SedaKit: add an in-process Swift API for macOS

[pnegahdar]

Outcome

A macOS app can add Seda through Swift Package Manager and transcribe microphone audio without starting a subprocess, opening a port, or speaking the HTTP protocol.

Public API

let seda = try await Seda.install(profile: .compact) { progress in
    modelProgress = progress
}

let microphone = try await seda.microphone(language: .english)

for await revision in microphone.transcripts {
    editor.replaceTranscript(revision.text, stable: revision.isFinal)
}

let final = try await microphone.stop()
await seda.close()

Architecture

• Add a stable, versioned C ABI with opaque runtime and session handles over the Rust core.
• Keep installation, model verification, inference, and session state in-process.
• Wrap the C ABI with a concurrency-safe Swift API using actors and AsyncStream.
• Capture through AVAudioEngine, convert to 16 kHz mono in Swift, and feed bounded chunks off the audio callback.
• Package the native core as a checksum-pinned SedaNative.xcframework; expose source-level SedaKit through Swift Package Manager.
• Do not expose C, JSON, runtime paths, localhost URLs, bearer tokens, or model download mechanics to normal Swift callers.

Acceptance criteria

• macOS 14+ on Apple silicon
• .compact, .balanced, and .quality profiles share the catalog used by CLI and Node
• automatic, resumable, checksum-verified model install with progress
• offline transcription and buffered/true streaming capability reporting
• microphone permission, start, stop, cancel, interruption, and device-change handling
• deterministic cleanup of audio taps, native sessions, model handles, and tasks
• fixture integration tests exercise the complete Swift to C to Rust path
• a real Parakeet model integration test runs in GitHub Actions
• XCFramework provenance, checksums, licenses, and release automation
• DocC and README examples begin with microphone capture

Deliberate boundary

This package owns local speech and microphone lifecycle. Global shortcuts, focused-app text insertion, and product UI remain opt-in app policy.