New detection rule: repeated-denials (agent probing a guardrail)

[blitzcrieg1]

An agent that hits a wall of denied tool calls is often probing a guardrail. A single denial is noise; a burst is a signal.

Task

In apps/orchestrator/core/audit/detection/rules.py, add a pure function rule_repeated_denials(events) and append it to REGISTRY. Fire when a session has >= 4 tool_called events with action.outcome == "denied". Aggregate into one Detection (severity high) listing all offending event_ids. Reuse the existing accessors (_action_type, _outcome, _event_id, ...).

Acceptance criteria

• Fires on a session with 4 denials; does not fire on 3.
• Tests in apps/orchestrator/tests/test_detection_engine.py (one fires, one near-miss), using the _event(...) helper.
• python -m ruff check core api tests and python -m pytest -q pass.

Hint: copy the shape of rule_autonomous_unapproved_write — it already aggregates offending events into one detection.