Harden AI approval defaults and first-run warning

[kyroskoh]

Context

AI can run commands on connected hosts; a “never ask” mode plus run_command is root-risk for default users.

Goal

Safer out-of-the-box AI approval behavior and clear first-run / settings messaging.

Acceptance criteria

• Default approval mode prefers ask-before-run (not “never ask”)
• First-run (or first AI use) warning explains what the agent can do and approval modes
• Settings copy documents each mode clearly
• Existing “never ask” remains available only as an explicit advanced choice (or justified if removed)
• Tests cover default / approval gating where feasible (test/assistant-approval.test.js)

References

• src/main/ai/tools.js, assistant approval UX