Define FormalPR-Bench seed fixture format

[fraware]

Goal

Create the initial benchmark format for evaluating OVK on agent-authored PR verification tasks.

Scope

• Define benchmark item schema.
• Add 5 seed fixtures based on the current examples.
• Include expected intents, expected backend class, expected evidence status, and expected merge decision.
• Add a simple scoring script placeholder.

Acceptance criteria

• Benchmark fixtures can be loaded and validated.
• Each fixture states the expected verification intent and decision.
• Scoring dimensions include intent recall, backend selection, evidence honesty, counterexample usefulness, and merge decision.

[fraware]

Progress update:

• Added seed cases for self-protection and authorization.
• Added benchmarks/formal_pr_bench/score_seed_cases.py.
• CI now runs the seed scorer after tests.
• Added capability registry, router, and attestation helper tests.

Next benchmark step: add adversarial cases for missing metadata, weak intent selection, unknown solver state, and multi-backend routing.