Refactor public estimator constructors for sklearn <=1.2 clone compatibility

[TheHiddenObserver]

Context

PR #79's Gate B found that sklearn.base.clone() under scikit-learn 1.2.2 fails for 26 public estimators because constructors normalize or defensively copy public parameters. The same failures reproduce on base SHA a4879fb, so this is pre-existing compatibility debt rather than a PR #79 regression.

The current regression test is version-limited and marked strict=True XFAIL for scikit-learn <1.3.

Goal

Restore legacy sklearn clone compatibility without weakening validation or changing normalized runtime behavior.

Required design

• preserve each public constructor argument exactly on the public attribute expected by get_params(deep=False);
• place canonicalized/validated values in private attributes, for example solver plus _solver_normalized;
• avoid str(), .lower(), bool(), int(), float(), list(), dict(), or np.asarray() replacement of public constructor attributes inside __init__;
• ensure set_params() refreshes private normalized state;
• retain __sklearn_clone__ for newer sklearn while also satisfying the legacy constructor identity check;
• centralize choice normalization and alias handling in shared helpers.

Work plan

1. Generate a structured mismatch report for all public estimators and parameters.
2. Group failures by normalization/copy pattern.
3. Refactor shared bases and validators first.
4. Update estimator families in small batches.
5. Test clone, nested clone, pipelines, grid search, mutable parameters, and absence of fitted state.

Compatibility matrix

• scikit-learn 1.2.2;
• scikit-learn 1.3.x;
• latest supported scikit-learn.

Acceptance criteria

• test_all_default_public_estimators_clone passes without XFAIL on all supported sklearn versions;
• get_params/set_params semantics remain stable;
• normalized estimator behavior and error validation remain unchanged;
• complete CPU and physical-GPU suites pass.

Relationship

Follow-up to PR #79. This is non-blocking for the GPU correctness and finite-input paths validated there.

[TheHiddenObserver]

Resolved by PR #87 and included in the 0.2.4 release tree.

Closure evidence:

• public constructor arguments are preserved exactly for get_params(deep=False) and the scikit-learn <=1.2 constructor identity check;
• normalized runtime values are retained in private state;
• set_params() reconstructs estimator state transactionally, refreshes normalized/private attributes, supports nested parameters, and clears fitted state;
• __sklearn_clone__ remains available for newer scikit-learn and recursively clones estimator-valued parameters;
• the prior strict XFAIL for scikit-learn <1.3 was removed from test_all_default_public_estimators_clone;
• compatibility CI runs the full default-public-estimator clone test on scikit-learn 1.2.2, 1.3.2, and the latest supported release;
• additional maintained coverage validates fitted-clone isolation, mutable parameter identity, nested set_params, pipelines, and GridSearchCV;
• the complete CPU and exact-head physical-GPU suites passed.

Implementation merge: PR #87, commit 249456610c9d8ab62e9f2749e1fc7a769250b08a.
Release merge: PR #88, commit 0aeeb95b60e3e274053b8f1b6427ae50c8eec015.

The legacy clone-compatibility matrix and the requested constructor/get_params/set_params contracts are now green.