Layout presets by graph size: large graphs in seconds, not minutes

[endrix]

The server's ELK layout now picks its option preset by the graph's node and edge counts, so large graphs lay out in seconds instead of minutes (or a stack overflow). Graphs up to the first threshold get exactly the options they did before, byte for byte.

What was measured

ELK standalone on the exact graphs the engine hands to elk.layout, four real netlist modules of a hardware product:

module	nodes	edges	default	middle	fast (POLYLINE)	fast (ORTHOGONAL)
storeack	212	445	5.8 s	~1 s	0.9 s
loadreq	344	936	41 s	1.7 s (2.6 s re-run)	1.3 s (0.9 s re-run)	2.7 s
priv	1528	3617	334 s	9.3 s (10.6 s re-run)	3.8 s	30.8 s
conv	3293	8112	RangeError: Maximum call stack size exceeded inside ELK	not run	11.5 s	2289 s

Two options carry the default preset's cost: elk.layered.nodePlacement.strategy: NETWORK_SIMPLEX (dominant: switching it alone takes 344 nodes from 41 s to 3.8 s) and elk.layered.considerModelOrder.strategy: NODES_AND_EDGES (ten-fold on the cheap presets). Crossing minimisation is not the cost. Orthogonal routing of a simply placed graph is: the fast preset was measured with both routings on this branch, and ORTHOGONAL is 38 minutes on conv, so the fast preset routes polylines. The middle preset stays orthogonal, where POLYLINE was the slower of the two.

Tiers

Constants in packages/diagram-server/src/server/layout-configurator.ts, with the measurements in their doc comment:

• default: up to 150 nodes and 400 edges. The options are unchanged.
• middle (MIDDLE_LAYOUT_OPTIONS): above 150 nodes or 400 edges. Default plus considerModelOrder.strategy: NONE, nodePlacement.strategy: BRANDES_KOEPF, thoroughness: 3.
• fast (FAST_LAYOUT_OPTIONS): above 1000 nodes or 3000 edges. Middle plus crossingMinimization.strategy: NONE, nodePlacement.strategy: SIMPLE, thoroughness: 1, edgeRouting: POLYLINE.

The count is recursive over the graph's children, GNodes that are not compartments and GEdges. The rule reads nothing else, so the core learns nothing product-specific.

The tier applies wherever the configurator does: the initial layout on open, Layout Diagram, boundary flow, Untangled, Free Ports and the Free Ports probe. The engine still merges a layout's own graph options (UNTANGLED_ELK_OPTIONS) over the preset, except in the fast tier, where it re-asserts the fast preset: Brandes-Koepf placement is among the options that take minutes there. In the middle tier Untangled's placement is the middle preset's already. Documented in both files.

Each layout logs [dialogram perf] layout <network>: tier=<default|middle|fast> nodes=<n> edges=<m> beside the existing elk= timing, and a fast-tier layout sends a StatusAction ("N nodes: laid out roughly. Run Layout for the full layout.", 10 s) through the ActionDispatcher the engine factory now passes in; the stock GLSP status module shows it, so no client code changed.

Tests

packages/diagram-server/test/layout-tier.test.ts, 13 tests on synthetic GGraph trees: the counting (depth, compartments skipped), the thresholds (at and above), the default preset pinned for a small graph and equal to an empty graph's, the middle and fast presets by node and by edge count, the fixed post-drag configuration untouched, Untangled overriding in the middle tier and overridden in the fast tier, the log line and the status action. The existing elk-layout-port-order.test.ts suite runs the real ELK on the 40-node speculation flow and reports tier=default, every crossing count unchanged. No netlist dump runs in the suite.

Gates

• npm run typecheck: PASSED (5/5 packages)
• npm run check:neutrality: PASSED (5/5 gates)
• npm run build: succeeded
• npm test --workspace packages/diagram-server: 68 files, 339 tests passed
• npm test --workspace packages/sidecar-toolkit: 42 files, 264 tests passed

https://claude.ai/code/session_01UtmwstZRcWJeg3C69s1TnS