Phase 3: Success-rate monitoring and self-healing procedures

[AlexMikhalev]

Parent Epic

Part of #692 (Operational Skill Store)

Summary

Track procedure confidence scores over time. When a procedure degrades (success rate drops below threshold), auto-pull it from replay and flag for re-learning. This is the "self-healing" layer that prevents agents from replaying broken workflows.

What Changes

Health tracking in terraphim_agent_evolution

• ProcedureHealthReport -- rolling success rate, total executions, failure patterns
• ProcedureHealth enum: Healthy / Degraded / Broken / NeedsRelearning
• Health transitions emit events for EIDOS prediction-outcome tracking (EIDOS episodic reasoning: prediction-outcome tracking for agent evolution #601)

CLI

terraphim-agent learn health                    # all procedures
terraphim-agent learn health --degraded-only    # only degraded/broken
terraphim-agent learn health --task-type "deploy"

Auto-pull logic

When replay is requested for a Broken procedure:

• Refuse replay, display health report
• Suggest re-capture: "This procedure has failed 6 of last 10 times. Re-capture with: learn capture-success ..."

EIDOS integration

Each health state transition generates a prediction:

• "Procedure X will succeed next time" (for Healthy)
• "Procedure X is likely to fail" (for Degraded)
• Observed outcomes feed back into EIDOS confidence scoring (EIDOS episodic reasoning: prediction-outcome tracking for agent evolution #601)

Affected Crates

• terraphim_agent_evolution (health monitoring)
• terraphim_agent (health CLI, auto-pull in replay)

Dependencies

• Phase 1 (Phase 1: Success capture -- extend CapturedLearning with procedural memory #693) -- procedures exist
• Phase 2 (Phase 2: Procedure replay engine -- skip reasoning for proven workflows #694) -- replay updates success/failure counts
• EIDOS episodic reasoning: prediction-outcome tracking for agent evolution #601 EIDOS episodic reasoning (prediction integration)

Acceptance Criteria

• Health report computed from rolling window of last N executions
• Procedures auto-transition between health states
• Broken procedures refused for replay with clear message
• Health transitions emit EIDOS-compatible predictions
• learn health CLI displays formatted report
• Unit tests for health state transitions