Documentation Changelog¶

Unreleased¶

Python 3.14 support: added the Python :: 3.14 classifier and extended the CI test matrix to 3.12/3.13/3.14 across Ubuntu, macOS, and Windows; requires-python stays >=3.12. The full runtime dependency tree provides cp314 wheels and curve_fit converges correctly on 3.14.
Correctness sweep (codex + agy review): fixed bounds validation in curve_fit input handling, covariance/SVD-rank edge cases, and several orchestration/result bugs.
Fixed common_jax over-aggressive denominator guard that broke rank-deficient TRF solves.
Fixed chunked-fit fallback path and made JIT cache keys include function code identity to prevent closure collisions across distinct models.
Test/CI hygiene: relaxed the too-strict gradient-precision optimality threshold, warmed up before the memory-leak baseline benchmark, and silenced expected non-convergence warnings so make verify is clean.

Multi-subsystem RCA hardening across GUI, core, orchestration, and security/infra.
GUI (Qt/PySide6): fixed a QThread GC crash (strong refs in _pending_threads), replaced the shallow FitWorker snapshot with a deep copy, made autosave writes atomic (tmpfile + rename), corrected bounds/array serialisation, and fixed NaN/Inf JSON export via safe_serialize.
Core: 11 fit()-workflow correctness fixes (dtype handling, CG bias, alpha overflow, theta bound, API compat).
Orchestration: corrected data_preprocessor sigma handling, optimization_selector method selection, covariance_computer SVD rank, and streaming_coordinator memory budgeting.
Security / infra: JIT-safety fix, safe_serialize data-integrity fix, tightened model_validation AST allowlist, GPU-fallback locale handling, and excluded graphify-out/ from codespell/detect-secrets.

Skip sparsity detection when tr_solver is already set (7–8ms per-call savings)
Fix Windows CI: catch AttributeError for POSIX-only os.sysconf in xdist worker cap hook
Align release.yml test step with ci.yml to prevent timeout on publish-triggered test runs
Replace warnings.warn() with logging.debug() in SmartCache for operational fallback paths
Reduce JIT cache clearing interval and add pytest_xdist_auto_num_workers hook for OOM prevention
Align config tested-on version comments with local environment
Update uv.lock for black 26.3.1

Added check_plugin_conflicts() for CUDA plugin dual-install and version-mismatch detection in device.py
Fixed misleading GPU warning header; expanded uninstall command to include CUDA plugin packages
Promoted xxhash, pyyaml, evosax, psutil, and dev/test tools to core dependencies (single uv sync)
Removed extra_requirements: [docs] from .readthedocs.yaml; docs deps covered by core
Added gpu-diagnose Makefile target and 3-step CUDA 13 install with conflict verification
Major architecture.rst overhaul reflecting v0.6.4+ orchestration components and facades
Trimmed ~1500 lines of redundant inline code blocks from tutorials 01–06
Refreshed 18 autodoc RST stubs and added 5 new API stubs (stability.guard, utils.diagnostics, callbacks, core.factories, interfaces)
Clarified mixed-precision fallback log message; added streaming docstring note
Moved deferred scipy.stats imports to module level in GUI export/results pages
Security: cryptography 46.0.4 → 46.0.5 (CVE-2026-26007), pillow 12.1.0 → 12.1.1 (CVE-2026-25990)

Fixed critical GUI startup crashes on macOS (SIGBUS) caused by OpenGL/Metal conflicts
Enhanced JAX platform guards and device detection reliability
Unified logging handlers to prevent file lock issues on Windows
Hardened test suite against environmental variability and timeouts
Fixed GUI cache directory cleanup on shutdown

Updated version references from v0.7.0 to v0.6.4 across all orchestration documentation
Added CurveFit decomposition architecture documentation (v0.6.4 feature)
Updated package structure references to v0.6.4

Migrated to Furo theme for modern documentation styling
Removed all deprecated functionality from NLSQ package
Comprehensive documentation update with AST-based code analysis (94.9% coverage)
Fixed pre-push hook CI interference
Fixed Windows resource module import compatibility

Replaced pyqtdarktheme with Qt 6.5+ built-in color scheme (setColorScheme())
Updated dependency versions: JAX 0.8.2, NumPy 2.4.0, SciPy 1.16.3, ruff 0.14.10, mypy 1.19.1
Added CHANGELOG entries for v0.5.0, v0.5.1, v0.5.2 release history
Reorganized benchmark suite into ci/components/microbench directories

Added nlsq.gui_qt API documentation to Sphinx
Fixed README link to GUI user guide (now points to ReadTheDocs)
Updated README GUI description to reflect pyqtgraph plots (was incorrectly referencing Plotly)
Fixed pre-commit issues: refactored complex functions, added contextlib.suppress, created .secrets.baseline
Added mypy overrides for PySide6, pyqtgraph, and qdarktheme in pyproject.toml

BREAKING: Removed Streamlit GUI in favor of native Qt desktop application
- The nlsq.gui package has been removed entirely
- Use nlsq.gui_qt and nlsq-gui command for the desktop GUI
- Removed gui optional extra from pyproject.toml; use gui_qt instead
- Install with: pip install "nlsq[gui_qt]"
Removed legacy StreamingOptimizer/StreamingConfig and the Adam warmup path from docs; use AdaptiveHybridStreamingOptimizer with HybridStreamingConfig (L-BFGS warmup) for large datasets.
Removed DataGenerator, create_hdf5_dataset, and fit_unlimited_data from examples/docs; use LargeDatasetFitter or AdaptiveHybridStreamingOptimizer instead.
Updated streaming workflow guidance to reflect hybrid streaming only.