Add adaptive scheduling and trustworthy benchmark tooling

[polishdataengineer]

What changed

• replace auto = CPU count with adaptive worker selection based on schedulable units, process-visible capacity, duration history, startup cost, and predicted makespan
• add testenix tune / testenix benchmark with safe configuration persistence
• add explicit fail-closed --shard-modules scheduling and trusted collection manifests to avoid the collection-side import on unchanged suites
• report the actual number of non-empty workers used
• add synthetic 100/500/1000/3000 scaling curves and a redaction-safe real-project benchmark harness
• update README, GitHub Pages sources, generated LLM references, changelog, and benchmark workflow
• label the historical 3.15x result with Testenix 0.1.0, 4 workers, and --no-history

Why

The previous defaults could oversubscribe high-core machines, hide module-affinity bottlenecks, import test modules twice, and make benchmark claims without enough provenance. The new paths are adaptive or explicitly opt-in, and publication gates fail closed on stale inventories, mismatched per-test outcomes, ambiguous runner targets, dirty repositories, or unverified runtimes.

Validation

• 461 tests passed locally with the non-editable install (excluding the user's unrelated duplicate test file)
• 76 focused benchmark/tuning tests passed
• GitHub Actions matrix passed on Linux, macOS, and Windows with Python 3.11–3.14
• Ruff check and format check
• mypy: 32 source files, no issues
• pyright: 0 errors
• generated documentation asset check
• Sphinx warning-as-error build
• wheel and sdist build
• clean-wheel CLI/run smoke
• 13-scenario unpublished scaling smoke covering 100/500/1000/3000 tests
• final adversarial multi-agent audit: no remaining blockers
• post-merge CI and GitHub Pages deployment passed on d355a457

Notes

This PR does not publish a new release and does not replace the historical 3.15x baseline with an unreviewed number. The local real-project measurements remain explicitly non-publishable because the target checkout was dirty and this code is unreleased.