fix(cli): pause wake-word listener before manual push-to-talk recording

[pierrenode]

Summary

The classic CLI's manual push-to-talk hotkey (Ctrl+B) can open a second, independent microphone input stream while an active wake-word listener's own stream is still open on the same device — a mic-contention race, the same class the desktop frontend already fixed for its button/hotkey path.

Root cause

• _on_wake_word (the wake-triggered capture path) already calls pause_listening(owner=self) before starting the recorder, because the wake detector owns the mic when it fires.
• The manual push-to-talk path (handle_voice_record's _start_recording closure, dispatched to a daemon thread so it never blocks the prompt_toolkit event loop) calls cli_ref._voice_start_recording() directly with no such handoff — grepping cli.py confirms pause_listening was only ever called from _on_wake_word.
• If a user has wake_word enabled (so WakeWordDetector holds an open sd.InputStream) and also presses the manual record hotkey, AudioRecorder._ensure_stream() opens a second independent sd.InputStream on the same device. On platforms where the capture device is effectively single-owner (Windows), this raises a PortAudio error surfaced to the user as "Voice recording failed: ...", while the wake listener silently keeps running untouched.

Fix

Mirror _on_wake_word's existing pattern inside the push-to-talk _start_recording closure: try pause_listening(owner=cli_ref) before starting the recorder and, on success, set _wake_suspended = True. The pre-existing wake watchdog (_start_wake_watchdog, already running whenever wake word is armed) picks this flag up and calls resume_listening() once recording goes idle — the exact same resume mechanism already used for the wake-triggered path, so no new resume logic is needed.

Tests

• tests/tools/test_voice_cli_integration.py::TestKeyHandlerNeverBlocks::test_start_recording_pauses_wake_word_first: source-order regression test (mirrors this file's existing static/AST-based tests for the same unblockable-closure code, since the handler is a keybinding closure defined inside __init__ and isn't practically invokable in isolation — the file's own _make_voice_cli() helper exists specifically to bypass __init__).
• Mutation-verified: reverting the fix makes the new test fail (pause_idx == -1).
• Full neighbor suite (test_voice_cli_integration.py, test_wake_word.py, test_voice_mode.py): 275 passed (up from 274), 15 pre-existing failures unrelated to this change (confirmed identical on unmodified code — real-audio-device/network-fetch/WSL tests that don't run in this sandbox).
• ruff check: clean.

Checklist

• Verified against live HEAD code, not just commit messages
• Regression test added and mutation-verified
• No rival PR found (push-to-talk wake microphone, wake word microphone contention, pause_listening cli, voice record wake — closest hit is feat(voice): chat UX polish — busy-aware silence, stop hint, thinking sounds, barge-in fix #74000, an unrelated Desktop/TypeScript voice-UX PR opened today that doesn't touch cli.py)

[teknium1]

Thanks for tracing this to the classic CLI path. The premise remains valid on current main: cli.py:15705-15713 starts the manual recording thread and calls _voice_start_recording() without the wake handoff, while cli.py:12614-12623 already uses pause_listening() plus _wake_suspended for wake-triggered capture.

Problems

• The new test reads cli.py and asserts source ordering. AGENTS.md:1382-1416 bans source-text tests because they validate implementation shape rather than runtime behavior. The test should not be carried forward as written.

Suggested changes

• Replace the source inspection with a behavioral seam: extract the handoff into a small testable helper, or invoke the registered keybinding with mocked pause_listening and _voice_start_recording and assert their call order plus the suspension flag.

Current main substantially restructured tests/tools/test_voice_cli_integration.py, and GitHub reports this PR as dirty, so the test needs rework during salvage. This is an automated hermes-sweeper review.

[teknium1]

Please replace this source-text assertion with a behavioral test. AGENTS.md:1382-1416 explicitly bans tests that read source files; this can pass while runtime wiring is wrong and fails under behavior-preserving refactors.

[pierrenode]

Rebased onto current `upstream/main`. This file (`tests/tools/test_voice_cli_integration.py`) was heavily restructured/pruned upstream in the meantime — the `TestKeyHandlerNeverBlocks` class (including this test) no longer exists there. The bug itself is still fully present on current `main` (`_start_recording`'s closure still has no `pause_listening` call), so re-applied the fix fresh against the current file rather than replaying the old diff through the rebase.

Also rewrote the test to be behavioral, per AGENTS.md's ban on source-inspection tests ("A test that reads a source file's text is testing the shape of the source code, not its behavior... banned outright"). The original test read `cli.py`'s text and string-searched for `pause_listening`/`cli_ref._voice_start_recording()` ordering — exactly the banned pattern, and it would happily pass on a correct-looking-but-wrong reformatting while missing an actual behavioral regression.

Fix: extracted the closure's pause-then-record logic into a standalone `HermesCLI._manual_ptt_start_recording()` method — mirroring `_on_wake_word`'s existing, already-real method pattern almost verbatim — instead of an inline nested function. `handle_voice_record`'s daemon thread now dispatches straight to it. This makes the logic independently callable in a test with real (mocked) collaborators, per AGENTS.md's own "Do write" example (extract into a small DI-testable function, call it for real).

3 new behavioral tests call `_manual_ptt_start_recording()` directly against a real `HermesCLI` instance (`_make_voice_cli()`, this file's existing helper) with a mocked `pause_listening`/`_voice_start_recording` pair, asserting on the actual call order and `_wake_suspended` side effect — not the source text. Mutation-verified: swapping the two try-blocks' order in `_manual_ptt_start_recording` breaks the ordering test; making `pause_listening`'s return value unconditional (dropping the `if`) breaks the "declines" test.

Full `tests/tools/test_voice_cli_integration.py` (34 tests) passes. Ruff clean. `tests/tools/test_wake_word.py` has one pre-existing, unrelated failure (missing `ai-edge-litert` tflite runtime on this Mac) — confirmed identical against clean `upstream/main`.

Textual-proximity note: open PR #79696 (wake-triggered continuous hands-free conversation) modifies the exact same `_on_wake_word`/`_start_wake_watchdog` boundary this PR inserts `_manual_ptt_start_recording` right after — different bug, will need a small rebase against whichever lands first, no semantic overlap.

Squashed to a single commit on top of current `upstream/main`.