v0.9.0

What's Changed

• Explain current behavior without tracker breadcrumbs by @jc-macdonald in #254
• Document execution roadmap and measured CPU baseline by @jc-macdonald in #262
• Honor fixed-loop controls across provider execution paths by @jc-macdonald in #263
• Preserve native state gradients on supported Torch 2.6 by @jc-macdonald in #265
• Integrate bounded forward parameter and measurement sensitivities by @jc-macdonald in #266
• Qualify optional bounded CPU Torch compiled sequential loops by @jc-macdonald in #267
• perf: reduce NumPy SSA event work and qualify stochastic baselines by @jc-macdonald in #268
• feat: adapt current compiler RHS to optional Diffrax terms by @jc-macdonald in #269
• Add checked linear actions and eager operator solve contracts by @jc-macdonald in #270
• Qualify eager banded and matrix-free SciPy solve adapters by @jc-macdonald in #271
• Route full-state RHS JVPs through SDIRK stages by @jc-macdonald in #272
• Add checked eager Newton–Krylov solves and stage diagnostics by @jc-macdonald in #273
• Add independent adaptive meshes and bounded JAX block replay by @jc-macdonald in #274
• Fix JAX/CUDA gradients through forcing schedules by @jc-macdonald in #276
• Add scoped local equilibrium stability diagnostics by @jc-macdonald in #278
• docs: index executable examples by @jc-macdonald in #279
• Decouple core release and harden documentation publication by @jc-macdonald in #280
• Release op_engine 0.9.0 (core only) by @jc-macdonald in #281

Full Changelog: v0.8.0...v0.9.0