feat: add reliable macOS audio capture streams

[marcusschiesser]

Closes #262.

Supersedes #257 with the reliable, pull-based stream design discussed in the replacement issue. File/container writing remains out of scope and is tracked separately in #263.

Summary

• add macOS 15+ system-audio, microphone-only, and combined capture using ScreenCaptureKit and AVFoundation
• expose separate, timestamp-aligned signed 16-bit PCM source bytes through reliable Zig effects and TypeScript commands
• keep accepted frames in a bounded runtime ring; stop with consumer_too_slow instead of overwriting when the app does not drain it
• enumerate microphone devices, report invalidation changes, and provide explicit non-prompting status/request permission operations
• journal and replay lifecycle/read deliveries, including the PCM bytes consumed by the app

Stream contract

• one active capture per app
• 20 ms paired blocks, reads bounded to 20–100 ms
• default 5 second ring, configurable from 1–30 seconds
• disabled sources return an empty byte slice
• timestamp gaps are zero-filled and counted independently from real acoustic silence
• stop seals producers but preserves buffered frames for draining; discard releases immediately
• no file output, mixing, encoders, gains, or mode option

Platform and packaging

• system and combined capture use ScreenCaptureKit on macOS 15+
• microphone-only capture uses AVFoundation without requiring Screen Recording access
• older macOS, Chromium, Windows, Linux, and mobile hosts report the features unsupported
• add system_audio manifest permission and microphone/system-audio privacy strings, generating the relevant Info.plist purpose keys
• leave App Sandbox entitlements app-owned and document com.apple.security.device.audio-input

Verification

• zig build test
• zig build validate
• npm run gate -- --skip-bench in packages/core
• pnpm --dir docs check
• Zig and TypeScript audio-capture example tests
• macOS audio-capture example build/link against AppKit, AVFoundation, ScreenCaptureKit, AudioToolbox, and CoreMedia
• managed and standalone example checks, including the post-rebase gpu-components target
• mobile layout/package checks
• Objective-C syntax/availability build for the audio backend

Local scripts/gate.sh fast has three environment/baseline limitations worth calling out:

• the aggregate native-example shard stalls in Zig while the terminal example resolves/downloads its pinned Ghostty dependency; the audio-relevant standalone and registered example targets pass independently
• bench-render misses the existing toggle-big-view 9 ms budget at 10.18 ms; the same command on origin/main fails at 10.22 ms
• the CEF link check skips because this checkout has no prepared CEF distribution; the Chromium host change is a signature-matched unsupported stub

Manual TCC prompt/restart testing remains a draft-PR follow-up because it requires interactive Screen Recording and microphone grants on a test machine.

[vercel]

@marcusschiesser is attempting to deploy a commit to the Vercel Labs Team on Vercel.

A member of the Team first needs to authorize it.

[marcusschiesser]

tested with https://github.com/marcusschiesser/desktop-apps/tree/main/meeting-notes apps see PR marcusschiesser/desktop-apps#2