[deferred] BENCH-001: Direct-frontier comparison benchmarks (children of #262)

[mrnicholasbcarter-code]

Scope (minimum v1)

Deterministic benchmark harness runs the same published task via DIRECT API model vs Verdict route, reports cost / latency / completion / regression / failover. Reproducible when seeded.

Do not duplicate

This issue surfaces the existing benchmarking.py + reproducible_benchmarks.py modules (already shipped) and the PR #22 final V1-007 ticket.

Deferred (per audit, not for V1)

• Multi-model tournament scoring
• Provider-rotation benchmarks
• GNN evaluation harness

Refs: evidence/V1_READINESS_AUDIT_2026-08-03.md #239