feat(voice): play berd-call status sounds

[johnmatthewtennant]

Summary

Adds VCCLI-style status sounds to berd-call start, so you can hear when the agent is working on what you said. Status sounds are off by default.

• With status sounds on, the call is Working from the moment a user utterance (or an Expert-Spokesperson handoff) is delivered until the agent's reply is admitted, and Waiting otherwise. working plays the Working cue (Pop); working-and-waiting also plays the Waiting cue (Purr). In conventional mode, cues use the call's input controls so they are not transcribed as user speech, and they hold off while speech is still being recognized.
• berd-call settings --status-sounds off|working|working-and-waiting changes the mode during a call, and the menu bar has a matching Status Sounds submenu. The mode and the current Working or Waiting status survive settings --restart.

Reviewer-reproducible example

Route the call through BlackHole: set the system output to BlackHole 2ch and the system input to BlackHole 16ch, then speak into the call with say.

berd-call start --port 5363 --stream --voice Aaron --language en-US &
berd-call status --port 5363                     # "statusSounds": "off"
berd-call settings --port 5363 --status-sounds working
say -a "$(say -a '?' | awk '/BlackHole 16ch/{print $1}')" "Tell me something interesting."
berd-call settings --port 5363 --restart --voice Aaron --language en-US
berd-call speak --port 5363 --re 1 "Octopuses have three hearts."

The utterance arrives as cursor 1 and the Working cue repeats on BlackHole 2ch, including after the restart. It stops once the reply is spoken:

1	user	Tell me something interesting.
0	lifecycle	session restarted; transcript cursors reset

In the menu bar, choosing Status Sounds → While Working and Waiting changes berd-call status to "statusSounds": "working-and-waiting".

[johnmatthewtennant]

🤖 Fanout on 8f736bb was dispositioned in bec6cf4. The fixes: poisoned settings locks now surface an error instead of being swallowed; the restart responder was separated from SessionStart; conventional-mode cues now use the call's input controls, so they aren't transcribed; and Working status now carries across settings --restart (verified live: the cue kept playing after the restart). Expert-Spokesperson mode has no VoiceInputControls, and Berd's own Expert-Spokesperson runtime (openai_realtime.rs) also plays status sounds without them, so I left that path at parity. The PR body now uses a menu-enabled example with restart evidence, and I removed the timing diary.

[johnmatthewtennant]

🤖 Disposition of the review at 68a97aa8:

• Suppress cues until recognition finishes (P2): accepted. Conventional-mode cues now wait while recognition is pending. The test failed first (status_cues_wait_for_pending_recognition_and_suppress_actual_audio). Expert-Spokesperson cues do not mute input, so that path is unchanged.
• Restart-response refactor (scope): accepted. SessionStart::restarted is restored; the diff only adds the two carried fields.
• Status-sound mode has two owners: declined. The control-side copy mirrors muted, and it is written only after the actor acknowledges. Moving it into the session snapshot would change the session protocol for a setting the child never reads.
• Conversation-status policy is scattered: declined for now. It is three one-line transitions next to the events that cause them. A policy function is worth doing when agent hooks add real working signals.
• PR body (P1/P2): updated with BlackHole routing and no cue timing.

Also, per John: status sounds are now off by default until agent hooks exist (calls_start_with_status_sounds_off_until_agent_hooks_exist, red first). I verified through BlackHole: a new call reports off; with working on, the cue repeats after an utterance, continues across settings --restart, and stops after speak.

[johnmatthewtennant]

🤖 Review at ffaf585a: accepted the settings-choice error missing --status-sounds (red-green, settings_choice_error_lists_every_setting) and the constructor default (status-sound state now passes through SessionActor::new). The summary no longer mentions hooks. Declined again: the two-copies finding (same reasoning as before), and splitting SessionStart, since restoring restarted there was the smaller diff the previous review asked for.

[johnmatthewtennant]

🤖 Review of e20ef98d (4 findings):

• Declined — keep the two status-sound assignments at the call site (scope-minimality). This conflicts with the earlier beauty finding, which I accepted: that finding asked for construction-time state rather than a patched-in default. Passing the values through the constructor removes an invalid intermediate state, and this crate already has too_many_arguments allowances.
• Declined, repeated from earlier — two host-side mode copies (beauty). The reason is unchanged: this mirrors muted, and the host copy is written only after the actor acknowledges the change.
• Declined, repeated from earlier — SessionStart mixes carried state with the restart reply channel (beauty). Splitting it would undo the earlier accepted scope revert. A carried-state struct can come later if restart state grows.
• Declined — exercise the Waiting cue in the reviewer example (pr-template). Waiting depends on conversation-status transitions that only agent hooks will drive. Status sounds are off by default until those hooks exist, so the example covers the mode that works today.

Separately, BlackHole testing found a menu bug, now fixed red-green in #345 (61192d97) and merged here as 8636c923. After a menu action, the menu kept up to a second of stale status. A quick second click on Mute Microphone or Mute Input During TTS then resent the same toggle instead of reversing it. Each action now refreshes status before returning.

[morgmart]

🤖 Automated code review

APPROVE. The exact prepared three-dot comparison was reviewed and has no publishable findings. Supplied GitHub check evidence is structurally valid but currently pending, so required checks still govern merge readiness.

Deterministic publication result: 0 blocking and 0 non-blocking inline finding(s) publishable; 0 duplicate(s) suppressed; 0 blocking screenshot-evidence requirement(s) in this review body.

Pending checks: 1 check(s) are not complete.

This approval reflects the completed code review only; merge readiness remains governed by the repository's required checks.

[morgmart]

🤖 Automated code review

REQUEST_CHANGES. The exact prepared three-dot comparison has one blocking lifecycle race: a newly acknowledged status-sound mode can be lost when a restart overlaps the settings response. Supplied GitHub evidence is structurally valid; several required checks were still in progress when captured, so those checks separately govern merge readiness.

Deterministic publication result: 1 blocking and 0 non-blocking inline finding(s) publishable; 0 duplicate(s) suppressed; 0 blocking screenshot-evidence requirement(s) in this review body.

[morgmart]

🤖 Automated code review

APPROVE. The exact prepared three-dot comparison was reviewed through the project code-review and user wes-review passes. The prior blocking restart race is fixed by sharing actor-owned acknowledged status-sound state and committing it before replying; no publishable findings remain. Supplied GitHub evidence is structurally valid, with several checks still in progress, so required checks separately govern merge readiness.

Deterministic publication result: 0 blocking and 0 non-blocking inline finding(s) publishable; 0 duplicate(s) suppressed; 0 blocking screenshot-evidence requirement(s) in this review body.

Pending checks: 5 check(s) are not complete.

This approval reflects the completed code review only; merge readiness remains governed by the repository's required checks.

[johnmatthewtennant]

🤖 Deferred at the user’s request along with agent hooks. Status sounds depend on agent binding and hook activity, so this PR is parked as a draft and excluded from the active migration stack. The existing implementation is preserved for later.