feat(ui): solid mode-displacement arrows + CIF/mCIF export from the 3D card

[drthyang]

Two UI requests, plus a data-integrity fix found while wiring the second.

1. Distortion-mode arrows (PDF page, irreps)

Two independent problems with the green arrows:

Too long. The scale was span * 0.34 where span is the bounding-box body diagonal — on GaTa₄Se₈ (10.36 Å cubic) that is 17.9 Å, so the longest arrow ran ~6 Å, well past the neighbouring site. The pattern read as a vector field over the cell rather than a displacement of each atom. Now min(shortestCellEdge · 0.26, span · 0.12) ≈ 1.8 Å, about one bond. Tying it to the cell edge rather than the diagonal also stops arrows growing when a supercell is displayed.

Stems too thin. The real culprit: THREE.ArrowHelper draws its shaft with LineBasicMaterial, which WebGL rasterizes at a single device pixel however close the camera gets — the stem was a hairline that disappeared against the lit spheres. New arrowBuilder emits a real cylinder shaft + cone head in MeshPhongMaterial, lit by the same lights as the atoms and bonds. Shaft and head sizes are constant across a set, so length alone carries the amplitude and the smallest arrow of a pattern stays legible; unit geometries are built once per overlay and scaled per arrow.

Verified in the running app on the GaTa₄Se₈ demo, mode 1 @ Γ (Ta1): the ⟨111⟩ arrows on the 16 Ta copies now read as solid local displacements.

2. CIF export from the 3D model

StructureView gains exports — a list of host-supplied {label, title, run} descriptors rendered as download buttons in its toolbar. Descriptors rather than a boolean so the viewer never learns which flavour it is writing and the label tracks the content; that is what leaves room for the mCIF work ahead (moments today, mode displacements once the writer carries them).

Host	Button	Writes
PDF card	↓ CIF	the phase on screen
Rietveld card	↓ CIF / ↓ mCIF	the phase the picker selects
Magnetic analysis card	↓ CIF → ↓ mCIF once a group is previewed	moment loop at the current amplitudes

The magnetic-analysis card is beyond the two pages originally asked for, but it is the view that actually holds the moments, so it is where mCIF means the most — and its export uses appliedMagnetic, the very model drawn as arrows, so file and picture agree.

Fix: every CIF export wrote the starting structure

Found while tracing the export path. structureToCif reads coordinates and cell straight off the model and consults the parameters only for their esds — but all three workbenches passed the pre-refinement structure. Exported files therefore carried starting coordinates annotated with refined uncertainties and refined R-factors.

All three now pass the model with parameters applied. The powder path additionally routes bindings by phase (as refinedPhases does), so a multi-phase session cannot attach one phase's esds to another phase's identically-labelled site. cif.test.ts pins the contract in both directions.

Verification

• Full suite: 1126 passed, 0 failed
• tsc -b --noEmit clean
• Both features exercised in the running dev server (GaTa₄Se₈ PDF demo for the arrows; Mn₃Ga + MnO for the CIF→mCIF label switch)

Not touched

The red magnetic moment arrows still use ArrowHelper, so they share the hairline-stem issue and sit at span * 0.42. Left alone deliberately — out of scope, and that rendering has a validated look plus test coverage. Routing them through the same builder is a two-line follow-up.

🤖 Generated with Claude Code