RayZhang2024/NEAT: NEAT v4.8.2

Changelog All notable user-facing changes to NEAT are recorded here. 4.8.2 - 2026-08-20 Added shared user-configurable lower and upper bounds for fitting parameters s, t and eta. Applied the same bounds to individual-edge, full-pattern and batch fitting. 4.8.1 - 2026-07-18 Changed Official standalone downloads now enable NEAT shared AI access automatically; users no longer need to configure a service URL or access token. Shared-access labels describe a limited daily request allowance without advertising a fixed number in the interface. Complete per-computer shared-service settings continue to override the bundled public configuration for development and testing. Packaging and security Release automation injects the limited public-client service credential from a GitHub Actions secret and stops the build when it is missing. The packaged smoke test verifies that automatic shared access is configured. The server OpenAI API key remains exclusively on the hosted laptop and is never written to the release package. The bundled public-client token is intentionally extractable; the atomic server-side daily allowance and OpenAI project budget remain the cost controls. 4.8.0 - 2026-07-18 Added Dockable NEAT AI Assistant with grounded answers and verified source links. Dedicated AI Assistant menu, settings window and Explain current screen action. Personal model access for OpenAI, Anthropic, Google Gemini, DeepSeek and Kimi/Moonshot, plus manually configured OpenAI-compatible endpoints. Local-model support through Ollama and LM Studio, including automatic model discovery and no-key localhost operation. Optional hosted shared access with a global daily request allowance. Curated software, UI, preprocessing, fitting, mapping, post-processing, troubleshooting and known-limitations knowledge for assistant retrieval. Local helpful/not-helpful feedback recording with path redaction. Rectangle and polygon mask-editing tools and a mask-applied preview. Uncertainty-planning estimator and additional fitting controls and tests. Full Process data and open-beam folder selection controls. Changed Normalisation defaults to adjacent-frame half-window m = 0; validated ranges are 0-100 for spatial half-window n and 0-10 for m. Positive spike cleaning uses a 10x threshold relative to the surrounding 5x5 mean, excluding the centre pixel. Clean expands to a 7x7 neighbourhood when the initial neighbourhood has no valid replacement value. FITS and TIFF image orientation is handled consistently across loading, processing and writing. Fitting-window, fixed-parameter uncertainty, phase-definition and d_110 naming behaviour have been clarified and tested. Assistant indexes, settings and feedback are written to per-user application data rather than the source or executable directory. Assistant retrieval falls back to bundled BM25 search if the optional ONNX semantic runtime or first-use model download is unavailable. Packaging and security Windows release builds include assistant providers, keyring support, Chroma, ONNX Runtime and the approved knowledge base. CI installs and tests both desktop-assistant and shared-server dependencies. The packaged release has a non-network import and knowledge-base smoke test. API keys and hosted-service secrets remain outside committed settings and release artifacts. Known limitations Shared access is disabled unless the service URL and access token are explicitly configured for the installation. Local LLM answer quality and performance depend on the selected model and hardware; users should validate models against representative NEAT questions. Scientific interpretation remains subject to the assumptions and review boundaries documented in the assistant knowledge base.