Add evidence-based benchmark regression limits

[Widthdom]

CLOSES #225

Summary

Replace the single 200% BenchmarkDotNet regression threshold with an evidence-based, per-benchmark-group policy and a repository-owned comparison script.

• compare the current BenchmarkDotNet mean with the median of compatible trusted gh-benchmarks history
• distinguish PASS, WARNING, FAIL, and WARMUP in the GitHub Actions job summary
• publish a failed run as a reviewed baseline only through an explicit main-only workflow-dispatch option
• keep English and Japanese performance, developer, and testing documentation aligned

Acceptance criteria

1. Measure normal variance per group: documented seven compatible hosted Ubuntu runs and each group's observed maximum slowdown (4.7% to 42.7%).
2. Document rationale: warning limits round the observed maximum up to 5 percentage points and add a 5-point guard band; failure limits are twice the warning limits.
3. Detect meaningful regressions: failure limits are 20%, 40%, 80%, 40%, and 100%, all substantially below the retired 200% threshold.
4. Avoid noise-driven failures: use the median of up to ten compatible runs, require five samples, and fingerprint benchmark definitions plus global.json.
5. Report warning/failure results: write a per-benchmark Actions summary table and emit ::warning / ::error annotations.
6. Define reset/update process: document fingerprint-based warmup, baseline-revision bumps, trusted-main publication, and runner-change remeasurement.
7. Test boundaries: cover exact warning/failure boundaries, intentional publication, definition/revision/base filtering, unmatched benchmarks, and rejection of 200% policies.

Validation

• python3 -m unittest discover -s scripts/tests -p 'test_*.py' -v — 13 passed
• dotnet test FolderDiffIL4DotNet.Tests/FolderDiffIL4DotNet.Tests.csproj -c Release — 2,044 passed, 1 expected environment-dependent skip
• dotnet format FolderDiffIL4DotNet.sln --verify-no-changes --no-restore --verbosity minimal
• parsed .github/workflows/benchmark-regression.yml with Ruby YAML
• validated benchmark-regression-policy.json with jq
• ran the new gate against the latest hosted BenchmarkCombinedReport artifact and published gh-benchmarks history — all 12 benchmarks passed
• git diff --check