Sentinel actions: NetworkIntent + deterministic Safety Kernel boundary

[tuxevil]

Follow-up to #17.

Goal

Replace model-authored raw UCI as Sentinel's long-term mutation interface with typed semantic NetworkIntent objects evaluated by a deterministic Safety Kernel.

NetworkIntent

The model expresses domain intent (for example wifi.channel.change) with target, parameters, evidence references, reason, and expected observable outcome.

OMEGA then:

1. resolves logical resources against current state/capabilities;
2. compiles the semantic intent to desired-state delta / exact immutable device plan;
3. evaluates deterministic invariants;
4. classifies risk and authorization;
5. applies controller-owned canary/staging policy;
6. executes through the canonical durable device-operation path;
7. verifies expected outcomes and health;
8. commits or rolls back;
9. records the complete audit chain.

Safety Kernel

Non-LLM authority boundary. Possible results include ALLOW, DENY, REQUIRE_APPROVAL, REQUIRE_CANARY, and REQUIRE_STRONGER_VERIFICATION.

Sentinel must never be able to modify:

• its own authority/policy roots;
• trust roots/device identity;
• invariant implementation;
• operation executor/rollback machinery;
• Tool Registry implementation;
• skill DSL interpreter.

Dependency

Do not enable broad Sentinel mutation until the controller has a single canonical mutation path, durable generation-bound operations, immutable preview/apply binding, canary rollout, capability-aware rendering, and management-path-aware verification.