Changelog

​

Workflow & Automation

• Reusable Release Prep workflow (.github/workflows/release-prep.yml) that gates release pipelines with tests, typecheck, Zig guard checks, benchmark checks, README benchmark sync checks, and npm-pack smoke validation
• Zig backend smoke example for users: examples/zig-backend-smoke.ts
• Per-kernel tree hot-path regression guard (bench:hotpaths:check) using bench/results/tree-hotpaths-baseline.json
• Release Guard workflow (.github/workflows/release-guard.yml) with tag/version/npm preflight checks and duplicate publish/native-run detection
• Concurrency dedupe guards on publish workflows to avoid duplicate tag releases

​

Machine Learning APIs

• Clustering: KMeans with deterministic randomState, fitPredict, transform, and score
• Clustering: DBSCAN and AgglomerativeClustering for density-based and hierarchical clustering parity
• Decomposition: PCA with explained-variance attributes, fitTransform, transform, and inverseTransform
• Decomposition: TruncatedSVD with deterministic power-iteration components, variance attribution, fitTransform, and inverseTransform
• Decomposition: FastICA with whitening, independent component extraction, and inverse projection support
• Decomposition: NMF with multiplicative-update factorization and reconstruction APIs
• Decomposition: KernelPCA (linear, rbf, poly kernels)
• Calibration: CalibratedClassifierCV with sigmoid and isotonic calibration methods
• Ensemble: VotingClassifier (hard/soft) and StackingClassifier baseline meta-estimators
• Ensemble: VotingRegressor, StackingRegressor, and BaggingClassifier meta-estimator parity additions
• Boosting: AdaBoostClassifier, GradientBoostingClassifier, and GradientBoostingRegressor
• Boosting: HistGradientBoostingClassifier and HistGradientBoostingRegressor
• Multiclass: Support across core classifiers/ensembles (GaussianNB, KNeighborsClassifier, LogisticRegression, SGDClassifier, LinearSVC, DecisionTreeClassifier, RandomForestClassifier, VotingClassifier, StackingClassifier, BaggingClassifier, CalibratedClassifierCV)
• Feature Importance: featureImportances_ API for tree/forest/boosting estimators, including histogram boosting

​

Testing & Fixtures

• sklearn snapshot fixtures at test/fixtures/sklearn-snapshots.json and fixture-based parity tests for calibration/ensemble/decomposition outputs
• Reproducible fixture generation script: scripts/generate-sklearn-fixtures.py

​

Documentation & Benchmarks

• README install docs now include a post-install Zig backend smoke check for DecisionTreeClassifier and RandomForestClassifier
• README benchmark section is now marker-driven and auto-generated from bench/results/heart-ci-latest.json via scripts/sync-benchmark-readme.ts
• bench:snapshot now also runs bench:sync-readme
• Benchmark Snapshot workflow now commits README benchmark updates

​

CI & Testing

• CI benchmark gating now enforces tighter zig/js slowdown limits and README benchmark sync
• CI and release-prep now run tree hot-path microbench + regression checks
• Hot-path predict retention guard for random-forest is relaxed to 0.55 in CI/release-prep to match observed stable baseline variance
• CI now uses fast PR checks and main-only heavy lanes (native matrix, zig-tree smoke, benchmarks) to reduce PR cycle time while preserving release strictness
• New CI parity job runs parity:check and enforces per-family sklearn drift thresholds
• sklearn fixtures and parity checks now include multi-seed drift baselines with fixture-defined threshold tables

​

Native Backend

• Release-native-prebuild workflow now runs only on published releases (plus manual dispatch) to avoid duplicate tag-triggered asset jobs
• Native tree/forest ABI now supports multiclass labels end-to-end in Zig (Uint16 labels, up to 256 classes) and is loaded via ABI version 2

​

Model Improvements

• HistGradientBoostingClassifier / HistGradientBoostingRegressor now support maxDepth, maxLeafNodes, and early-stopping controls (earlyStopping, nIterNoChange, validationFraction, tolerance)

​

Zig Performance Optimizations

• Zig tree codebase split into modules: zig/src/tree/split.zig, zig/src/tree/fit.zig, and zig/src/tree/predict.zig
• RandomForest Zig fit uses dynamic atomic work scheduling across threads instead of static chunking
• RandomForest Zig predict supports threaded row-chunk execution for larger inference batches
• DecisionTree Zig predict includes a SIMD threshold-compare traversal path for larger row batches
• Tree splitter threshold bin cap increased to 128 with exact-threshold fallback on small nodes