Add evaluation harness for reasoning scenarios

[sparklkt]

What changed

This PR adds the v0.7 evaluation harness and benchmark suite without expanding the reasoning architecture itself.

• adds structured evaluation models, scenario loading, metrics aggregation, and failure analysis
• adds deterministic baseline runners for cot, reflection, multi_agent_debate, and risk_rule
• adds benchmark scenarios covering offline prototyping, networking, self-modification, privileged execution, creative design, safety-creativity tradeoffs, low-value manipulative input, and long-horizon research
• adds CLI evaluation commands to generate a metrics dashboard, structured JSON output, and Markdown failure reports
• tunes safety-critical attention calibration so high-risk boundary requests surface as deep-attention inputs during evaluation
• adds tests for scenario loading, harness execution, high-risk gating, low-risk over-gating control, and failure report rendering

Why it changed

The next stage of the project needs evidence, not more unchecked architectural growth. This PR creates a repeatable offline benchmark harness so InnerBrain-Factor can be compared against simpler baselines on risk recognition, long-term goal retention, attention stability, creativity-safety balance, evidence-gap detection, human-judgment correctness, and over-gating.

Validation

• /Users/qianjinus/Documents/类脑/innerbrain-factor/.venv/bin/pytest -> 28 passed in 0.10s
• innerbrain eval --scenario-dir evals/scenarios --output-path evals/results/latest.json --failure-report-path evals/results/latest_failures.md

Observed comparison on the current benchmark set:

• innerbrain_factor composite: 1.00
• multi_agent_debate composite: 0.66
• reflection composite: 0.59
• risk_rule composite: 0.58
• cot composite: 0.37

Safety boundary

• remains offline and rule-based
• no LLM integration
• no network participation in reasoning
• no real external actions
• no bypass of human judgment
• evaluation baselines are deterministic local heuristics only; they do not call external models or tools