[feat] Support system messages in governed one-shot runs

[willsarg]

Problem

ara run exposes only a user prompt. Many instruct models and normal automation workflows distinguish a system instruction from user content, but ARA users must currently flatten both into one user message and lose the model's native chat-role semantics.

Proposed change

Add an optional --system TEXT to ara run across native and Ollama lanes.

ARA should pass the system and user messages through the selected model's supported chat-template/runtime interface. Both messages and all template expansion must be included in the existing governed token accounting.

Acceptance criteria

• Runs without --system remain byte-for-byte compatible at the public CLI/backend boundary.
• Supported models receive distinct system and user roles rather than concatenated text presented as equivalent behavior.
• The fully rendered message sequence is counted before generation and refused if it cannot fit within the effective governed context.
• A model/template/runtime that cannot represent a system message refuses clearly before generation; ARA does not silently concatenate or drop it.
• Native transformer, native GGUF, and Ollama adapters use equivalent public semantics while retaining their engine-specific rendering.
• JSON indicates that a system message was supplied without echoing its content.
• Tests cover supported templates, unsupported templates, token-expansion refusal, Ollama request construction, and unchanged user-only behavior.

Non-goals

• Persistent chat sessions or conversation history.
• A prompt-template registry.
• Persisting or displaying system-message contents.
• Changing sampling defaults.

Independence

This is a message-role feature over the current governed run path. It does not require stdin/file prompts, run --ctx, provenance receipts, or any other open issue.