Stabilize C# allocation regression probes

[Widthdom]

Summary

• Stabilize the C# regex allocation regression probes with warmed, alternating median sampling.
• Allow a fixed 64 KiB measurement-noise margin while retaining deterministic material-regression detection and the existing structural/symbol-parity checks.

Validation

• dotnet build CodeIndex.sln -c Release -p:UseSharedCompilation=false --no-restore (0 warnings, 0 errors)
• SymbolExtractorCSharpRegexProbeTests in Release, three repetitions each on net8.0 and net9.0 (12/12 passed per run)
• Full Release net8.0 suite with -m:1 -nodeReuse:false (12,219 passed, 9 skipped, 0 failed)
• Full Release net9.0 suite with -m:1 -nodeReuse:false (11,730 passed, 435 skipped, 0 failed)
• dotnet format CodeIndex.sln --verify-no-changes --no-restore
• dotnet run --project tools/CodeIndex.Changelog -- check (14 fragments validated)
• Root index and workspace manifest health checks (fresh and complete)
• Codex adversarial review: No blocking/actionable issues found.

Documentation and changelog

• Documented the allocation-probe sampling contract in both English and Japanese in TESTING_GUIDE.md.
• Added the bilingual changelog.d/unreleased/5244.fixed.md fragment.

Follow-up candidates

• None.

Fixes #5244