SCPN Fusion Core v4.0.0

Major release: the differentiable predictive free-boundary equilibrium line —
an SI free-boundary Grad–Shafranov solver with exact coil-current gradients,
a fully compiled forward (13.0 ms warm-started, fully converged 129² FP64 on a
dedicated H100; evidence chain in RESULTS.md), a cached batched (vmap)
variant for ensemble/MCMC workloads, real-DIII-D validation, and an IMAS
equilibrium bridge. The major bump reflects behavioural changes (fail-closed
input hardening, IMAS COCOS handling, MG preconditioner default, re-measured
disruption operating point), not API removals. The full-fidelity evidence
boundary is unchanged (GENE/CGYRO/GS2, full electromagnetic, DREAM kinetic,
mechanistic Aurora/STRAHL, and production MPI/multi-GPU parity remain
fail-closed).

Added

• Predictive free-boundary GS solver (jax_free_boundary_predictive): SI
  units, general p′/FF′ profile bases, von Hagenow wall coupling, Anderson
  fixed-point acceleration, smooth sub-cell O-point/X-point extraction (the
  coupled fixed point stays differentiable).
• Implicit-function-theorem adjoint with Jacobi-preconditioned solve —
  coil-current gradients agree with warm finite differences to ≤ 1.718e-05
  relative (committed sweep evidence artifacts/coilgrad_adjoint_fd_evidence.json).
• Compiled forward (jax_predictive_forward_compiled): the whole Anderson
  iteration in one lax.while_loop under jax.jit; warm-start support;
  MG-Richardson inner solver; iteration diagnostics (return_iterations);
  Anderson normal_eq Gram variant.
• Cached batched solve (solve_predictive_equilibrium_batched): lru_cache
  runner factory (a per-call jax.vmap re-jit was recompiling the whole
  batched graph every call), element-equivalence guarantees at span tolerance.
• Geometric multigrid preconditioner (jax_multigrid_precond): vertex-centred
  hierarchy, red-black Gauss-Seidel smoother matching the solver stencil,
  ring-exact identity rows.
• Differentiable compact profile bases: B-spline (read-only frozen cache) and
  Chebyshev ψ representation, with an end-to-end IDA-pattern differentiation
  test through the solve.
• IMAS equilibrium IDS bridge via OMAS with COCOS audit (solver COCOS 3 →
  IMAS 11) and fail-closed read/write validation.
• Real-data validation lane: DIII-D shot 145419 EFIT reproduction (full-domain
  0.72 %, shell-pinned 0.051 % span-relative RMS; executable cross-check and
  negative-control lanes; runtime + corpus provenance in the artifact).
• MAST forced-VDE open-data disruption pilot manifest (CC-BY-SA-4.0 source
  data; per-object SHA-256 digests) and reference-data provenance policy rules
  for both tracked PROVENANCE manifests.
• Evidence artifacts with generator/logic-source/pinned-environment digests and
  a stale-artifact freshness guard suite binding each artifact to its
  generator; dedicated-hardware *_h100.json / *_a100.json snapshots.

Changed

• Compiled forward defaults to the MG preconditioner ON (plain BiCGSTAB does
  not reach the inner tolerance at 129² within any practical cap — a
  robustness requirement, recorded in the iteration-count evidence).
• IMAS equilibrium bridge and predictive-solver inputs are fail-closed:
  non-uniform grids, malformed settings, wrong shapes, unknown solver names,
  and unsupported COCOS conventions now raise instead of proceeding.
• Disruption threshold evidence regenerated in the hash-pinned environment:
  the operating point moved from the stale (bias −5.0, FPR 0.50, pareto)
  record to the measured feasible point (bias −4.0, threshold 0.99, recall
  1.00, FPR 0.00) over the same 16-shot corpus; RESULTS.md and the claims
  manifest are bound to the current values.
• Dependency refresh across ecosystems: uv-pinned Python locks regenerated
  (superseding 7 pip Dependabot PRs), Rust serde/serde_json/thiserror/bytemuck
  bumps, pinned GitHub Actions updates.

Fixed

• Predictive adjoint convergence at fine grids (was 4× wrong at 65²) and
  grid-insensitive Jacobi preconditioning of the adjoint solve.
• Requirements regeneration honours each .in header (was mis-locking every
  extra to py3.12).
• Claims-manifest bindings retarget the disruption-sweep optimal block and the
  threshold-sweep prose; a new optimal-bias binding prevents silent staleness.
• Evidence hygiene: end-of-file newlines on tracked JSON artifacts, trailing
  newline writers in generators.

Artifacts (PyPI, verified at source)

File	SHA-256
scpn_fusion-4.0.0-py3-none-any.whl	3f9015473f1cf8ec42ad318b73af40fb0013de7b6c2a47705c20ed3971b9ceed
scpn_fusion-4.0.0.tar.gz	cb1a117cd85cc604fe2798b40fa89709cc8a736baecb8e82e455045807efc539

Verified: https://pypi.org/pypi/scpn-fusion/4.0.0/json