auto mpsi under-resolves the edge, giving spurious dW and Delta-prime

[d-burg]

With grid_type = log_asymptotic and automatic radial resolution (mpsi = 0), the
optimal-mpsi heuristic allocates very few edge grid points (about a dozen near the
boundary). The resulting edge under-resolution produces spurious ideal dW and
Delta-prime values, especially when a near-separatrix rational surface is captured.

Workaround: set a fixed mpsi (e.g. mpsi = 256), which resolves the edge and gives
stable dW / Delta-prime.

The fix should make the automatic mpsi selection allocate adequate edge resolution
(or warn when the edge is under-resolved). Note this is an auto-mpsi problem, not
log_asymptotic itself: log_asymptotic with a fixed mpsi is fine.

[d-burg]

@logan-nc I scanned psi_accuracy with log_asymptotic + auto-mpsi and compared against ldp + fixed mpsi=256, checking the least-stable eigenvalue et[1] (the correct value, cross-checked against fortran DCON and the parallel-integration test, is ~0.80):

grid / setting	N_edge	mpsi	et[1]
ldp, mpsi=256	—	256	0.8006	correct
log_asymptotic auto, psi_accuracy=1e-3 (default)	5	210	0.6851	wrong, −14%
log_asymptotic auto, psi_accuracy=1e-4	8	631	0.8074	~ok
log_asymptotic auto, psi_accuracy=1e-5	13	1012	0.8020	converged
log_asymptotic auto, psi_accuracy=1e-6	23	1060	0.8021	converged

So you were right that psi_accuracy is the knob -- the default 1e-3 insufficiently populates the edge (N_edge=5) and gives a spurious δW, tightening to 1e-5 converges to the ldp value.

Cost is the problem -- the converged auto-grid needs mpsi ≈ 1012, and that total is dominated by N_mid=950, not the edge. The edge count grows much more slowly with psi_accuracy than the mid does, so to get enough edge points you over-resolve the mid and end up ~4x the total knots for the same answer that ldp gets in 256. In other words the inefficiency here is the N_edge/N_mid imbalance in make_optimal_mpsi, not core-packing in ldp.

Until this issue is resolved in a new PR, I'll keep ldp + fixed mpsi=256 for the baseline in our #295 merge since it's both correct and ~4x cheaper. Once auto-mpsi resolves the edge efficiently, log_asymptotic + auto-mpsi can go back to being the default and I'll re-pin the regression. Does this sound good to you?