[undivisible]

The existing bridge controls FaceTime calls but exposes no media API. This adds opt-in audio support while leaving BridgeClient, helper IPC, and call control unchanged.

• audio provides validated, owned Float32 PCM frames, a bounded drop-newest channel with statistics, synchronous/poll/async consumption, and capture/injection backend contracts. Callers own processing; no runtime or new third-party dependency is added.
• coreaudio-capture adds a macOS 14.2+ process-output tap scoped to an explicit PID, output-device UID, and stream index. A preallocated native ring feeds a Rust worker. Capture defaults to refusal; explicit startup, source-change checks, cancellation, and cleanup errors are documented. Non-macOS builds expose an unsupported-platform stub.
• Injection remains a caller-supplied implementation that must write to a real configured audio route. No virtual driver, FaceTime input-selection API, or end-to-end injection implementation is claimed.
• Add synthetic examples/tests and feature/platform CI coverage. The development manifest is 0.2.0 with Rust 1.85 minimum and resolver 3 to prefer MSRV-compatible dependency versions; features remain disabled by default. Public APIs/workers are Rust; the native shim is Objective-C/C11.

Validation on exact head df8b8e80395eb02471052ecf30432621efcee19a: all 18 serial, locked/offline local gates passed with the configured sccache wrapper unchanged. This includes strict Clippy, default/private-api/CLI/audio/capture/all-feature checks and builds, 13 audio-only tests, 19 capture/all-feature tests, the generated-audio example, actual Rust 1.85 all-feature checks and 19 tests, and Rust 1.85 Linux cross-target checking of all targets. The native sources also passed ASan/UBSan synthetic-buffer and mocked-cleanup tests plus ThreadSanitizer concurrent-ring stress; those native files are unchanged. Independent exact source review found no remaining findings. GitHub Actions CI passed both jobs on this exact head: the full macOS job and fresh Rust 1.85 Linux resolution, 16 portable runtime tests, and the generated-audio example. All audio tests use generated PCM, refused permission, or mocked teardown. No call, live audio, permission prompt, driver installation, or default-device change is involved.

Known limits: process output may include ringing and mixed participants; it is not internal FaceTime/RTP/per-participant PCM. Real permission/device behavior, macOS/SDK 14.2, and FaceTime end-to-end capture/injection remain unverified. Severe HAL teardown failure is surfaced as CleanupFailed and can retain HAL resources until process exit; callbacks then discard input. See docs/audio.md for the full ownership and routing contract.