Hands-free: Stop steps down during a take (exit lock / cancel), noise-loop exit, stale-audio clear

[KeithCu]

Problem

When hands-free (hold Record about 2 s) is on, the gap between turns is just another take. Stop was greyed out during every take (send_state.py RECORD_CLICKED had stop_enabled=False, and the deal contracts required that). So the only ways out were Clear (which wipes the chat) or closing the sidebar. Headed tester report: Stop is greyed out while hands-free is listening between turns. Stop Rec can't be the exit, because it is the send, and in loud rooms where silence auto-stop never fires you need it to send.

Change

Stop during a take now steps down one level per click:

State	Stop click
Locked (hands-free) take	Lock off, status "Hands-free off". The take keeps recording, so Stop Rec or silence auto-stop still sends it once, and nothing re-arms afterwards.
One-shot take	Cancelled. The new CANCEL_REC_CLICKED / CancelRecordingEffect runs recorder cleanup and deletes the WAV. Nothing is sent, the status is "Recording cancelled" (not "Error"), and the button goes back to Record.
Transcribing / model replying / speaking	Unchanged. It still ends hands-free, and Stop during a reply does not re-arm.

• Stop Rec is unchanged: it sends and keeps the lock.
• Contracts: the send_state deal contracts now allow Send and Stop to be enabled together only while recording. The why is in the comment above next_state.
• Where the step-down runs: StopButtonListener.on_action_performed, via SendButtonListener.stop_during_take. The GTK mousePressed hook still only acts while a send is busy, so one click is one step.
• Speech finished: the speech-finished callback no longer greys out Stop when a take is already recording (barge-in or a hands-free restart).
• Status while locked: silence progress reads "Hands-free… (N ms silence)".
• Tooltips: Send and Stop now have tooltips (HelpText) explaining Stop Rec = send and Stop = leave hands-free / cancel. No label changes, so the 1x button row is unaffected.

Other fixes

• Noise loop: an empty transcript ("[No speech detected.]") finishes as Stopped, so hands-free re-armed every time and noise could loop forever. Hands-free now ends after 2 empty takes in a row (EMPTY_TAKES_EXIT). A non-empty transcript resets the count.
• Stale audio: if Stop landed after Stop Rec but before the send drain started, the WAV stayed on audio_wav_path and the next typed Send transcribed or attached it. The skip branch now calls clear_pending_audio_wav().
• Outdated comments: comments in send_handlers._transcribe_audio said Stop kills local Whisper. It doesn't. Stop keeps the transcript for the Ask box, and only sidebar teardown kills the child.
• Docs: docs/chat/audio-architecture.md gets rows for Stop during a take and for empty takes.

Tests

• test_send_state.py:
  • Record enables Stop; typing during a take keeps it enabled.
  • Cancel drops the take with no send, no error status, typed text kept, no-op when not recording.
• test_fsm_verification.py: oracle and Hypothesis invariants updated to the relaxed contracts; CANCEL_REC_CLICKED added to the oracle.
• test_record_gesture.py: step-down decision, exiting the lock does not swallow the next Stop Rec, hands-free silence text, EMPTY_TAKES_EXIT.
• test_panel.py (TestHandsFreeRecord):
  • Stop is enabled during a locked take.
  • Locked take + Stop: lock off and still recording; Stop Rec then sends once and nothing re-arms.
  • One-shot take + Stop: cancelled, no send, no "Error", WAV deleted.
  • Two Stops from a locked take cancel it; mousePressed alone does not step down.
  • Hands-free silence text.
  • Speech finishing does not grey out Stop during a take.
  • Two empty transcripts exit hands-free; a non-empty one resets the count.
  • Stop before the send starts clears the pending WAV.
• Mock sidebar G11 (test_mock_llm_sidebar_uno.py): with an in-process listener, Stop on a one-shot take now asserts cancelled (not recording, not busy, nothing sent).

Local: the related pytest files pass (286), and ruff, basedpyright, ty and mypy are clean on the changed files.

Live checks for the headed tester

• Locked take: Stop is enabled. First click shows "Hands-free off" and keeps recording; Stop Rec sends once and nothing re-arms. A second Stop cancels.
• GTK: one Stop click during a take is exactly one step.
• Tooltips: they show on hover, and the 1x button row isn't clipped.

Mock sidebar hooks fix (2e2a173): in the URP path, a Record click now counts as taken as soon as the label leaves Record or Stop is enabled. Before this, G4's stub auto-stop could send the take before any poll saw Stop Rec, so the fallback started a second take. Master never caught it because URP wait_idle reads "Stop off + Send on" as idle, and Stop used to be greyed out during takes.

CI on 2e2a173: PR CI 37492341929 passed; mock sidebar 37492364772 passed (111 passed, 0 failed).