[daniele21]

Product intent

Make speaker-attribution quality measurable and selectively improve local meeting transcripts without risking the established Community-1 experience.

Scope

• Keep Community-1 as the default; add opt-in Nemotron 3 CoreML (fast128 / c128-split-w8a8) in Advanced settings.
• Reuse canonical Swift helper, lease, local model cache, speaker timeline/persistence and the existing API flow. Never transmit meeting audio or silently fall back to a remote backend.
• Invalidate recording pipeline cache on local engine/preset changes; retain manual transcript content and error isolation.
• Add privacy-safe optional RTTM A/B benchmark and deterministic tests.
• Upgrade pinned FluidAudio to v0.17.7 (new runtime API).

Risk and gates

Product: FEATURE | Delivery: INTEGRATION | Validation: STRONG due to native inference, model resource lifecycle, dependencies & speaker persistence semantics | Execution: REMOTE_AUTOMATED for source gates + REAL_ENVIRONMENT for final native model/build/quality proof.

Requested automatic preflight: repo-selected source tests/typecheck only. No CI build, DMG or large artifacts.

PENDING (must not be labeled PASS)

• SwiftPM package resolution + actual Swift compile on a target Mac with Xcode/toolchain.
• Representative Apple Silicon inference (both presets), source-aware diarization quality, memory/thermal and packaged helper smoke.
• Review model OpenMDW 1.1 attribution/packaging obligations before public distribution.

No merge until automated gates pass and native/toolchain compatibility is established; preserve default until measured superiority.