v3.0.0

A major refactor centred on a JAX-backed model, a new class hierarchy, and an opt-in gradient-based MCMC sampler. All user-facing call signatures from 2.x are preserved.

Highlights

• JAX backend for the rotation map model and likelihood: JIT-compiled, autodifferentiable, and GPU-aware.
• numpyro NUTS sampler as an opt-in MCMC backend via mcmc='numpyro' in fit_map() and Annulus3D.get_vlos_GP(). emcee remains the default. Legacy nwalkers/nburnin/nsteps kwargs map onto numpyro's num_chains/num_warmup/num_samples internally.
• Class hierarchy split: datacube → imagecube (shared FITS I/O) + momentmap (2D) + linecube (3D); Annulus → Annulus2D / Annulus3D. Legacy import aliases retained.
• imagecube.to_fits() and linecube.to_momentmap() (via bettermoments).
• GP backend swapped from celerite to tinygp.quasisep.Matern32 for JAX compatibility.
• Performance: ~16× cumulative speedup in warm fit_map calls via JIT-closure + vmap'd batch log-probability.
• Test suite (tests/) and GitHub Actions CI (pytest matrix on Python 3.11 + 3.13, ruff lint gate).
• Tutorial 6 demonstrates the numpyro path end-to-end.

Breaking changes

• Minimum Python is now 3.10 (was 3.8).
• celerite is no longer a dependency (replaced by tinygp). Drop it from pinned environments.
• A few previously-hidden DeprecationWarnings may surface (notably from emcee 3.x sampler.chain).

See the full CHANGELOG entry for the complete list of added, changed, removed, and fixed items, plus migration notes.