Release STREAMLINE v1.0.0: P1-P11 pipeline, multiclass/regression support, new features and overhaul

[raptor419]

Release STREAMLINE v1.0.0: P1-P11 pipeline, multiclass/regression support, config runs, and reporting overhaul

This PR promotes the current v3 branch into main as the STREAMLINE v1.0.0 release. Compared with the current main/v0.3.4 release line, v1.0.0 is a major release that reorganizes, expands, and modernizes STREAMLINE while preserving the goal of transparent end-to-end AutoML for tabular data.

Major Changes From v0.3.4/prev main

Architecture And Run Workflow

• Refactored STREAMLINE into explicit P1-P11 phase modules: data processing, imputation/scaling/balancing, feature learning, feature importance, feature selection, modeling, classification ensembles, summary statistics, dataset comparison, replication, and reporting.
• Added config-driven full-pipeline execution with .cfg files, while keeping phase-by-phase command-line and notebook workflows available.
• Standardized parameter names across config files, command-line arguments, notebooks, and saved run-command metadata.
• Added run-command pickle support so repeated phase calls can reuse the parameters that were actually used, while allowing users to override or ignore saved values when needed.
• Added Parallel local multiprocessing-style execution in addition to Serial, local Dask through Local, and supported cluster submission modes.
• Reorganized code around registries so users can add or swap methods more cleanly across phases, including Phase 2 imputers/scalers, Phase 3 learners, Phase 4 feature-importance methods, Phase 5 selectors, Phase 6 models, and Phase 7 ensembles.

Binary, Multiclass, And Regression Support

• Extended STREAMLINE beyond the earlier primarily binary-classification workflow to support binary classification, multiclass classification, and regression as first-class task types.
• Added UCI-based demo datasets and matching held-out replication splits for binary classification, multiclass classification, and regression tutorials and tests.
• Updated P1 task handling so user-specified outcome_type is respected instead of being re-inferred only from the number of outcome values.
• Updated P6 model loading and evaluation around task-specific model registries for binary, multiclass, and regression runs.
• Added multiclass evaluation support, including macro/micro metrics and multiclass ROC/PR curve summaries where applicable.
• Added regression evaluation support, including regression metrics such as explained variance, Pearson correlation, MAE, MSE, median absolute error, max error, residual outputs, and actual-vs-predicted style reporting.
• Updated reports and summary statistics so metric names, no-skill baselines, curves, and plots are task-aware instead of assuming binary classification.
• Clarified that P7 ensemble modeling is currently classification-only, so regression workflows proceed from P6 modeling to P8 summary statistics.

Data Processing, Feature Types, And Preprocessing

• Added clearer feature-type handling for categorical and quantitative features, including optional user-supplied feature type files and inferred feature types.
• Added optional SMOTE/SMOTENC balancing after Phase 2 imputation and scaling, with SMOTENC used when categorical features are present.
• Added support for bypassing one-hot encoding when using models that can handle categorical features natively.
• Added explicit native-categorical model controls so unsupported models are rejected when one-hot encoding is disabled instead of silently receiving an incompatible representation.
• Kept preprocessing extensible through the Phase 2 registry, so users can add scalers such as MaxAbsScaler or custom imputers without changing the rest of the pipeline.

Feature Learning, Importance, And Selection

• Added P3 feature learning as a dedicated phase, including PCA-style learned feature outputs and manifests.
• Updated replication handling to use the saved training workflow artifacts and feature manifests more consistently.
• Updated Phase 4 so feature-importance methods write scores without mutating shared CV datasets, avoiding order-dependent outputs and parallel race risks.
• Updated scikit-rebate integration for the current package behavior, including passing STREAMLINE's categorical feature indexes to ReBATE methods.
• Added MultiSWRFDB and MultiSWRFDB* feature-importance methods.
• Updated default feature-importance/selection behavior around mutual information, MultiSWRFDB, and MultiSWRFDB*.
• Improved Phase 5 feature selection so rankings can be combined across all feature-importance methods that were run, not only a fixed pair of methods.
• Added or restored instance_subset support for expensive feature-importance methods so large runs can be controlled from config/CLI parameters.

Modeling And Native Categorical Algorithms

• Reworked Phase 6 modeling around clearer dataset/model/CV execution units for serial, parallel, Dask, and cluster runs.
• Added Decision Tree support for classification workflows.
• Added or updated HEROS wrappers and optional TabPFN wrappers.
• Added TabPFN token handling: if TABPFN_TOKEN is not set, requested TabPFN models are skipped with a warning while HEROS and other requested models continue.
• Updated ExSTraCS categorical initialization so categorical and continuous attributes can be passed through its supported discrete_attribute_limit and specified_attributes parameters.
• Improved native categorical model handling for compatible algorithms such as CatBoost/CGB and ExSTraCS.
• Added Optuna trial accounting so reports can show how many trials actually ran within the requested n_trials and timeout budgets.
• Standardized model defaults across config, command-line, and notebook run modes, including removal of deprecated/default-only legacy methods where appropriate.

Reporting, Replication, And Outputs

• Added standard and replication PDF report improvements, including clearer first-page summaries, experiment names, resolved/default parameter display, and report-mode-specific text.
• Added dedicated replication report naming and clearer replication report content so replication outputs are not confused with standard CV reports.
• Improved report language around categorical handling, scaling/imputation, feature learning, feature selection, and metric interpretation.
• Added feature-learning and feature-selection summary tables modeled after the data-processing/feature-engineering summary style.
• Improved feature-importance figure layout, including more square plots and clearer ordering emphasis.
• Updated metric highlighting to use shading rather than only bold text.
• Added no-skill ROC/PR legend notes and label-aware baseline handling for multiclass or non-stratified/random CV settings.
• Improved output organization and report data JSON so reports are easier to debug and regenerate.

Notebooks, Documentation, Tests, And Release Readiness

• Updated the Google Colab notebook and local Jupyter notebook for the v1.0.0 workflow, including parameter blocks for binary, multiclass, regression, and custom dataset runs.
• Rebuilt the documentation website around the v1.0.0 workflow as the new main documentation site.
• Added Sphinx/autodoc documentation build support.
• Added GitHub pytest workflows for Python 3.10, 3.11, and 3.12. Python 3.9 was skipped because TabPFN does not support it.
• Added optional TabPFN-specific pytest coverage for no-token skip behavior and token-gated wrapper fitting.
• Added macOS installation guidance for conda-forge compiled dependencies such as Graphviz, WeasyPrint/Cairo/Pango, LightGBM, XGBoost, and CatBoost.
• Removed or de-emphasized old main-era documentation paths and files that are no longer part of the maintained v1.0.0 workflow.

Branch / Release Notes

• This PR merges the current v1.0.0 branch into main.
• The release should be tagged as v1.0.0 after merge.
• The current main state has been preserved separately as the v0.3.4 branch for the previous tested/stable v0.3.4 release line.
• Branch names are intentionally not being changed in this PR; they can be renamed or reorganized manually after the merge if needed.