Add macOS and Windows support

[bnovik0v]

Goal

Make ghst run on macOS and Windows, not just Linux/PipeWire.

Current state

Only one place in the code is hard-coded to Linux: startCapture / findDefaultSink / ensureMonitorVolumeFull in src/main/index.ts (lines ~66-110). It spawns pw-record and shells out to pactl. The platform guard at line 68 throws on non-Linux.

Everything else is already cross-platform:

• setContentProtection(true) is already gated on darwin / win32 (src/main/index.ts:161).
• Hotkeys use CommandOrControl+....
• keyStore.ts uses Electron safeStorage — Keychain (macOS) / DPAPI (Windows) work out of the box.
• The VAD / Whisper / copilot pipeline in src/core/* and the worker renderer is pure web APIs.

Missing:

• mac and win targets in package.json#build.
• Cross-platform audio capture strategy.
• CI matrix / signing / notarization.

Proposed approach: electron-audio-loopback in the worker renderer

The README and CLAUDE.md already reference this as the planned cross-platform path. It exposes getDisplayMedia({ audio: true }) with system-audio loopback on macOS 13+ (via ScreenCaptureKit) and Windows 10+ (WASAPI loopback). The captured MediaStream plugs directly into the existing AudioWorklet → VAD → Whisper pipeline that already lives in the worker renderer.

Pros: pure JS, no native module to build/sign/notarize.
Cons: macOS requires Screen Recording permission; a system picker may appear on first capture per session.

Native-module alternatives (ScreenCaptureKit addon on macOS, WASAPI addon on Windows) are possible but add a large build/sign burden. Skip unless the loopback UX is unacceptable.

Plan

1. Refactor startCapture into a CaptureStrategy interface under src/main/capture/, with linux-pipewire.ts (existing logic) and a new strategy that signals the worker renderer to start a getDisplayMedia capture. Pick strategy by process.platform at app.whenReady().
2. On non-Linux, the worker renderer owns the audio source: it pipes the loopback MediaStreamTrack into an AudioContext at 16 kHz and feeds the existing VAD pipeline. Main just sends cmd:capture-start / cmd:capture-stop. Linux keeps the existing pw-record path in main.
3. Extend package.json#build with:
   • mac: dmg + zip, arm64 and x64, hardened runtime, entitlements for audio input + screen recording, notarization config.
   • win: nsis, x64.
4. extendInfoPlist: NSMicrophoneUsageDescription and NSScreenCaptureUsageDescription.
5. Extend .github/workflows/release.yml matrix with macos-14 (arm64), macos-13 (x64), windows-latest. Add secrets for Apple ID notarization (and optionally Windows code-signing cert).
6. README: drop the Linux-only badge, add per-OS install sections, document the macOS Screen Recording permission step.
7. Tests: unit test for the strategy selector. Capture strategies themselves are integration-only.

Open questions

• Minimum macOS version: recommend 13+ (clean ScreenCaptureKit loopback). Drop 12?
• Apple Developer ID for notarization ($99/yr) — required to ship without Gatekeeper warnings. Acceptable?
• Windows code-signing cert ($200-400/yr EV) — optional but avoids SmartScreen warnings.

Out of scope for this issue

• Linux/Wayland ghost-mode (still no per-window screencast exclude API).
• Migrating away from Groq Whisper.