Add repeated large-repository benchmark samples and percentile summaries

[aliengineering-byte]

The current benchmark is a useful scale smoke test, but one-shot timings are noisy and should not be treated as a performance curve.

Improve the existing benchmark harness so contributors can collect repeated workspace-setup samples and report useful distribution summaries.

Scope:

• add a configurable repeat count without adding a runtime dependency;
• report individual samples plus median and a clearly defined upper percentile when the sample size supports it;
• keep the existing quick benchmark path fast;
• produce stable machine-readable output in addition to concise terminal output;
• document the host, filesystem, Node, Git, file count, and sample count required for comparable reports;
• add focused tests for summary calculations and argument validation.

Do not use the result to claim that one isolation strategy is universally faster. The goal is better measurement quality and reproducibility on Windows, Linux, and macOS.