff on exclusive input devices and opt-in full-duplex HFP transports.

Changes

• Release the CLI AudioRecorder after every wake-suspended capture, before transcription or follow-up capture reacquires the device.
• Add optional wake_word.duplex_output_device support: keep a silent 16 kHz output stream open while wake capture uses a bounded callback queue.
• Close both wake streams before dispatching the wake callback, preventing the command recorder and first cue from racing stream teardown.
• Add voice.output_device for explicit cue routing and a 250 ms silent preroll for the first cue after an HFP wake.
• Preserve ordinary non-wake recorder reuse and the existing input-only wake path.

Deployment-specific selectors, launchers, STT-provider changes, and local wake engines are intentionally excluded.

Verification

• Six focused regressions: 6 passed.
• Full affected slice on the PR branch: 96 passed.
• Same patch applied cleanly to current upstream main: 116 passed; ruff and git diff --check clean.
• Raspberry Pi 5 + Jabra Speak2 55 over Bluetooth HFP/mSBC: first wake phrase detected on the first attempt, first cue audible, command/TTS/follow-up completed, wake resumed after the 60-second cap, and no BlueALSA overruns or device-busy errors were observed.

Related context

Complementary to #74152 and #74996.

Type of Change

• Bug fix
• Tests
• New feature
• Security fix

Checklist

• Scope is limited to the audio lifecycle fix and its regressions.
• Current main compatibility was verified.
• Real-device end-to-end validation completed on Raspberry Pi OS 13/aarch64.
• No deployment credentials or device-specific values are included.

[teknium1]

Thanks for the focused lifecycle fix. Current main confirms the premise: cli.py:12027 calls AudioRecorder.stop(), and tools/voice_mode.py:1112-1127 documents that this preserves the stream. The wake path sets _wake_suspended at cli.py:12617-12626, while its watchdog later calls resume_listening() at cli.py:12682-12701.

The proposed finally release is correctly scoped to wake-suspended, non-continuous captures, preserves ordinary recorder reuse, and runs even when stop() fails. The added tests cover release-before-transcription, finalization failure, and regular-mode persistence. The PR is one commit ahead of current main with no divergence, so it should cherry-pick cleanly.

Automated hermes-sweeper review.

[seppegadeyne]

Closing this draft — we won't have time to continue this line of work for the foreseeable future. Keeping the branch around in case we pick it back up later.

[seppegadeyne]

Closed — no time to continue this work for now.