[Newarr]

Scribe retained roughly 4 GB of MLX GPU buffers after local transcription. Local transcription and language detection now release model references, finish GPU work, and clear the cache before returning, including error paths.

Pin the tested dependency graph and supply the compiled Metal library to GPU tests in CI. MLX 0.32.3 and LM 3.32.3 failed compilation in the audio fork, so retain MLX 0.31.3 and LM 3.31.3.

Declare the immutable model download worker Sendable so CI can build with Swift 6.3.

Validation:

• CI passed: 613 tests passed, one optional integration test skipped, and the full app built.
• Real native transcription peaked at 5.35 GB, then retained 0 cached bytes and 16 active bytes before process exit.
• The full app built with the pinned versions.
• Source and dependency reviews passed with no actionable findings.