Add external benchmark validation

[sparklkt]

Summary

• add external-style benchmark adapters and local benchmark task packs for ARC-style, reasoning, safety, long-horizon, and distraction-robustness validation
• add benchmark runner unification across innerbrain_factor, cot, reflection, multi_agent_debate, and risk_rule
• add benchmark-only ablation toggles for value_field, collision_lanes, attention_sovereignty, and growth_memory
• add richer metrics and typed failure logging for robustness, consistency, long-horizon stability, distraction resistance, safety correctness, over-gating, false positives, false negatives, value collapse, attention hijack, and unstable decisions
• add CLI commands for external-benchmark and external-benchmark-report, plus tests and light documentation updates

Testing

• pytest -> 40 passed in 0.52s
• innerbrain external-benchmark --benchmark-dir evals/external_benchmarks

External Benchmark Snapshot

• innerbrain_factor | 1.00
• innerbrain_factor.growth_memory | 1.00
• innerbrain_factor.value_field | 0.98
• innerbrain_factor.collision_lanes | 0.91
• innerbrain_factor.attention_sovereignty | 0.84
• multi_agent_debate | 0.78
• reflection | 0.76
• risk_rule | 0.75
• cot | 0.70

Ablation Reading

• attention_sovereignty shows the largest drop on this pack, especially on distraction robustness
• collision_lanes materially reduce robustness when disabled
• value_field still matters, but the current pack shows a smaller delta than attention and collision structure
• growth_memory is neutral on this benchmark pack right now, which is useful feedback about where memory is and is not yet pulling weight

Boundary Check

• remains offline and rule-based
• no LLM integration
• no network reasoning in the pipeline
• no real external actions
• human judgment remains required for high-risk boundaries