Add transparent adversarial Benchmark v1

[LE0-Lin]

Why

AgentConfigScore had deterministic tests and a three-repository smoke benchmark, but no labeled mutation suite that quantified both expected detections and known misses. This made it too easy to over-interpret A 100 or a small set of demos.

What changed

• add an offline labeled corpus expanded across six instruction formats
• evaluate 74 deterministic contract cases, including 49 expected detections and 26 clean negative controls
• add eight explicitly labeled, non-gating semantic and unmodeled danger challenges
• report exact-case accuracy, precision, recall, F1, clean controls, and per-category results in JSON and Markdown
• commit the generated report and test that it stays synchronized with scanner behavior
• validate fixture paths so benchmark cases cannot write outside their temporary repository
• document why closed-suite 100% metrics are not real-world accuracy estimates

Current result

• deterministic contract: 74/74 exact matches
• expected detections: 49, with 0 missing and 0 unexpected rule IDs
• clean controls: 26/26
• open challenge set: 0/8 detected and clearly reported

The challenge tier intentionally does not control the exit code. It is a visible research backlog, not a hidden failure or an inflated headline metric.

Verification

• 122 unit and contract tests pass locally
• repository self-scan: A 100, zero findings
• pinned real-repository benchmark: all three repositories still match reviewed expectations
• benchmark runs fully offline and uses no LLM/API tokens