feature: add patching test result type checking (#2971)

* feature: add patching test result type checking

* fix: add cpnp and dpctl check

* fix: comments

* fix: comments

* fix: comments

* fix: add transform test

* fix: transform test

* fix:comments

* fix: versionb

* feature: add dtype checking

* fix: comments

* Address review: predict dtype vs y, skip fp16 dtype check, forest apply exceptions

* Add dtype check exceptions for KMeans predict and SVM estimators

* Add ElasticNet/Lasso predict and path to _DTYPE_CHECK_SKIP in test_patching

* Add more estimator-method pairs to _DTYPE_CHECK_SKIP

Add LinearRegression, Ridge, IncrementalRidge, KNeighborsRegressor,
ExtraTreesRegressor, RandomForestRegressor predict methods and
LogisticRegression decision_function, KNeighborsClassifier predict_proba
to dtype check skip set - these return float64 for integer/float32 inputs.

* Add fitted attribute checks and automate dtype/type skip logic

- Add _check_fitted_attributes() to verify fitted attributes preserve
  input array namespace and dtype (mirrors sklearn's new array API checks)
- Automate regressor predict skip via is_regressor()
- Automate clusterer predict/fit_predict skip via is_clusterer()
- Automate SVM decision_function and fitted attr skips via BaseLibSVM
- Automate tree apply skip via method name
- Reduce hardcoded skip entries from ~70 to ~12

* Add missing fitted attr skip entries for CI failures

* Add LogisticRegression.n_iter_ to FITTED_ATTR_NUMPY_OK

* Remove n_iter_ from _INTEGER_FITTED_ATTRS (no-op, dtype check already skips non-float)

* Add compat fallback for is_clusterer/is_regressor (sklearn < 1.6)

* Add PCA.singular_values_ to FITTED_ATTR_NUMPY_OK

* Add PCA.explained_variance_ratio_ to FITTED_ATTR_NUMPY_OK

* Address reviewer feedback: use sklearn_check_version, update PCA comment

- Replace try/except ImportError with sklearn_check_version('1.6') for
  is_clusterer/is_regressor imports
- Keep PCA singular_values_ and explained_variance_ratio_ in skip set:
  array_api path produces correct types, but dpctl/dpnp path still
  returns numpy fitted attrs. Updated comment to clarify.

* fix: use get_namespace instead of __sycl_usm_array_interface__ for device detection

* fix: remove PCA fitted attrs from skip list - they preserve type correctly

* fix: update comment for _FITTED_ATTR_NUMPY_OK_NON_NUMPY

* fix: remove ElasticNet/Lasso/TSNE from skip list - now fall back to sklearn after #2996

* fix: use standard array API .device for device alignment check

* fix: use standard .device in _check_fitted_attributes device check too

* fix: stricter device and dtype checks per reviewer feedback

- Device check: assert res has .device when X does (catches non-array output)
- Device comparison: use direct == instead of str()
- Dtype check: remove 'float in res.dtype' guard to catch wrong dtypes

* fix: skip dtype check for sparse results (decision_path returns int64 indices)

* fix: skip dtype check for integer structural outputs (decision_path n_nodes_ptr)

* fix: add decision_path to dtype skip list instead of gating on res dtype

decision_path returns (sparse_matrix, n_nodes_ptr) — n_nodes_ptr is
inherently integer and should not match X's float dtype.

* fix: simplify assert messages

* fix: skip device check for attrs accepted as numpy in skip lists

* fix: also skip device check on sklearn fallback

* fix: use continue instead of flags for early exit, remove duplicate dtype skips

* fix: add PCA fitted attrs back to skip list - numpy on dpnp/dpctl CPU path

* feat: add sparse class checks for outputs and fitted attributes

* fix: only skip NON_NUMPY attrs when array_api_dispatch is off, add sparse TODO

* fix: remove PCA and IncrementalEmpiricalCovariance from output skip list after #3021

* fix: skip PCA.score_samples and covariance.mahalanobis output check when array_api_dispatch is off

* fix: add IncrementalEmpiricalCovariance.mahalanobis to dtype skip list

* fix: add output conversion to support_sycl_format, remove skip list

support_sycl_format now converts output back to dpnp/dpctl matching
support_input_format pattern. Removes _NUMPY_OUTPUT_OK_NO_DISPATCH
skip list since PCA.score_samples and covariance.mahalanobis now
return correct types.

* Revert "fix: add output conversion to support_sycl_format, remove skip list"

This reverts commit 46d1cce.

