Problem

The chaos module has strong adversarial testing (8 attack types: prompt injection, policy bypass, privilege escalation, data exfiltration, tool abuse, identity spoofing) and infrastructure faults (latency, errors, timeouts). However, it lacks behavioral conflict scenarios:

• No deadlock detection (two agents waiting on each other's output)
• No contradictory instruction injection (agent receives conflicting directives)
• No dynamic trust perturbation mid-task (trust score changes during execution)
• Enterprise fault types (delegation_reject, llm_degraded) raise NotImplementedError

What Exists

• adversarial.py — 8 attack techniques + OWASP playbooks ✅
• engine.py — FaultType enum with infrastructure faults ✅
• test_scenarios.py — Trust degradation detection (event-based) ✅

Proposed Solution

1. Add DEADLOCK_INJECTION fault type — simulate circular dependency between agents
2. Add CONTRADICTORY_INSTRUCTION fault — inject conflicting directives mid-task
3. Add TRUST_PERTURBATION fault — dynamically change trust scores during execution
4. Implement the stubbed delegation_reject and llm_degraded enterprise faults
5. Add corresponding experiment templates to ChaosLibrary

Acceptance Criteria

• Three new fault types in FaultType enum
• Deadlock scenario with timeout detection
• Contradictory instruction injection with conflict resolution test
• Trust perturbation during active task execution
• Experiment templates in ChaosLibrary
• Unit tests for each behavioral fault

/cc @imran-siddique