EQUIL - NEW FEATURE - Pin located kinetic-resonance surfaces into the auto psi grid

[logan-nc]

Stacked on #414. The kinetic-grid follow-up chosen after the certified-grid postmortem: instead of adaptive compression, make the auto grid resonance-aware statically — the cheapest layer of the safety-net design, and per the tests below, the only one currently justified.

What

When a run builds calculated kinetic matrices, locate the Ω_ℓ = 0 kinetic-resonance surfaces (same kinetic_resonance_psi_nodes locator the NTV quadrature panels at — one source of truth) and pin them into the two-pass auto ψ grid as plain knots via merge_mandatory_nodes — knot-at-node, no cleared zone, inserted before rational bracketing so the Δ′ clean-interval treatment wins wherever a resonance sits inside a rational's bracket. Nodes inside the near-axis validity region (ψ < ψ_c, #414) are suppressed anyway and are not pinned. ~30 lines, no new knobs.

Evidence

• DIII-D auto grid: 6 nodes located (the ψ=0.021 node — which explains the previously mysterious "sharp increment feature at ψ≈0.022" — correctly filtered below ψ_c), +2 net knots after pedestal nodes yield to rational brackets, et[1] unchanged to 2e-6, EL steps −14% (4,246 → 3,665: the grid finally knows about the ψ=0.17 resonance).
• Fixed-mψ convergence ladder (closes the review question "was the full grid converged above 512?"): fixed-1024 → 1.004073−0.259566i, within 0.12% of auto-288; fixed-512 is 1.6% off (fixed grids converge slower than the auto grid). Five grids across three families now cluster at 0.1–0.2%; nothing approaches the parked certified grid's 0.84–0.86.
• Mid-radius resonance stress test (toroidal_rotation_factor = 0.2 parks Ω_ℓ=0 nodes at ψ = 0.578, 0.812): naive auto grid vs 1024-knot gold agrees to 1.24e-3 — the profile-curvature term already tracks the reshaped ω_E — and pinning reproduces it to 1e-6. The feared naive-grid physics risk is bounded at the convergence-scatter level for this machine; pinning is cheap insurance plus the step-count win.
• Collisionless stressor (nufac = 0.02, layers narrowed ~50×): naive auto grid vs 1024-knot gold again agrees to 1.244e-3 — the same margin as the collisional case. The sharp-resonance risk did not materialize under the strongest stressor constructible on this equilibrium.

Design provenance

This replaces the parked adaptive approach (#410): detection of possible resonances is essentially free from the frequency profiles, so the grid can be made resonance-aware without any kernel-driven refinement loop. The remaining safety-net layers (a free post-hoc residual detector; one bounded probe-and-insert round) are designed and recorded in handoff/issue376/RESULTS.md §31 but not implemented — every stress test so far says the seeded auto grid is sufficient.

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

🤖 Generated with Claude Code

https://claude.ai/code/session_01LzbLFQKyuRE5DYZmLokKmk