Changelog

Brought four plasma kernels to a single canonical contract across the Rust and NumPy tiers with fastest-first dispatch. The tearing-mode model now runs the full Modified Rutherford equation (bootstrap-current drive, seedable Gaussian process noise, and disruption labelled only on an actual island-width threshold crossing), bit-exact with NumPy on the deterministic growth step and statistically equivalent on the stochastic trajectories. The geometric multigrid V-cycle gained a coarse-grid correction sign fix — the residual now collapses by orders of magnitude within a few cycles instead of diverging — plus a Rust full-solve binding. The magnetic-probe diagnostic (endpoint-inclusive probe placement, bilinear flux interpolation) and the minimum-norm/ridge coil-current solve were unified into shared free-function kernels across both tiers. Capability inventory: 278 Python modules, 547 classes, 12 Rust crates, 436 test files.

Improved the public documentation path across the README, project overview, onboarding, and API overview; reframed the applications and market context around control iteration, validation packaging, and acceptance evidence; refreshed the generated capability snapshot. Capability inventory: 277 Python modules, 547 classes, 13 Rust crates, 425 test files.

Native same-case profile-source metrics, finite signed-q sanity, and machine-readable strict threshold checks for free-boundary equilibrium; whole-plant fault-tolerant scenario campaign (10/10 reduced-order lanes); fail-closed acceptance benchmarks for native nonlinear GK, runaway electrons, and impurity transport.

QLKNN-10D data acquisition and conversion pipeline (QuaLiKiz flux database to training tensors); neural transport surrogate (test rel_L2 = 0.094) and FNO turbulence surrogate (val rel_L2 = 0.055).