Summary

• expose ongoing_exploration_period, max_components, and min_elite_samples through Rust config, Python, YAML/CLI documentation, effective study configs, and checkpoints
• preserve the incumbent values when omitted (5, 3, and 1) and keep the existing five positional Python Gmm(...) arguments compatible
• keep post-warmup concurrent asks on Sobol until an empirical GMM fit is installed, without losing the historical route of legacy checkpoints
• enforce the feasible-elite floor for scheduled refits, objective updates, and leaderboard-only recovery; an inadequate replacement leaderboard now invalidates any stale fitted model

This PR adds the controls needed for calibration. It does not adopt dynamic defaults, execute a calibration campaign, rerun paper benchmarks, or change the incumbent 5/3/1 setting values.

Validation

• cargo fmt --all --check
• cargo clippy --locked --workspace --all-features --all-targets -- -D warnings
• cargo test --locked -p hola --tests --all-features (83 active unit + 105 integration tests)
• cargo test --locked --workspace --all-features (full workspace pass before the final restore-path hardening; the affected crate suites were rerun afterward)
• minimal-feature checks for opt_engine and hola
• RUSTDOCFLAGS='-D warnings' cargo doc --locked --workspace --all-features --no-deps
• full Python suite: 430 passed, 4 skipped
• changed Python API suites after the final native rebuild: 88 passed
• Ruff, ty, and documentation examples: 40 passed, 4 skipped

All monitored local runs stayed below 1 GiB peak RSS and reported zero swap operations.