Changelog¶

All notable changes to this project will be documented in this file.

1.0.0 - 2026-07-23¶

Publish the stable customer-commitment release after the 0.x feature, validation, and hardening line.
Freeze the public scope around native CPU/MPS tree estimators, explicit optional backend diagnostics, project-local real-world dataset caches, and customer-facing environment guidance.
Add final release, known-issue, and performance reports that distinguish current HEAD evidence from historical pre-v2/v3 benchmark baselines.
Keep HIGGS as an explicit large local-file performance-boundary dataset rather than packaging or automatically downloading multi-gigabyte raw data into user environments.
Require fresh PyPI install verification and versioned artifact hashes before closing the release audit.

Harden portable backend behavior so explicit external backend policies warn and continue through the native CPU compatibility path instead of stopping user workflows.
Record both requested and effective portable backend decisions in adapter training summaries.
Add versioned release audits and release pages through 0.5.0 while preserving earlier release history.
Move long-form project documentation into docs/ and AI usage guidance into ai-skills/, leaving README as the root entry point.
Add the 0.5.0 known-issue audit as the handoff gate before any future 1.0.0 commitment.

Add explicit portable-backend diagnostics, optional xgboost, sklearn, and cuda extras, and observable backend selection summaries.
Add the official SHAP integration boundary with mpsboost[shap], native tree export payloads, and a clear semantic-validation gate for TreeExplainer.
Add opt-in real-world acceptance tests for MNIST, Titanic, Adult Income, and HIGGS using project-local ignored caches instead of user-global cache paths.
Add the S18 real-world acceptance report and documentation-site navigation for the report.
Keep native CPU/MPS as the default implementation and correctness oracle; external backends remain explicit, observable, and optional.

Add sklearn-style binary and multiclass classifiers, including native CPU softmax and explicit OvR compatibility.
Add decision tree, random forest, ExtraTrees, and CatBoost-like numeric estimator families on the shared native tree engine.
Add quantile, Poisson, and Tweedie regression objectives with model-format round trips.
Add gain, split-count, permutation, and controlled SHAP-like explanation helpers.
Add CPU-suitable isolation forest and pointwise learning-to-rank estimators with observable backend routing.
Add import-time MPS environment guidance with copy-paste setup commands and MPSBOOST_SKIP_ENV_CHECK=1 for CPU-only workers.
Expand real-world dataset smoke coverage and CPU baseline records for the 0.x release line.

Replace preview mock APIs with a real compiled MPS/Metal backend foundation.
Add native device diagnostics and a real GPU vector smoke kernel.
Build Metal shaders at package build time and bundle the resulting library.
Add deterministic native quantization for strided float32/float64 buffers.
Add compact uint8/uint16 feature-major bins and validated serialization.
Reject invalid values, unsupported layouts, corrupted data, and size overflows.
Add real MPS split-scan and stable partition/compaction kernels.
Add histogram subtraction and layer-aware training hooks.
Add an L1 temporary buffer pool for MPS histogram and hot-path workspaces.
Add S6 end-to-end regressor benchmarks with documented small-data regressions and large-wide wins.
Add versioned L2 cache keys, atomic writes, checksum validation, and damaged-cache invalidation.
Add public cache diagnostics, explicit cache directory creation, and safe cache clearing.
Tighten CI cache boundaries so self-hosted GPU jobs do not write or upload user-level pip caches.