feat(skill): add ROS agent integration and permission recovery

[guima-why]

Summary

• add the alicloud-ros-agent external Skill with StartChat code/CLI transports, endpoint/profile controls, SSE event handling, and local relay coverage
• add durable permission waiting/checkpoint recovery across A2A, Web, pipeline, and local Skill integrations while preserving sub-pipeline timeout behavior
• classify cloud tools by read/write intent so read-only operations can be approved automatically while resource mutations remain gated
• extend pipeline handoff/event routing, session backup recovery, and focused offline/real E2E coverage

Validation

• uv run ruff check scripts/a2a/e2e/run_recovery_scenarios.py tests/a2a_e2e/test_run_recovery_scenarios.py
• uv run pytest -q tests/a2a_e2e/test_run_recovery_scenarios.py (64 passed)
• real A2A regressions passed: scenario1, scenario1-performance-backup, fault-after-snapshot, and ask-waiting
• all four ROS stacks created by those regressions were deleted and verified as DELETE_COMPLETE

Follow-up observation

The corrected selection-during-backup runner now proves that the selection request is dispatched inside the injected backup window. The current product routing classifies that concurrent selection as a supplement interrupt instead of candidate input, so that real-cloud scenario remains a focused follow-up rather than a passing validation.