Cross-run compounding via lesson injection and validation (Phase 3)

[AlexMikhalev]

Summary

Enable agents to compound improvements across runs by injecting top-confidence lessons into prompts and validating them through success/failure outcomes.

Approach

• Before each spawn: load agent's LessonsEvolution, inject top lessons into prompt
• After run: validate injected lessons (success increases confidence, failure decreases)
• Add periodic memory consolidation every 100 reconciliation ticks (~50 min)
• Lesson confidence scores create natural selection pressure (AVO "single-lineage sustained evolution" pattern)

Critical Files

• crates/terraphim_orchestrator/src/lib.rs -- add load_prior_context() method
• crates/terraphim_agent_evolution/src/lessons.rs -- use existing validate_lesson() with Evidence
• crates/terraphim_orchestrator/src/lib.rs -- add consolidation call

Acceptance Criteria

• Second+ runs of same agent see prior lessons in prompt
• Confidence scores change after success/failure
• Memory consolidation runs without errors
• cargo test --workspace passes

Dependencies

• Blocked by: Phase 1 + Phase 2
• Related: Phase 1: Success capture -- extend CapturedLearning with procedural memory #693 (success capture), Phase 2: Procedure replay engine -- skip reasoning for proven workflows #694 (procedure replay), Phase 4: Nightly extraction -- automated workflow crystallisation via ADF #696 (automated workflow crystallisation)