feat(mcp): Phase 4 — oneshot pipeline orchestration via MCP

[staging-devin-ai-integration]

Summary

Follow-up from #358 (Phase 1 landed in #359). Replaces #362 which incorrectly proposed routing media through MCP.

Key insight: MCP is the control plane — the agent designs and validates pipelines, but the actual data plane (media processing) runs through proper channels (curl, skit-cli, or the HTTP API).

Use case

A user asks an agent in natural language: "convert file X speech into language Y". The agent:

1. Uses MCP list_nodes to discover available nodes (e.g. Whisper, NLLB, Kokoro)
2. Uses MCP validate_pipeline to design and validate a working oneshot YAML
3. Triggers the actual oneshot run via skit-cli or curl POST /api/v1/process — media flows through the proper HTTP data plane, not MCP
4. Reports the result back to the user

Proposed MCP tools (control plane only)

• generate_oneshot_command — given a validated pipeline YAML and input file path, produce the correct skit-cli / curl command to execute the oneshot
• explain_oneshot_result — help interpret oneshot HTTP response codes and error messages

The agent may also use shell access to execute the generated command directly, keeping MCP purely for coordination and the HTTP API for data transport.

What this is NOT

• MCP does not route media or binary data through tool calls
• MCP does not replace the /api/v1/process endpoint
• The agent is the orchestrator, not the data plane

Dependencies

• Phase 1 MCP infrastructure (feat: implement MCP server #359)