[juw12]

Talk mode: a real-time spoken conversation with Buddy (Ctrl+Alt+T)

Press Ctrl+Alt+T (or Talk with Buddy in the tray) and talk, no key to hold.

• Knows when you are done: keyit_talk.TurnTaker reads the words at every pause and waits 0.3 s after a complete
  sentence, 0.7 s when it cannot tell, 1.6 s when you trail off ("and", "because", "um"). English, Indonesian and Chinese
  rules; other languages get the plain 0.7 s pause.
• Speaks mid-answer: the voice starts on the first sentence while the brain is still writing the rest.
• Barge-in: EchoGate compares the microphone with what Buddy is playing and stops the voice when you talk over it.
• Ends with Ctrl+Alt+T, Esc or "goodbye". Push-to-talk and dictation are unchanged.

Gates (frozen rubric buddy-talk-v1)

gate	result
G1 tests pass locally	PASS: 203 passed (main: 177 passed, 1 skipped)
G2 CI green on this PR	PASS: core and windows
G3 barge-in, onset to stop requested, p95 <= 500 ms	PASS in simulation (<= 0.2 s) with a headset, or when the user is 3x (natural voice) / 4x (Windows voice) louder than the echo. Stop on the real sound device: not measured
G4 one decision, p95 <= 300 ms	FAIL: 0.87 s with Whisper at about 90% CPU load (rules alone: 0.005 ms)
G5 balanced accuracy >= 85% and above fixed silence at the same delay	PASS on the audio timeline: 85.4% vs 50.0% (48 synthetic clips, a 0.4 point margin). With compute time under this load: 72.2%, about equal to the best fixed pause (72.9%)
G6 zero false barge-ins at >= 3 echo levels	PASS in simulation: 0 in 24 min, 4 levels x 2 voices
G7 push-to-talk and dictation unchanged	PASS: existing tests
G8 honest claims	README states where Buddy does not beat Jev and what was not measured

Not merged automatically: G4 fails. Giving Whisper only the last 1.2 s still takes 0.43 s (p95) at 100% CPU, so it needs a streaming recognizer or a quiet PC. Council score (3 judges): 1.3.2 4.0, this PR 7.62, blind pair 3-0. Jev was not run; its figures are the vendor's published ones.

Not verified

A real person at a real microphone; a real loudspeaker-to-microphone echo; voice stop on the real output device
(bench part d did not start); the talk mode UI running end to end; latency on a quiet PC; brain latency on CPU-only Ollama.

Bench: bench/talk_turns.py, results in bench/results_talk/, write-up in bench/README.md (Talk mode) and
README.md (Talk mode vs Jev).