Draw the vacuum field a start-up is read from: one cached evaluation, four quantities, a time slider

[HongSik-Yun-Fusion]

An interactive vacuum-field map

vacuum_field draws the field the coils and the vessel make with no plasma in
it: the flux, |B_p| in gauss, the field decay index, and the breakdown figure
of merit E_t B_t / B_p. field= chooses among them, time_index= steps along
the PF time base, and interactive=True gives both as widgets.

All four come from one evaluation, compute_vacuum_field_map, so the panels
of a start-up figure cannot end up describing four slightly different instants.

Why it is fast enough to slide

A grid's response to each coil and vessel loop is a pure function of geometry,
so it is built once and cached; a different time then costs a contraction with
that instant's currents. The vessel eddy solve is cached the same way, because
a recipe hands the evaluator a private copy of the ODS on every frame and would
otherwise repeat the same two-second solve each time.

	packaged shot
first evaluation, 65×65	0.67 s
another time, same grid	44 ms
a whole redrawn frame from an untouched ODS	0.22 s

The map is confined to the limiter, and that is the load-bearing choice

VEST's limiter bounding box contains 330 of the machine's 1480 current
filaments, so a map over the box is dominated by conductors — at the breakdown
instant the entire vessel fell in the bottom contour bin while a handful of coil
points took the whole colour range. Not one filament lies inside the limiter
outline, so masking to it leaves exactly the field a discharge would see, and
the levels then describe that field.

Levels come from a percentile of what is left, and Field2D gained an extend
policy so the points the percentile excludes saturate at the end colours rather
than appearing as holes indistinguishable from missing data.

A dense time base is a slider, not a list

This is the first plot to emit a range control. The kind was declared and unit
tested but never instantiated, so the widget path was unexercised; it works in
both the Matplotlib and ipywidgets backends. A times block names the option it
drives, so a stored equilibrium's handful of slices keeps its radio buttons.
Discovery resolves the default instant the way the builder does — from the
magnetics, before the private copy that drops them — so opening the controls
does not move the figure.

What the plan asked for and measurement refused

The plan had equilibrium_field_psi_vacuum move onto the same evaluator. It
should not: on the 129×129 equilibrium grid that plot draws, the exact elliptic
kernels cost 18 s against the polynomial path's 5.8 s, and their cached matrices
would be 591 MB. That trade buys a slider, not a single figure. It stays where
it is, and the evaluator now refuses a grid whose response would not fit its
budget instead of quietly allocating half a gigabyte and keeping it.

Also here

• Three reusable kernels in vaft.formula.equilibrium:
  poloidal_field_magnitude, decay_index_from_bz (nan where B_Z vanishes,
  not a spike), toroidal_electric_field (full weber in, per Formula psi helpers assume Wb/rad silently while the Green's functions return full Wb #354).
• Display policy: gauss for magnetic_field, a new electric_field quantity,
  and the decay index as dimensionless. Units are three vocabularies here, so
  the existing applies_to gate offers the flux units only while flux is drawn.
• _vacuum_sources factored out of compute_point_response_matrices_ods, so
  the cache key is derived from the same extraction the matrices are built from
  rather than restated beside it.

The ids declaration is load-bearing

With no time=, the map is drawn at the breakdown onset, and that timing reads
H-alpha from spectrometer_uv alongside the plasma current. My first version
left it out of the plot's ids. The OMAS adapter hands over the whole ODS, so
it never noticed; the IMAS adapter loads only declared IDSs, got no spectrometer
channels, and resolved an onset one PF sample earlier — during the ramp a 1.3%
difference in the drawn flux, not just the title. test_imas_omas_plot_equivalence
caught it; equilibrium_field_psi_vacuum has declared both all along for this
reason, and a direct test now pins it.

30 new tests across test_vacuum_field_map.py and test_vacuum_field_plot.py;
the formula catalog count moves 107 → 110 for the three new kernels.

Test run

Full suite on this branch, rebased onto 399f4e13: 9855 passed, 1 failed. The
one failure is test_flux_loop_voltage.py::test_legacy_flux_is_the_documented_voltage_composition,
which fails identically on a clean checkout of develop and is the single
failure in develop's own Package CI at the same commit: #679 made a degenerate
baseline window raise, and that test still feeds a 2000-sample signal with the
default windows at samples 3499:5000 and 11999:15000. Not touched here; it needs
its own fix.

Part of #487.

🤖 Generated with Claude Code