SwiftLM engine: subprocess integration + model loading

[magicnight]

v0.1 detects SwiftLM presence only. Wire up actual subprocess launch, OpenAI-compatible HTTP pass-through, 100B+ MoE model loading. Spec: features/inference-engines.md Engine 2.

[magicnight]

Deferred out of v0.2. Reason: the subprocess launch path (/opt/homebrew/bin/SwiftLM etc.) is blocked by App Sandbox, which we kept ON in v0.1 (#4). Un-sandboxing to enable this engine would reverse that decision and loosen Gatekeeper posture, so we're holding off until there's a concrete 100B+ MoE user asking. Revisit when sandbox policy is reviewed (v0.3 candidate).

[magicnight]

Deferred indefinitely (2026-04-17).

Root cause is structural: SwiftLM is a standalone executable with its own HTTP API — integrating it requires spawning a child process, which macOS App Sandbox blocks without either disabling the sandbox entirely or wrapping it in an XPC helper. Both are substantial engineering commitments that don't pencil out against the actual use case (100B+ MoE models, which in turn require Mac Studio-class hardware).

Not closing as "won't fix" — reopening is welcome if:

• A 100B+ MoE model becomes practical on more common hardware
• An alternative native-MLX 100B+ solution emerges that integrates in-process (avoiding the subprocess requirement)
• The project decides to accept unsandboxed distribution

Analysis + options discussed in the v0.3 retrospective; see .omc/plans/v0.3-overview.md Batch 3 notes.