fix(voice): rebuild inactive recorder streams

[andyst-dev]

Summary

• reuse the recorder's cached InputStream only while its active liveness signal remains true
• close and rebuild inactive or unprobeable streams before the next recording, using the existing timeout-bounded shutdown path
• still attempt close() when stop() rejects an already-inactive stream
• add behavioral regression coverage for active reuse, inactive replacement, and failed liveness probes

This addresses the dead cached recorder mechanism isolated in #75417. It is intentionally complementary to #74152, which covers the separate manual push-to-talk wake-listener handoff and does not touch tools/voice_mode.py.

Refs #75417

Validation

• scripts/run_tests.sh tests/tools/test_voice_mode.py -q -k 'ensure_stream' — 3 passed
• related voice/wake/STT suites — 118 passed
• four unrelated platform-specific failures were excluded from that related-suite run after reproducing each on clean upstream/main (two WSL playback mocks, the macOS beep gate, and the missing local TFLite runtime case)
• uv run ruff check tools/voice_mode.py tests/tools/test_voice_mode.py
• python -m py_compile tools/voice_mode.py tests/tools/test_voice_mode.py
• git diff upstream/main...HEAD --check

[teknium1]

Thanks for the focused regression fix. The premise is confirmed on current main: tools/voice_mode.py:896-897 reuses every non-null cached stream without checking whether it remains active, and tools/voice_mode.py:1096-1101 skips close() if stop() raises. The PR's liveness gate and independent close attempt directly cover both paths, with behavioral tests in tests/tools/test_voice_mode.py:427-480.

The only production caller of _ensure_stream() remains AudioRecorder.start() at tools/voice_mode.py:1082; no sibling recorder lifecycle path was found. The checked-out main is an ancestor of the PR base, so no surrounding-code rework is indicated.

This is an automated hermes-sweeper review.

[andrexibiza]

Vox Lockin lane 09 cross-reference receipt (desktop lane; this PR touches the CLI recorder in tools/voice_mode.py, outside the desktop hook lane):

• Merge-clean against current main (70db671).
• Reviewed: AudioRecorder._ensure_stream rebuilds an inactive/broken sounddevice stream (reuse active, close+rebuild inactive, tolerate broken liveness probes) — the CLI voice-loop analogue of the desktop's mic re-arm class. Note transcribe_recording in the same file serves the desktop /api/audio/transcribe endpoint, but the recorder-stream change does not affect the desktop transcription path (desktop recordings come from the browser's MediaRecorder, not AudioRecorder).
• Tests on PR head: 3 new TestAudioRecorder cases cover reuse/rebuild/probe-failure; consistent with the existing mock harness.

No desktop-class gap; recommend merge. (Lane 01/03 may want to run the full tests/tools/test_voice_mode.py suite for the CLI loop.)