* fix: enable array_api_dispatch for dpnp/dpctl output checks, skip dpctl linalg gaps

* fix: address review - remove caplog.clear, move imports to top, update comments

* fix: merge main, update sparse checks after #3029, move RF/ET classes_ to common skip list

* fix: add missing get_namespace import

* fix: replace hardcoded classes_ entries with programmatic check by attr name

* fix: update comment for public fitted attr check

* fix: check public attrs as starts with letter and ends with _

* fix: remove redundant classes_ entries from skip list

* fix: update comments - SVM/clusterer attrs are skipped, not handled

* fix: narrow SVM skip - check type on CPU, skip GPU/device/dtype

* fix: update comments for skip lists and SVM checks

* fix: use queue.sycl_device.is_gpu instead of string parsing for GPU check

* fix: change caplog to INFO level, remove SVM GPU skip - fell_back now detects fallback

* fix: remove probA_/probB_ from type skip - they are dpnp not numpy on CPU

* fix: remove unused None defaults for est and method in _check_output_type

* refactor: extract helper functions, rename skip lists for clarity

* docs: shorten comments

* docs: shorten output skip list comments

* refactor: move imports to top, update docstring with call sequence

* refactor: rename data_input to y for clarity

* refactor: remove defensive result is None check

* feat: add array_api output type check for pairwise_distances, restore est=None

* docs: update _check_output_type docstring with type/device/dtype checks

* fix: filter caplog records by sklearnex logger to avoid extra INFO messages

* feat: add must-be-array check for fitted attrs like n_iter_, coef_

* fix: only skip classes_ type check when array_api_dispatch is off

* fix: add allow_sklearn_fallback marker - pairwise_distances falls back for int dtypes

* fix: revert to WARNING level, remove allow_sklearn_fallback, add SVM GPU skip back

* docs: fix docstring - classes_ only skips without dispatch

* fix: remove array_api from pairwise test - returns numpy on CI Python 3.13

* fix: skip second pass when estimator doesn't support GPU for this data, remove SVM GPU skip from _should_skip_all

* docs: remove SVM GPU from docstring skip list

* fix: rename helpers, merge main

* fix: change _dpctl_has_linalg default to False

* fix: simplify dpctl skip - always skip, linalg will never exist

* fix: detect decision_path programmatically instead of listing each estimator

* fix: remove regressor predict dtype skip - regressors preserve dtype

* fix: skip regressor predict dtype only for integer inputs, not float

* fix: skip regressor predict dtype for integer inputs only

* fix: remove n_iter_ from must-be-array - scalar for KMeans per sklearn docs

* fix: rename input_type to input_type_y for clarity

* fix: remove _SCALAR_METHODS, rely on np.isscalar in loop

* fix: remove 0-d array check - never triggered

* fix: rename _INTEGER_FITTED_ATTRS to _DTYPE_SKIP_ATTRS

* fix: remove _ATTR_SKIP_DTYPE_BY_NAME - redundant with is_clusterer/BaseLibSVM/classes_ inline checks

* refactor: unify 7 skip lists into 1 dict + 1 set + 1 helper

* fix: swap return order to result, X, y for conventional ordering

* fix: revert pairwise result to _ - not used

* fix: add LogisticRegression.classes_ to _SKIP - no array API support yet

* fix: update dpctl skip message

* feat: add pickle serialization check for dpnp/dpctl without dispatch

* fix: remove defensive 0-d and isscalar checks from fitted attr filter - no estimator returns these

* fix: add back isscalar check - some fitted attrs are scalars with dtype

* fix: add KNeighborsClassifier.classes_ to _SKIP - numpy with dispatch on

* fix: move pickle import to top, remove nosec - matches existing test patterns

* fix: remove PCA no_dispatch entries - first pass doesn't check attrs

* fix: simplify _should_skip - remove unused dispatch_on parameter

* Update requirements-test.txt

Co-authored-by: david-cortes-intel <david.cortes@intel.com>

* fix: move pandas/polars imports back inline

---------

Co-authored-by: david-cortes-intel <david.cortes@intel.com>

Author: yuejiaointel

conda-recipe/meta.yaml

@@ -93,6 +93,7 @@ test:
     - treelite
     - array-api-compat
     - array-api-strict
+    - polars
   source_files:
     - .ci
     - setup.cfg

requirements-test.txt

@@ -19,4 +19,5 @@ treelite==4.7.0
 array-api-compat==1.14.0
 array-api-strict==2.3.1 ; python_version <= '3.11'
 array-api-strict==2.5 ; python_version >= '3.12'
+polars==1.39.3
 packaging==26.0