v2.0.0 — v4.3 plugin-only scope trim

BREAKING. Trims Verdict to its plugin-only scope per the 2026-05-03 v4.3 runbook §scope-reset block.

What changed

Removed (16 rubrics, 6 adapters, 7 scripts, 3 integrations/exporters):

• Rubrics: agentic-sast-confidence, browser-agent, clinical-agentic-workflow, code-review-aider-polyglot, eu-ai-act-audit-trail, function-hijacking-robustness, gpt-5-5-differential, model-spec-compliance, owasp-mcp-top-10-beta, project-deal-commerce, routine-execution, ship-readiness, skill-compliance, swe-bench-pro, terminal-bench, tool-output-rewrite
• Adapters: gemini_cli, gemini_deep_research, mlflow_trace, inspect_ai_log, terminal_bench, browser_harness
• Scripts: replay_bfcl_attacks, ship_gate, judge_replay, eu_audit_export, benchmark_gaming_detector, audit_export, bench_gaming_check (+ Berkeley RDI signature pack)
• Integrations: cloudflare_ai_gateway, lighteval_shim. Exporters: openai_evals
• scorecard.adjustments reduced to {deduction, bonus} only

Added:

• CLAUDE.md §v4.3 Scope Contract — pins the in-scope inventory
• tests/test_v43_scope_contract.py — forcing function for any future re-add attempt
• tests/test_skill_md_conformance.py — pins SKILL.md to the Claude Code skill spec
• benchmarks/README.md — regression-gate scope clarifier (NOT a public eval bench)

Changed:

• README.md: rubric count 27 → 11, compliance table gone, architecture tree trimmed
• SKILL.md: allowed-tools: [Read, Write, Edit, Bash] (least-privilege)
• GitHub topic benchmarking removed

Migration

If you depended on a removed rubric, adapter, or scorecard.adjustments.* field, pin to v1.4.2 or fork. None of the removed surfaces will be re-added without a runbook spec change.

See CHANGELOG.md for the full breaking-change list.