Module runtime 

RFC 1 (issue #20) dedup window + cycle suppression engine.

Pure logic, no IO. AgentHarness::handle_trigger (follow-up sub-PR) wraps this engine into the agent loop entrypoint, but the dedup / cycle decisions live here so they are independently testable.

Behaviour matches RFC 1 §5:

• Dedup window: same idempotency_key seen twice within TriggerRuntimeConfig::dedup_window (default 5 minutes) → outcome depends on the first trigger’s ReplacementPolicy (per RFC 1 §11 fixed decision #4 — sources declare per-event; the runtime trusts the first arrival’s declaration to set the window’s collapse semantics).
• Cycle suppression: when the same trace_id exceeds TriggerRuntimeConfig::cycle_hop_limit (default 5) → forced EvaluationOutcome::CycleSuppressed. Each accepted trigger bumps the per-trace hop counter; the runtime calls TriggerRuntime::record_follow_up_hop before spawning sub-triggers that share the parent’s trace.

Structs

TriggerRuntime

In-memory dedup + cycle registry shared across all NotificationHook sources for a single agent / daemon. Cloning is cheap; the actual state lives behind an Arc<Mutex>.

TriggerRuntimeConfig

Tunable knobs for TriggerRuntime. The runtime never mutates these; callers can swap them via TriggerRuntime::new or TriggerRuntime::with_config.

TriggerRuntimeSnapshot

Point-in-time view of the runtime’s dedup + cycle bookkeeping. Cheap to copy; surfaced via TriggerExecutor::notification_status_snapshot for status banners and /triggers rendering.

Enums

EvaluationOutcome

Result of running a Trigger through TriggerRuntime::evaluate. Subsequent runtime state (state machine transitions, session audit, permission evaluator) consumes this.