VMEX 0.11.0

Changed

• The interior field is built from the native VMEC form (#403).
  B^theta = (chi' - lambda_zeta)/sqrt(g) and B^zeta = (phi' + lambda_theta)/sqrt(g)
  with sqrt(g) from the same R, Z series as the position, and a C2
  radial interpolant. On the breathing-circle oracle at ns = 41 the second
  derivative of B improves from 2.9e-2 to 2.2e-4 relative error and
  |div B|/|grad B| is at round-off (8.5e-18). Spectra without lmns,
  phipf and chipf keep the previous fitted path.
• Minimum versions (#410): booz_xform_jax>=0.4.0 (its Booz_xform
  class rebuilds cached grids when the resolution changes and reports a
  correct __version__; the JAX kernel is unchanged from 0.3.0),
  solvax>=0.21.0, and gkx>=1.8.0 for the turbulence extra.
• Force-balance polishing is compared on one mesh (#412). The example and
  README now state that the gain is concentrated near the axis (RMS force
  2.9e3 -> 61 N/m^3 for rho < 0.2) with a small volume-average change
  (2.27e-3 -> 1.91e-3), and that the summary plot's force panel cannot
  resolve it.

Fixed

• surface_field_data_from_state raised TypeError for every caller
  after #403 added three keys to the live-state spectra (#409).
• Free-boundary recovery depended on trial history (#416). A shared
  cold-rebuild budget meant that, after unrelated rejected trials, a
  recoverable point returned the stalled restart instead of its converged
  root. Each stalled restart now gets exactly one deterministic cold retry.
• Radial lifts accepted under-determined spline fits (#414), inventing
  curvature in unsampled spans; they now raise.
• Warm workflow profiles timed only the first stage (#420); schema 2
  repeats every stage.