Minimal mode: SGD stress layout with constraint-graph straightening

[hadley]

Experimental solution to #214.

Adds a minimal mode to the rendered dictionary page (toggle in the diagram toolbar, persisted per dictionary in localStorage): each table draws as its name only, one wire per relationship, laid out by a full stress model rather than dagre.

Layout pipeline (minimal mode)

1. Stress model solved by SGD — Zheng, Pawar & Goodman, Graph Drawing by Stochastic Gradient Descent (IEEE TVCG 2019, arXiv:1710.04626, s_gd2). All-pairs shortest-path ideal distances, weighted 1/d², inflated by box half-extents (Gansner & Hu) so the stress term itself keeps boxes apart; s_gd2 exponential annealing schedule; 8 deterministic restarts (seeded PRNG) scored on wire crossings of the wires as actually drawn.
2. Constraint-graph align/straighten — the coordinate-assignment half of Shape-Metrics (arXiv:2508.19416) reduced to what rectangle nodes need, no SAT: wires are compass-labelled, H/V wires become centre-equality constraints, non-overlap becomes separation arcs in a per-axis DAG, longest path assigns coordinates and compacts. A merge is refused when an arc already runs between the two groups (they overlap on the other axis, so sharing a centre would stack the boxes) — that wire keeps its diagonal, which the renderer draws happily.
3. Hill-climb polish on a cost of wire crossings + deviation from ideal length − alignment bonus.

Wire anchoring

Minimal-mode wires anchor to the nearest pair of compass points (3×3 grid: corners + edge midpoints) on the two boxes — aligned boxes get facing edge midpoints, so straight wires draw exactly straight. The layout's crossing scoring mirrors the same anchor choice.

Notes

• One shared BOX_GAP constant feeds the solver, the hill-climb's overlap test, and the separation pass, so they don't re-move each other's placements.
• tidy discards the saved arrangement but preserves the mode.
• Reviewer findings addressed: literal NUL byte replaced with a \0 escape (git treats the file as text again), dead acyclicity merge check replaced with the correct arc-between-groups infeasibility test (verified on the triangle repro), latent union-find revert corruption eliminated by checking feasibility before mutating parent.