Default the bend actual field from g_ref

[DavidSagan]

Implements the Kn0_from_g_ref defaulting of bend.md, the behavior that replaced the g_actual / bend_field_actual pair retired in #72.

Unless the author says otherwise, a bend bends its own reference particle along the reference curve, so the actual bending field is the reference one: Kn0 = g_ref (equivalently Bn0 = Bn0_ref). resolve_bend_actual_field runs between the bend and multipole resolvers, seeding the dipole component that resolve_magnetic_multipoles then completes into all four forms.

b1:
  kind: Bend
  length: 1.5
  BendP:
    g_ref: 0.1
    Kn0_from_g_ref: true    # materialized default
    ...
  MagneticMultipoleP:       # created to hold the actual field
    Kn0: 0.1
    Kn0L: 0.15000000000000002
    Bn0: -0.33356405164803304
    Bn0L: -0.5003460774720496

The default stands down whenever the author has stated the field: any of Kn0, Kn0L, Bn0, Bn0L counts, since the integrated forms say the same thing over the length. A skew dipole does not — that is a different component. Kn0_from_g_ref: false declines the default outright, and a straight bend has no reference field to copy, so nothing is written and no group appears.

Two notes on where things land:

• MagneticMultipoleP is created on a curved bend that carries none. That is a third exception to "an absent group is left untouched", after the parser-added ReferenceP and FloorP; without it the default would be unobservable, since the common bend writes no multipole group at all. The comment on kGroupDefaults and the "leaves absent parameter groups alone" test both now say so.
• Kn0_from_g_ref: true joins kGroupDefaults, a non-false default of a group the element already carries, alongside ref_geometry: ARC.

Six new bookkeeper tests: the default itself, the false setting, an author-set Kn0 kept as-is, Kn0L counting as set, the field joining a group of other orders, and the straight bend. 174 tests pass.

🤖 Generated with Claude Code