Epic: Ouroboros-inspired agent infrastructure patterns

Start with the existing agent areas named in the issue—supervisor, registry, messaging, evolution, goal_alignment, task_decomposition, kg_agents, and kg_orchestration—and read plans/terraphim-ouroboros-rebuild-assessment.md plus the referenced knowledge files. Work through subissues #596–#602 in dependency order; completion is the set of exit criteria covering guardrails, event logging, budgets, learning capture, verdicts, EIDOS tracking, and KG promotion.

Summary

Absorb operational patterns from joi-lab/ouroboros and vibeship-spark-intelligence into terraphim's existing Rust architecture. Not a rebuild -- selective pattern adoption to close gaps in the agent layer and complete terraphim_agent_evolution.

Context

Ouroboros and Spark Intelligence demonstrate infrastructure patterns that terraphim's agent system (supervisor, registry, messaging, evolution, goal_alignment, task_decomposition, kg_agents, kg_orchestration) currently lacks. The foundation layer (types, automata, rolegraph, persistence, middleware, hooks, MCP) is solid; the agent layer is incomplete.

Ouroboros Patterns

• #596 Constitutional Guardrails (CONSTITUTION.md) -- ~2 hours
• #597 Event Sourcing for Agent Decisions -- ~1 day
• #598 Budget-Aware Multi-Model Routing -- ~4 hours

Spark Intelligence Patterns

• #599 Enhanced Learning Capture (multi-hook pipeline + importance scoring) -- ~1 day
• #600 Dimensional Verdict Scoring / Meta-Ralph (6-dimension judge scores) -- ~4 hours
• #601 EIDOS Episodic Reasoning (prediction-outcome tracking) -- ~1 day
• #602 Confidence-Driven KG Promotion (EIDOS threshold gates KG entry creation) -- ~1 day

Execution Order

#596 Constitutional Guardrails ─────────────────────────────────┐
                                                                 │
#597 Event Sourcing ────┬──── #598 Budget-Aware Routing          │
                        │                                        │
                        ├──── #599 Enhanced Learning Capture     │
                        │         |                              │
                        │         v                              │
                        ├──── #600 Dimensional Verdict Scoring   │
                        │         |                              │
                        │         v                              │
                        └──── #601 EIDOS Episodic Reasoning      │
                                  |                              │
                                  v                              │
                              #602 Confidence-Driven KG Promotion│

#596 and #597 have no prerequisites and can start in parallel.
#602 is the capstone: Event Capture -> Event Store -> EIDOS -> Promotion -> KG Entry -> Aho-Corasick Automata.

Patterns NOT Adopted (and why)

Design Principle

Terraphim's identity is Aho-Corasick + local-first data sovereignty. Agent evolution is not about self-modifying code (Ouroboros) -- it's about self-improving knowledge quality (Spark). The evolution happens in confidence-gated KG promotion, not in source code mutation. The Aho-Corasick engine stays constant; the knowledge graph grows autonomously when confidence thresholds are met.

Reference

• Analysis: plans/terraphim-ouroboros-rebuild-assessment.md (in cto-executive-system)
• KB entries: knowledge/ouroboros-self-modifying-agent.md, knowledge/spark-intelligence-self-evolving-companion.md

Exit Criteria

• Constitutional guardrails enforced by agent_supervisor before execution
• Agent decisions logged as immutable event stream with drift detection
• LLM calls tracked per-task with budget limits and fallback chains
• Learning capture covers all three hook types with importance scoring
• Judge verdicts use 6-dimensional scoring with feedback loops
• EIDOS prediction-outcome loop tracks advisory accuracy
• High-confidence insights auto-promote to KG entries via Aho-Corasick indexing