Fix verdict-invariant unknown comparisons

[maximalfocus]

Closes #101

Outcome

Unknown pairwise comparison results now remain visible without forcing abstention when bounded counterfactual evaluation proves the final verdict invariant across better, equivalent, and worse. Verdict-changing unknowns retain require-evidence, while enumeration above the documented 81-assignment bound fails closed as material.

Contract coverage

• adds stable comparison-result-unknown-non-decisive rule with field, verdict-set, and evidence trace
• persists non-decisive gaps in canonical records and distinguishes them in executive summaries
• keeps all four public example verdicts; updates the insufficient-evidence representative rule to its genuinely material candidate-test gap
• publishes method 1.3.0 / ruleset 1.9.0 and a governed two-entry architecture-knowledge evolution from immutable v1
• registers method 1.2.0 / ruleset 1.8.0 review cohorts as superseded historical bindings

Verification

• python -m pytest — 1,175 passed
• python -m benchmarks.large_dossier --max-seconds 2.0 — 0.267s
• python -m ruff check .
• python -m ruff format --check .
• python -m mypy src benchmarks tests/test_performance.py
• python -m build
• wheel contains the new method and graph-evolution artifacts
• CI workflow parses as valid YAML