Make Calls audio-first with optional video

[nikshilov]

Result

Adds explicit Calls modes: Don’t record / Audio only / Audio and video. Call audio has absolute priority, uses independent microphone and system PCM tracks, survives GUI failure through the same-signed launchd helper, and does not restart when video is toggled. Call video is fixed-display HEVC/H.264 hardware-only capture with crash-forward segments, explicit spans/gaps, export/query/delete/trim coverage, and native-screenshot yield.

Closes #82, #83, and #84 only after exact installed-candidate qualification.

Exact candidate

• Source: 2f925e3e7dae2779b090b53ed583ad336c37c62b
• Version: 0.9.0 (39)
• Notarization: Apple Accepted, submission 55774fbd-71d4-4544-bf5f-68b7fd3131d4
• Installed locally from the stapled candidate ZIP
• Automated insurance: Debug GUI build, capture protocol self-test, 1366 unhosted tests; 19 skipped, 0 failed

Physical evidence already passed

• Real microphone and system PCM both advance and finalize.
• Killing only the GUI leaves both PCM tracks advancing; relaunch adopts the same Call.
• Killing the audio helper creates explicit bounded gaps and resumes the same Call in a new epoch.
• Live audio → video → audio → video stays in one audio epoch and creates exact separate video spans.
• Build 39 records real continuous video: 274 HEVC frames over 18.675 s at 1920×1080/15 fps; a longer Call produced six crash-forward segments.
• Opening the native Screenshot panel stops Call video before capture and does not interrupt either audio track.
• Failed candidates 36–38 and their causes are recorded in Capture Call system audio without any ScreenCaptureKit video leg #82/Call video retains only one frame with ScreenCaptureKit queue depth 1 #84 and project docs.

Hard blockers before ready/merge/release

• A physical Shift–Command–3 sample on installed build 39 has not happened; two two-minute observation windows expired without user input. No synthetic event substitutes for it.
• Direct command-line native captures on build 37 were still 1.60–1.84 s against a 0.33–0.44 s baseline; exact user-hotkey latency must be ≤ baseline +250 ms.
• Remaining physical matrix still includes longer audio-only/video Calls, two monitors and display removal, lock/unlock, low disk, crash/relaunch on the exact final candidate, delete/export/trim, and REST/MCP/UI agreement.

This PR stays draft. Do not merge or publish until the installed physical gate is complete.

[nikshilov]

Build 39 physical result blocked merge/release: the first user screenshot failed because macOS started it 58 ms after Eye began asynchronous Call-video stopCapture(); the retry was delayed. Audio stayed continuous. Commit e8d163c5af19380cf0f98cbd47e62366090ff4dd prepares build 40 by pausing frame admission without tearing down the SCK stream during the screenshot window. Local verification passed: Debug GUI build, 1,367 unhosted tests (19 skipped), Call fixture gate, and capture coexistence self-test. PR stays draft until exact installed build 40 passes the real hotkey latency/audio-gap check.

[nikshilov]

Release remains blocked by #82. Exact notarized build 40 preserved both Call audio legs and continuous video, but Shift-Command-3 still took 1.49 s versus a 0.47 s no-video baseline. A same-signed output-detach probe took 1.57 s and was reverted. Bare ScreenCaptureKit and AVCaptureScreenInput probes both bottom out around 0.80-0.84 s while active, still above the +250 ms gate. Branch is clean again at e8d163c; no merge or public release.