[matt-pharr]

Release note

• Audience: developers
• Numerical impact: none (harness @ 0cd8239)
• Migration: none

The deck (main on a gpec.toml) and the scripting API are now checked against each other on the same inputs for Galerkin, Riccati Δ′, ascii/hdf5 file forcing, an analytic equilibrium on the two-pass auto grid, deck matching under gal_match_flag, and the fixed kinetic source. Where the two paths are known to disagree today, the test is marked @test_broken, so the fix for each drift shows up as an unexpected pass to flip.

Regression report

SDCC fan-out at 0cd823948 (vs develop @ 2842dfc2d) is running after the rebase onto develop 2842dfc2d;
the full report replaces this block when it lands.

Notes for reviewers

Tests only (test/runtests_solve_api.jl, +83/−2); no src changes. Rebased onto develop 2842dfc. Every new run uses the coarse Solovev fixture deck; the file's testset time goes from 39.5 s to 48.2 s (+8.7 s).

Parity that holds today (rtol 1e-12, exact in practice):

• Galerkin (gal_nx=32): Δ′ matrix and raw from main equal solve(..., Galerkin(nx=32)).
• Riccati: with the deck's own automatic chunking, Δ′ matrix, raw and coil columns equal the deck run. The existing nchunks=40 check at 1e-6 is kept.
• Matching: a deck with gal_match_flag (galerkin inner model) equals solve(MatchProblem(gal), GGJModel(solver=:galerkin)) on a Galerkin(cut_solution=true) result, for bpen, the matched ξ and the composite inner profiles. Since develop's :ray model returns no inner profiles by default, the galerkin model is the one that shows the cut-solution drift.
• File forcing: since the file-forcing normalization fix, the deck and a fresh RMPField agree on permeability and forcing_b for ascii and hdf5.

Known drifts marked @test_broken (2), and the change expected to flip each:

Drift	Flipped by
Analytic equilibrium on the two-pass auto grid: solve cannot re-form it and stays on the 1025-knot pass-1 grid, while the deck re-forms on a refined grid	storing the analytic config on PlasmaEquilibrium
Cut solution implied under gal_match_flag: the API without cut_solution=true has no composite inner profiles, where the deck has them	the change that routes deck matching through a MatchProblem clause (#554)

Two drifts this PR first marked broken were fixed on develop before the rebase. The file-forcing normalization now holds, so the hdf5 check here is a plain @test and the ascii case uses develop's own test. The fixed kinetic source is now accepted by solve, and develop's deck/API parity test for it replaces the broken one this PR had.

The two-pass case runs with psi_accuracy=3e-2 and no local-stability scan. At the defaults the local-stability scan on the 1024-interval pass-1 grid made it about 34 s.

🤖 Generated with Claude Code

[logan-nc]

Read in full. Ready to approve from my read; nothing pushed.

One cross-PR hazard to settle at merge time: the two-pass test sets "local_stability_flag" => false as a [ForceFreeStates] override. #568 removes that key without a deprecation entry, so once both land the override is an unknown keyword. On #568's semantics the fixture deck carries a [LocalStability] section, so the cheap form is delete!(inputs, "LocalStability") in _stage_deck rather than the flag. Whichever of the two merges second takes that line. The fixed-kinetic @test_broken flips with #550, as the body says.

[logan-nc]

Approving at 0ad81a2 on the comment review above. The one open item (the local_stability_flag override in the two-pass test once #568 lands) is a merge-order note, not a blocker.