fix: harden runtime and estimator contracts

.github/workflows/maintenance-compatibility.yml

@@ -0,0 +1,45 @@
+name: Maintenance compatibility
+
+on:
+  push:
+    branches: [master]
+  pull_request:
+    branches: [master]
+
+permissions:
+  contents: read
+
+jobs:
+  sklearn-compatibility:
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        sklearn_spec:
+          - scikit-learn==1.2.2
+          - scikit-learn==1.3.2
+          - scikit-learn
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.11"
+
+      - name: Install compatibility environment
+        run: |
+          python -m pip install --upgrade pip
+          python -m pip install "numpy<2" scipy pytest packaging pandas patsy "${{ matrix.sklearn_spec }}"
+          python -m pip install -e . --no-deps
+
+      - name: Validate maintenance benchmark syntax
+        run: |
+          python -m py_compile dev/benchmarks/benchmark_torch_compile_maintenance.py
+
+      - name: Run maintenance regressions
+        run: |
+          python -m pytest \
+            dev/tests/test_maintenance_024_025.py \
+            dev/tests/test_legacy_sklearn_integration.py \
+            dev/tests/test_second_full_review.py::TestEstimatorCloneAndFeatureSelectionBackend::test_all_default_public_estimators_clone \
+            -q --tb=short

.gitignore

@@ -103,6 +103,11 @@ dev/tests/verify_*.py
 dev/tests/REMOTE_GPU_TEST_GUIDE.md
 dev/tests/TORCH_BACKEND_CV_TESTING.md
 
+
+# Maintained pytest modules must always be visible to git.  Move manual,
+# remote, benchmark, or exploratory scripts to dev/manual instead.
+!dev/tests/test_*.py
+
 # Build artifacts
 build/
 dist/

CHANGELOG.md

@@ -1,7 +1,79 @@
 # Changelog
 
+- Removed universal ElasticNet backend thresholds, coefficient tolerances, and fixed speedup claims that were not established for the current exact-head environment; the model guide now requires workload-specific benchmarking and dtype/solver-specific validation.
+
+- Corrected ElasticNet/Ridge scaling documentation and added a regression test confirming that `ElasticNet(alpha, l1_ratio=0)` matches `Ridge(alpha)` under the shared average-loss convention.
+
+- Reconciled the ElasticNet API documentation with the implementation by correcting constructor defaults, removing nonexistent parameters, and replacing stale strict/approx guidance with the actual FISTA and post-fit inference semantics.
+
+- Completed the public ElasticNet inference contract: the standalone wrapper now exposes and forwards inference options, and ElasticNetCV honors `compute_inference=True` on its final full-data refit with NumPy/CuPy/Torch matrix tests.
+
+- Integrated transactional CV reset with the shared public finite-input guard, so NaN/Inf refit attempts invalidate stale RidgeCV, ElasticNetCV, LogisticRegressionCV, and unified penalized-CV state before validation raises.
+
+- Made dedicated RidgeCV, ElasticNetCV, and LogisticRegressionCV refits failure-safe: every fit attempt clears stale fitted state, and CV selections are published only after the final model refit succeeds.
+
+- Pinned AUTO-mode RidgeCV, ElasticNetCV, and LogisticRegressionCV final refits to the backend selected during CV, preventing silent Torch/CuPy backend drift after parameter selection.
+
+- Preserved `device='auto'` through public RidgeCV, ElasticNetCV, and LogisticRegressionCV dispatch so GPU-resident inputs retain their owning backend; LogisticRegressionCV now validates 0/1 responses without a full GPU-to-CPU copy.
+
+- Logistic and ElasticNet CV default regularization grids now use analytic weights and satisfy integer-weight row-replication equivalence; CV GPU-array device inspection no longer masks runtime failures.
+
+- Dedicated Ridge, ElasticNet, and Logistic CV routines now preserve explicit Torch versus CuPy backend requests, normalize Device enum values consistently, and validate analytic weights before grid generation or degenerate returns.
+
+- Corrected analytic-weight LogisticRegression IRLS across NumPy, CuPy, and Torch: weights now enter WLS curvature rather than the working-response denominator, and weighted likelihood/inference use the same objective. Narrowed penalized-CV alpha-grid and exact CuPy Ridge fallbacks so programming, CUDA OOM, and device errors propagate.
+
+- Completed penalized-CV fallback hardening: optional Lipschitz recovery now recognizes NumPy/CuPy/Torch rank failures consistently, while alpha-grid estimation no longer hides memory or GPU infrastructure failures.
+
+- Preserved the declared validation objective in penalized CV: non-Gaussian losses no longer silently fall back to MSE, weighted squared-error fallback retains validation weights, and GPU infrastructure failures propagate through layered CV fallbacks.
+
+- Narrowed GPU linear-algebra fallbacks so only genuine rank/definiteness failures use least-squares, pseudo-inverse, ridge, or zero-block recovery; CUDA OOM, device, index, and programming errors now propagate.
+
 All notable changes to statgpu are documented here, organized by release and date.
 
