[P2] V1-007: Reproducible benchmarks (BENCH-001 minimal)

[mrnicholasbcarter-code]

Acceptance criteria

• BenchmarkRunner runs DIRECT API model vs Verdict route on the same task
• Reports cost / latency / completion / regression / failover
• Deterministic when seeded

Refs: evidence/V1_READINESS_AUDIT_2026-08-03.md