[FEAT] Wave 1: Go embed-sidecar binary (status API, llama-server mgmt, model download, queue)

[four-bytes-robby]

Part of EPIC #117

Tasks:

• W1.1 Repo skeleton + CI (four-opencode-embed-sidecar)
• W1.2 Status API server (always responsive, phase/progress/queue_depth)
• W1.3 llama-server lifecycle (download, spawn, health poll, restart)
• W1.4 Model detection + download (CPU features, GGUF variant selection, HuggingFace)
• W1.5 Embed endpoint + queue (batch, backpressure, non-blocking)
• W1.6 Single-instance guard (port bind detection + connect-to-existing)
• W1.7 Idle timeout 30min + graceful shutdown
• W1.8 GitHub Releases CI (build all 5 platforms, upload artifacts)