+## Unreleased — maintenance hardening
+
+- Removed the package-initialization cycle between `statgpu.glm_core` and the Cox loss export by lazily exposing `CoxPartialLikelihoodLoss`; GLM internals can now be imported first in a fresh interpreter.
+- Removed the over-broad Armijo `out of range` numerical marker so index and device programming errors propagate instead of being mistaken for recoverable trial-point domain failures.
+- Made proximal-Newton Armijo backtracking treat recognized numeric-domain ValueError trials consistently with Newton while still propagating input-contract and infrastructure failures.
+- Narrowed the shared backend linear-system fallback to genuine rank failures; CUDA OOM, device, and unrelated RuntimeError failures now propagate instead of being silently retried with least squares.
+- Made shared NumPy zero/conversion helpers honor a floating reference array dtype, matching the existing CuPy/Torch backend contract while retaining float64 defaults for integer references.
+- Normalized FISTA/FISTA-BB warm starts to the preprocessed design and converted smooth proximal-Newton sample weights to the active backend, device, and dtype before loss evaluation.
+- Narrowed Newton-family Armijo trial exception handling to expected numeric-domain failures so CUDA OOM, device, and infrastructure errors remain visible to callers.
+- Preserved backend RuntimeError failures (including CUDA OOM/device errors) during solver sample-weight validation instead of rewriting them as ordinary invalid-input ValueError exceptions.
+- Aligned the executable loss/penalty/solver matrix with the maintained compatibility contract: Elastic Net precision is tested through FISTA, while smooth solvers are tested to reject it explicitly.
+- Smooth Newton/L-BFGS solvers now reject Elastic Net and other non-smooth penalties before preprocessing instead of silently omitting their non-smooth objective component.
+- Normalized Newton, proximal-Newton, L-BFGS, L-BFGS-B, and ADMM warm starts onto the preprocessed design backend, device, and dtype; added physical Torch/CuPy regression entry points.
+- Removed the incorrect Euclidean-prox Newton shortcut that duplicated smooth penalty terms and solved the wrong non-smooth objective. Smooth L2/no-penalty requests retain Newton updates; non-smooth requests now explicitly use FISTA, and FISTA-LLA requires a future metric-prox capability.
+- Completed ADMM's legitimate Cholesky-to-iterative fallback and kept L-BFGS-B directions/bounds feasible and backend-native.
+- Hardened adjacent Newton, proximal-Newton, ADMM, FISTA-BB, L-BFGS, and L-BFGS-B contracts: validate weights before curvature work, only downgrade true singular systems, preserve dtype/device for proximal Newton and CuPy bounds, and use the correct squared-gradient Armijo slope.
+- Kept direct solver and penalized-CV sample-weight checks backend-native, validated weights before weighted Lipschitz operations, rejected overflowing weight totals, and made HC1 analytic-weight inference invariant to global weight rescaling.
+- Fixed Issue #45 by routing statgpu-owned Torch compilation through a
+  centralized policy that avoids CUDA Graph lifecycle hazards for iterative
+  solvers; compile decisions are observable, and only the known lifecycle
+  failure falls back to eager execution. Performance comparison with
+  `reduce-overhead` remains explicitly deferred.
+- Addressed Issue #81 with backend-native finite-value validation at public
+  estimator boundaries without full GPU-array transfers.
+- Aligned formula sample weights after missing-row filtering across linear,
+  GLM, and penalized estimators; retained Torch/CuPy weights on device; and
+  corrected Gaussian GLM FISTA to use weighted centering and the intended
+  weighted squared-loss intercept.
+- Unified analytic-weight GLM semantics across IRLS ridge scaling, line search,
+  pseudo-loglikelihood, AIC/BIC, dispersion, and sandwich inference; centralized
+  active GLM Torch compilation; narrowed singular-system fallbacks; and added
+  backend-native response-domain validation for every supported GLM family,
+  including penalized estimators and cross-validation entrypoints; scalar
+  GLMs now normalize single-column responses and reject empty, non-real,
+  multicolumn, or length-mismatched responses before solver/fold dispatch;
+  GLM design matrices and analytic weights now share backend-native real,
+  finite, shape, length, and non-empty validation across model, CV, formula,
+  and direct IRLS entrypoints.
+- Addressed Issue #82 by preserving exact raw constructor arguments for
+  legacy scikit-learn clone identity while retaining normalized runtime
+  attributes and `set_params` bookkeeping.
+- Addressed Issue #83 by making maintained `test_*.py` files visible to git,
+  documenting the manual GPU diagnostic boundary, and adding maintained
+  regression coverage.
 ## 0.2.3 — 2026-08-04
 
 ### Added

dev/README.md

@@ -39,6 +39,15 @@ The canonical developer architecture entry point is
 defines the current Cox module ownership, canonical statistical source,
 CV/refit reuse contract, and inference boundaries.
 
+
+## Maintained tests versus manual diagnostics
+
+Every `dev/tests/test_*.py` file is maintained pytest coverage and must be
+tracked, discoverable, side-effect free at import time, and runnable from a
+clean checkout.  Hardware-specific experiments and one-off reproducers
+belong under `dev/manual/gpu_diagnostics/`; `.gitignore` is not a test
+ownership mechanism.
+
 ## Remote GPU Testing
 
 ### Server Access