test(ai): reject unsafe and ambiguous generated action plans #91

Exercise the boundary between untrusted model output and deterministic PDF commands.

Coverage

• Unknown tools and options.
• Path traversal and paths outside user-selected inputs and outputs.
• Invalid, overlapping, reversed, or excessive page ranges.
• Prompt content that asks for shell commands or hidden actions.
• Ambiguous requests that require a user choice.
• Model output that claims an operation succeeded before execution.

The test suite must prove rejected plans cannot invoke a Tauri command.