feat(mac): on-device transcription via Apple's SpeechTranscriber

[z33b0t]

Adds a second transcription engine to the Mac client: macOS 26's on-device
SpeechAnalyzer + SpeechTranscriber, selectable in Settings ▸ Transcription
alongside the existing blurtd path. Server mode is unchanged and stays the default.

Until now the client was a thin pipe to a Parakeet server — great once you have
the GPU box, nothing until you do. Apple's model is in the same league on
accuracy (~2.1% WER clean vs Parakeet TDT v3's ~2.5% on LibriSpeech test-clean),
runs on the Neural Engine, and ships zero bytes with the app.

Reach

Tier	Share	Engine
Apple silicon + macOS 26+	~84%	local or server
Everything else	~16%	server, exactly as today

Purely additive — LSMinimumSystemVersion stays 13.0 and nobody loses anything.

Design

AppDelegate already talked to DictationClient through a closure surface, so
that surface became TranscriptionEngine; DictationClient conforms nearly
as-is. Apple's result model lines up with Blurt's, so the HUD needed no changes:

Blurt	SpeechAnalyzer
partialLive	volatile result (ReportingOption.volatileResults)
partialCommitted	finalized results (isFinal) accumulated
onVad / "Hearing you…"	SpeechDetector(reportResults: true)
onInfo "loading"	AssetInventory status / install progress

AudioCapture grows a negotiable output format (the server wants its declared
16 kHz Int16, SpeechAnalyzer names its own) and hands over an
AVAudioPCMBuffer rather than packing Data itself.

Why no #if arch(...) anywhere

SpeechTranscriber.isAvailable answers both halves of "can this Mac do it" —
the API only exists on macOS 26, and even there it reports false without a
Neural Engine, i.e. on every Intel Mac. And because the symbols ship in the
x86_64 slice of the SDK too, the universal build compiles unchanged and
Intel simply takes the server path at runtime.

Settings shows the local option greyed out with the reason rather than hiding
it, and Settings.engine refuses to return .local on a Mac that can't run it,
so a preference restored from an Apple silicon Mac can't strand this one.

Verified

On a MacBookPro16,2 / Intel i7-1068NG7 / macOS 26.5.2 — the negative case,
run against the production source files:

LocalTranscription.isSupported = false
LocalTranscription.unavailable = Needs a Mac with Apple silicon.
stored engine preference = local ← forced
Settings.engine reads back = server ← gate holds

Driving the real SettingsWindowController:

Blurt server (Parakeet, on your GPU box) enabled=yes selected=yes
This Mac (Apple Speech, on-device) enabled=NO selected=no
note: "Needs a Mac with Apple silicon."

Universal build is intact (lipo: x86_64 arm64), signs with Developer ID +
hardened runtime, and the build is warning-free.

Not yet verified — needs an Apple silicon Mac

I only have the Intel machine, so the positive path is unexercised. Before
merging, on Apple silicon + macOS 26:

• Dictate with Wi-Fi off and no blurtd running — the feature test and the
  privacy proof in one
• Volatile → final renders correctly in the HUD
• "Hearing you…" tracks SpeechDetector
• Esc mid-dictation injects nothing
• First run in a locale with no installed model shows "Loading model…"
• Engine switch takes effect next dictation, both directions
• Confirm whether SpeechAnalyzer needs Speech TCC — NSSpeechRecognitionUsageDescription
  is added defensively; if it's required, Onboarding.swift wants a third
  permission row

CI

mac.yml moves macos-15 → macos-26: the engine needs the macOS 26 SDK to
compile and the Sequoia image doesn't ship it.

https://claude.ai/code/session_01SUWiz8CtyyAGuxKgAEVZJu