Version 0.2.7

BrainState 0.2.7 modernizes the experimental compilation stack, deepens the transformation APIs, and tightens runtime infrastructure across the project.

Experimental Compiler and Visualization

• Introduced the experimental neuroir compiler built on dataclass-based graph IR elements and an explicit CompilationContext, improving dependency tracking, hidden-state mapping, and ClosedJaxpr fidelity even for self-connections and delay buffers.
• Added GraphDisplayer and TextDisplayer backends with hierarchical and force-directed layouts, plus richer diagnostics and tests that cover large sample networks and neuro-graph visualizations.

Transformations and Autodiff

• Added the jit_named_scope decorator and supporting utilities so nested transformations emit meaningful names inside traced functions, together with _make_jaxpr refinements that separate dynamic/static arguments and improve caching semantics for StatefulFunction.
• Expanded the gradient toolkit by exporting the new Jacobian (forward and reverse), Hessian, and SOFO transforms, unifying gradient handling for classes, auxiliary returns, and state-aware updates through the transform module.

State and Runtime Enhancements

• Replaced the experimental ArrayParam with a dedicated DelayState, propagating the new state through the compiler, delay modules, and neuro-IR so historical buffers participate in tracing and optimization just like other states.
• Environment helpers can now run against injected EnvironmentState instances, enabling sandboxed or per-thread configurations while DelayState-aware unit tests extend coverage of the updated modules.

Experimental and Infrastructure Updates

• Completed the neuron IR → neuroir rename, aligned the GDiist BPU codebase with the new terminology, and added new sample networks plus placeholder skips to keep the growing compiler/displayer test surface manageable.
• Added braincell to the development requirements, refreshed documentation wording, and kept CI dependencies current for the GitHub Actions runners.

Bug Fixes

• Hardened caching, randomness, and initialization logic by fixing get_arg_cache_key, removing stale decorator parameters, validating truncated normal draws, and correcting the exported version metadata.
• Declared Python 3.14 support and cleaned up compiler import ordering to keep linting noise low.

What's Changed

• Add Jaxpr optimization passes and integrate constant folding by @chaoming0625 in #114
• Enhance compilation with BPU parser, IR optimizations, and device wrappers by @chaoming0625 in #115
• deps(deps): bump actions/checkout from 5 to 6 by @dependabot[bot] in #116
• update document by @xinzhu-L in #117
• Support custom EnvironmentState instances in environ helpers by @chaoming0625 in #118
• Add unified gradient transforms with Jacobian, Hessian, SOFO by @chaoming0625 in #119
• Refactor(state): remove ArrayParam, add DelayState by @chaoming0625 in #122
• feat(neuroir): add NeuroIR core module and API by @chaoming0625 in #123

Full Changelog: v0.2.4...v0.2.7