v5.3: Saved Runs, the explanations and the review's hardening

v5.3 "impressive features", the first part: the full leaderboard and Saved Runs, what a model relies on and why it predicts what it does, and the hardening an outside review asked for: what the evaluation and the explanations can and can't say, remote access, the outside services, files published whole, one web server per data folder, deleting and storage. Project version: 5.3.0.

Test result: 1,864 automated tests pass; ruff, mypy and pip-audit are clean. The gate (record): the review's conditions (1), (2) and (6) checked again on the real server; the full NSL-KDD files through the trainer with the keyboard only, with a cancel, what the model relies on and Why?, and the web and the CLI label KDDTest+ the same, byte for byte; the ML report on the v5.3 code gives the 5.0 numbers (all 1,246 values equal); a fresh install on Windows passes every menu path; every page at 8 sizes in both themes, with no contrast failure. Windows 10 Enterprise LTSC 2021, Python 3.14.3, scikit-learn 1.8.0.

Everything that changed, with the known limits: CHANGELOG.

An older version, kept for the record: the latest is v5.4.