[Essoz]

This prepares TrainCheck 0.2.0 from the reviewed main commit 205a687 after #151. PyPI currently publishes 0.1.2. The accumulated inference, tracing, and validation changes warrant a new minor beta release.

The package's four console-entrypoint declarations fail standard importlib.metadata.EntryPoint.load() in fresh interpreters, although pip-generated wrappers happen to work via import side effects. Change them to direct module:function targets: 4/4 standard loads fail before; 4/4 pass after. A typed installed-release validator checks actual distribution metadata, environment location, and independent command loading.

The source archive previously included test files without their fixtures or imported tools. Include those source-only dependencies and the release documentation in the sdist; keep wheel package selection unchanged. Add five negative metadata/location tests, bump the package version, and document reproducible build/publication steps. No training, tracing, or inference algorithm changes are introduced here.

Validation:

• Wheel and source archive build; the wheel is built from the sdist. Both pass twine check --strict.
• 173 regressions pass from the extracted sdist, including a fresh rerun in a fully isolated Python 3.10.16 environment with the recorded tested dependency versions (Torch 2.2.2, pandas 2.2.3, NumPy 1.26.4).
• Clean environment pip check: no broken requirements.
• All four real installed entrypoints load in fresh isolated interpreters. All four installed commands accept --help outside the checkout.
• Required static-analysis tables and new production modules are present. Archive inspection found no credentials, traces, checkpoints, or experiment outputs.
• Configured changed-file hooks, strengthened Mypy over both reusable validator/test modules, and diff checks pass.
• Signed commit: 967070c7e3d421e368de6dde4fd731cab7d9750f.

The known three unrelated repository-wide formatting failures remain outside this change. Installed tests use the recorded runtime versions; this does not establish compatibility with every version allowed by the currently unbounded dependency declarations. Benchmarks/fresh-validation utilities remain source-only.

Release notes retain the original measured performance comparisons and their limitations; this packaging change makes no new speedup claim. The new direct inference comparison and tracing hotspot audit are recorded separately.

Evidence and inspected distributions are retained under outputs/release-20261026/. After this main-targeted PR lands, create a signed v0.2.0 tag, publish the GitHub release, and upload the exact inspected artifacts to PyPI using the maintainer's configured project token. PyPI authentication remains pending.