feat(tooling): aggregate win-rate summaries for result comparisons (external-first)

[christso]

Canonical Plan

This issue body is the source of truth for implementation.

Objective

We provide aggregate win/loss/tie summaries that make comparison results decision-ready at benchmark level.

Why This Matters

• Per-case tables are useful but slow for high-level model selection.
• Teams need compact benchmark-level signals for release and routing decisions.
• Consistent aggregation improves comparability across runs.

Why This Location

• Win-rate aggregation is a reporting/analytics layer concern.
• It does not require changing runtime scoring semantics.
• External-first implementation keeps core stable and composable.

Architecture Boundary

External-first aggregation/reporting layer.

Deliverable Location

• Primary location (in-repo tooling path): examples/features/benchmark-tooling/
• Script location: examples/features/benchmark-tooling/scripts/ (win-rate aggregation utility)
• Usage/docs: examples/features/benchmark-tooling/README.md

Design Latitude

We decide representation details (overall rates, per-metric rates, optional uncertainty) while keeping outputs script-friendly.

Acceptance Signals

• We produce overall win/loss/tie rates from aligned comparisons.
• We support per-metric breakdown when metric-level data exists.
• We make tie policy explicit and documented.

Non-Goals

• No new core evaluator behavior.
• No forced changes to existing comparison semantics.