Add anti-self-bias evaluation suite

[sparklkt]

What changed

This PR adds the v1.2 anti-self-bias evaluation suite as an independent black-box benchmark path for innerbrain.

Key additions:

• evals/anti_self_bias_dataset.json with 35 base tasks across 7 categories:
  • raw reasoning
  • adversarial instruction
  • distractor-rich
  • long-horizon planning
  • hidden rule
  • cross-domain transfer
  • calibration and uncertainty
• automatic clean, misleading, and overloaded variants for every task
• external baseline runners in innerbrain/evals/baseline_runner.py
  • simple_prompt
  • cot_prompt
  • reflection_prompt
  • debate_prompt
  • rule_heuristic_solver
• independent evaluator in innerbrain/evals/evaluator.py
• anti-bias orchestration and rendering in innerbrain/evals/anti_self_bias_suite.py
• CLI commands:
  • innerbrain anti-bias run
  • innerbrain anti-bias report
  • innerbrain anti-bias compare
• extended anti-bias metrics in innerbrain/evals/metrics.py, including ASBI
• explicit external failure typing for:
  • false positive
  • false negative
  • value collapse
  • attention hijack
  • unstable decision

Why it changed

The goal is to test whether InnerBrain-Factor is improving on independent tasks, or only improving on benchmarks that already align with its own internal structure.

This suite keeps scoring external to the core pipeline:

• benchmark input is plain text
• benchmark output is a plain text answer
• scoring is done by an external evaluator
• baselines do not reuse InnerBrain scoring logic

Validation

• pytest -> 58 passed in 0.59s
• innerbrain anti-bias run --dataset-path evals/anti_self_bias_dataset.json --output-path /tmp/anti_bias_full.json --report-path /tmp/anti_bias_full.md --internal-benchmark-reference 1.0

Bundled anti-bias snapshot from the local deterministic run:

• rule_heuristic_solver external performance: 0.97
• reflection_prompt external performance: 0.96
• cot_prompt external performance: 0.96
• debate_prompt external performance: 0.79
• innerbrain_black_box external performance: 0.64
• simple_prompt external performance: 0.35
• ASBI: 0.27

Safety / scope

This PR remains within the project boundaries:

• offline
• deterministic and rule-based
• no LLM
• no network reasoning
• no real external actions
• no bypass of human judgment boundaries

[sparklkt]

Superseded by #12. This draft branch was opened from an older linear development branch and pulled in previously merged history, so the diff was much larger than the actual v1.2 scope. Please review #12 instead for the clean main-based v1.2 change set.