[ling101w]

Summary

• add an evidence-gated intervention profile with risk_only, attack_signal, confirmed_attack, and safety_boundary evidence classes
• preserve the raw policy decision while gating the effective intervention on causal certainty and confidence thresholds
• keep hard safety boundaries enforceable independently of attack attribution
• expose evidence metadata through plugin records and the native AgentDojo bridge, with operator documentation
• add focused unit, bridge, property, and native protocol coverage; enforce portable network-path boundaries across Windows and Linux

Verification

• npm run ci (478 passed, 1 skipped)
• python -m pytest -q (144 passed)
• GitHub Actions CI: https://github.com/ling101w/AgentSentry-/actions/runs/32045451074