Implement policy-gated action loop v0 with audit hooks

[mdheller]

Goal

Implement the first deployable Agentplane runtime-control loop with explicit policy and audit boundaries.

Scope

Implement minimal runtime structures for:

• action proposal
• policy check
• audit event
• intervention record
• runtime trace

Use Behavior Tree / BDI-inspired structure where practical, but keep implementation scope minimal.

Required concepts

• ActionProposal
• PolicyDecision
• RuntimeTrace
• InterventionOutcomeRecord
• TraitBaseline
• TraitDriftMetric

Runtime rules

• No action executes without a policy decision.
• Every action emits an audit trace.
• Runtime interventions must be recorded.
• Trait drift metrics are observational only in v0.
• Runtime must distinguish research-only evidence from implementation-safe evidence.

Validation

Add smoke tests/examples showing:

• action without policy decision fails
• action with policy approval succeeds
• runtime trace emitted for every action
• intervention record emitted when policy blocks or modifies action

Source substrate

Seed corpus PR: SocioProphet/sociosphere#334

Important limitations

• The misalignment-contagion trigger paper is still only normalized as a preprint seed.
• No autonomous adversarial execution should be implemented in this tranche.
• No self-modifying runtime behavior.
• No external side effects beyond bounded test actions.