[max-models]

The 3D field (step R4) gets coils of real wire and the network resistance of a driver.

• Feed3D: a current source along grid edges across a gap between the terminals of an ImpedanceConductor3D (a delta-gap source). The current returns on the metal's surface, so the terminals' impedance includes the coil's own Joule loss and proximity effect. Terminals that are not joined through one piece of metal on the grid are refused (the source would not be compatible with the gauge). Tests (test_feed3d.py): a coaxial line of copper against its closed form, reactance to 0.13 % and resistance to 4 % at a / h = 4; the gauge's eps and the feed's direction do not matter; a ring of wire.
• The slotted-cylinder test as unit tests (test_cutcell.py), against NGSolve with the same surface impedance: the z-invariant slotted tube to 0.1 % with cut cells (staircase 20 % off), the finite tube with no, closed and through slits to 6 % at h = 2.5 mm, and the closed slits' effect on the loss.
• benchmarks/rnet_driver.py: R_net of the BUG driver, a six-turn helical coil with leads, a Faraday shield with 24 closed slits and a conducting housing (simgeometry's Helix and PolarRepeat, Add Helix and PolarRepeat simgeometry#46), cut cells cached on disk. At h = 4 mm: coil 0.071 Ω, shield 0.196 Ω, R_net 0.266 Ω (measured 0.6 Ω, with the same 25/75 split). Several dimensions are assumptions, the shield is thicker than a real one and the grids are reduced: the converged value needs a production run.
• InductiveField3D(cut_cell_samples=...), and a fix of the multigrid's coarsest level, which could be exactly singular.
• Docs: the 3D field tutorial (the coax and the slotted cylinder) and the discretisation page (the feed, R_net).

The benchmark needs simgeometry#46; the tests do not.