PE/EQUIL - BUGFIX! - Withhold unvalidated kinetic tearing diagnostics and free the grid they constrained

[logan-nc]

Stacked on #425. Two coupled changes, both driven by the same finding: GPEC has no validated kinetic Δ′.

1. Withhold the unvalidated kinetic tearing diagnostics

Delta_prime, C_Delta_prime, resonant_current, resonant_area_weighted_field, island_width_sq, island_half_width and chirikov_parameter all descend from one quantity: the outer-solution jump in ∂b^ψ/∂ψ across the rational surface, read as a tearing index through ideal outer-region asymptotic matching at an ideal rational surface. Kinetic terms displace the surfaces (kinsing) and change the layer response, so that reading has no validated basis. In kinetic runs these datasets are now withheld with a warning naming exactly what is missing and why; surface metadata (ψ, q, m, n, area) and the inner-layer penetrated field are kept, and all ideal runs are bit-unaffected.

Opt-in [PerturbedEquilibrium] unvalidated_kinetic_tearing = true restores them for research use, with a warning that they must not be published without independent verification.

(The Riccati Delta_prime_matrix path was already unreachable in kinetic runs — kinetic requires the forward integrator — so PE's singular coupling was the only leak.)

2. Free the grid that constraint was paying for

Rational-surface bracketing exists solely to keep the ideal Δ′ stencil's 3rd derivative consistent across the surface (GridRefinement.jl). When that Δ′ is withheld, the bracketing is skipped and the fine near-rational knots survive for the kinetic matrices instead; opting back into the diagnostics restores the bracketing, so the grid always matches the physics being delivered.

Measured effect on the torque cross-check: none (DIII-D rot=1.0, deterministic): PE 0.1322 / fgar 0.1655 before and after, grid 290 → 288 knots. The bracket's cleared zone is only ~2e-4 wide and already contained 17–18 knots, so this is a cleanup of a now-purposeless constraint, not a fix — reported as the null result it is.

What the residual actually was

Chasing it produced the more useful result. The PE-vs-NTV gap on rot=1.0 is concentrated at the resonant layers — at ψ = 0.9933 fgar collects +0.0358 within ±0.002 while the EL solution collects +0.0026 — and it is a formulation difference amplified by a numerical smoothing width:

	PE (EL dissipation)	fgar	full domain	layers excluded (±0.002)
reg_spot = 0.05	0.1322	0.1666	20%	5.2%
reg_spot = 0.01	0.1322 (identical)	0.1402	5.7%	0.8%

PE is exactly reg_spot-independent (it is the ODE's own dissipation); fgar's layer contribution collapses 0.0453 → 0.0138 as the regularization narrows. At a rational surface the resonant torque enters the outer-region description as a surface term — the jump / resonant current, i.e. precisely the quantity part 1 withholds as unvalidated, while fgar integrates a reg-broadened volume approximation of it. So the two agree wherever that term is absent (0.8%), and cannot agree on the full domain until a validated kinetic layer formulation supplies it (issue #423).

Practical guidance this establishes: compare PE and NTV torque either outside the resonant layers or at small reg_spot; treat any full-domain difference at large reg_spot as the smoothing width, not physics.

Regression harness — and an honest coverage gap

regress --cases diiid_n1,solovev_kinetic_calculated,solovev_kinetic_ntv --refs bugfix/kinetic-clebsch-xi-s,local: 47 / 14 / 6 unchanged — nothing moved anywhere.

That clean result is partly a coverage gap, and should be read as one: the Solovev kinetic decks run fixed mpsi (16 and 128), so they never enter the two-pass auto grid where the bracketing lives, and they track no Δ′-family quantity, so nothing went missing. The harness therefore confirms only that ideal runs are bit-identical (47/47) and that this PR is inert on fixed-grid kinetic decks. The bracketing change and the withheld datasets are exercised only by the DIII-D kinetic-calculated auto-grid runs reported above. This is the third time this configuration has been the sole coverage for a change — a DIII-D kinetic-calculated FFS-only harness case remains the outstanding test-infrastructure item (see #398's disclosure and #423).

⚠️ REQUIRES THIRD-PARTY HUMAN REVIEW BEFORE MERGING — NON-NEGOTIABLE.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LzbLFQKyuRE5DYZmLokKmk

[logan-nc]

Closing and deleting: overstepped. The PE singular-coupling Δ′ is valid and must not be withheld — the untrusted-with-kinetics quantity is the Riccati Δ′, which is already unreachable in kinetic runs (kinetic requires the forward integrator, hard error at EulerLagrange.jl:159), so no guard was needed. The grid half of this PR is dropped with it; its measured result stands on the record as a null (removing the bracketing changed the PE-vs-NTV torque by nothing: 0.1322/0.1655 before and after).