feat(tooling): paired significance testing for result-file comparisons (external-first)

[christso]

Canonical Plan

This issue body is the source of truth for implementation.

Objective

We add statistically grounded significance checks for baseline-vs-candidate comparisons of result files.

Why This Matters

• Raw deltas can be misleading when variance is high.
• We need a principled way to distinguish real improvements from noise.
• Better statistical signals make benchmark decisions more trustworthy.

Why This Location

• Significance testing is analysis over result artifacts, not evaluation execution logic.
• Tooling can evolve methods faster than core runtime contracts.
• External-first keeps core focused on primitives and execution.

Architecture Boundary

External-first post-processing (wrapper/tooling), not core evaluator runtime.

Deliverable Location

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

Design Latitude

We choose the MVP method and interface (for example, paired bootstrap or McNemar first), then iterate.

Acceptance Signals

• We support at least one paired significance method on aligned test IDs.
• We output machine-readable statistics plus a clear human-readable verdict.
• We handle missing/unmatched pairs deterministically and transparently.

Non-Goals

• No new built-in evaluator primitives.
• No coupling to runtime scoring/execution pipeline.