FFS - NEW FEATURE - Certified adaptive psi grid for the calculated kinetic matrices

[logan-nc]

Stacked on #408 (base branch performance/decoupled-el-matrix-grid; a force-push trimming that branch back to the cap commits is pending, until then this diff shows only the seed-rework commit).

What

Adds kinetic_grid_tol ([ForceFreeStates], default 0 = off, bit-identical to today when off): when kinetic_source = "calculated", the expensive bounce-average kernel is driven over a certified adaptive ψ grid instead of every equilibrium knot.

• Certificate on the total matrices the EL solver consumes (ideal + kinetic, max-element residual vs tol · max|T|), including the adjoint combination kw₃ − kt₃ — never the kinetic increments in isolation.
• Physics-skeleton seed: the same structure the NTV ψ quadrature panels at — rational windows, located kinetic-resonance surfaces (kinetic_resonance_psi_nodes, one source of truth with the torque quadrature), and four interior points per inter-rational span (log-spaced in the axis span, matching the core Frobenius structure).
• Batched certify-or-refine rounds: every round evaluates all uncertified midpoints as one threaded batch (bulk-synchronous — the existing Threads.@threads kernel pattern is untouched). Spacing floors: the Frobenius cap in the core, RATIONAL_RES_SPACING outside, ÷4 inside rational windows (anti-aliasing where layers are expected). Certification exhaustion warns loudly with the first uncertified location.
• Kernel takes an arbitrary ψ list (psis kwarg through the KineticForces callback); everything else sizes from it.

Why

The shipped calculated-kinetic decks run mpsi = 16 and carry percent-level interpolation error on the B/C total matrices (ε 6.8e-2 → 5.8e-4 across a 16/64/256 ladder — mostly unresolved real structure, not noise). At production grids the kernel costs ~1.3 s/surface: minutes of matrix formation spent on knots the totals don't need, while coarse decks are silently under-resolved. Certification fixes both directions: it adds knots where the deck is starved and skips them where the equilibrium grid over-packs.

Evidence

• Knob-off bit-identity, re-verified on this branch's actual head (bef73a6, i.e. including the seed rework): regress --cases solovev_kinetic_calculated,solovev_kinetic_ntv,solovev_kinetic_nuzero,diiid_n1 --refs performance/decoupled-el-matrix-grid,local → 81/81 tracked quantities unchanged.
• Solovev smoke (m16 deck, tol=1e-3): 15 seed → 74 knots, 133 kernel evaluations, fully certified — the certifier correctly refuses the deck's under-resolution (more evals than the deck's 17 knots, because the deck was wrong).
• DIII-D survival test: FAILED its pre-registered physics gate — this PR stays draft until the root cause is decided. Full record in handoff/issue376/RESULTS.md §28. Summary: the DIII-D kinetic-calculated configuration (never run before) is pathological on the full grid too — 223k EL steps (94% at ψ<0.01) and 17.5 GB solution dumps. A discriminator run (kernel tolerance ×10 tighter) reproduces the full-grid et[1] to 2.7e-7 with steps unchanged, so the pathology is real near-axis structure, not kernel noise: the calculated kinetic increments diverge toward the axis (~ψ^−0.7; the kernel has no near-axis validity clamp). The certified grid smooths that structure (et[1] off by 17→14% across the tol ladder, converging toward the full-grid value with cost exceeding the full grid). Two design consequences recorded for the fix: certify the post-Schur matrices the ODE consumes (f0/K/G amplify the raw-increment certificate ~20×), and the Frobenius core floor is unjustified for kinetic increments. A denser-equilibrium-grid check confirms the full-grid reference is converged (1.6e-4), so the certified deviation is unambiguously certification error; the identified design flaw is the certificate's GLOBAL max|T| scale, which is blind to the small-amplitude but dynamically dominant near-axis structure (1.6e-4 of the G scale, yet 94% of the ODE work). The blocking question is physics: should the drift-kinetic contribution be tapered/cut near the axis where the model degenerates? If yes, both the step pathology and the certification failure disappear together.

⚠️ REQUIRES THIRD-PARTY HUMAN REVIEW BEFORE MERGING — NON-NEGOTIABLE. DO NOT MERGE WITHOUT AN APPROVING HUMAN REVIEW.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LzbLFQKyuRE5DYZmLokKmk