[AI Assistant] Classify tools by confirmation policy

[astandrik]

Goal

Reduce unnecessary approval prompts while keeping explicit confirmation for tools that may change state.

Behavior

• Explicitly classified safe tools, such as documentation search and current-time lookup, run without confirmation.
• Potentially state-changing tools, including restarts and query execution, always require confirmation for each invocation.
• Unknown or unclassified tools require confirmation by default.
• Auto-approval rules and user preferences must not bypass mandatory confirmation for state-changing tools.

Acceptance criteria

• The exposed tool catalog has an explicit confirmation policy.
• Safe, state-changing, and unknown tools follow the same policy in all supported AI Assistant modes.
• Tests cover automatic execution, mandatory confirmation, rejection, and attempts to bypass mandatory confirmation.

Reference

YDB-3538 — AI Assistant architecture