BPMN geometry

layout(view, policy=PreferDi, limits=Limits::new()) is a pure operation over moonrockz/diagrams/bpmn. It returns optional immutable geometry and diagnostics. It has no XML parser, DOM, clock or I/O dependency.

• PreferDi uses complete supplied geometry or computes the entire view when bounds or waypoints are missing. Invalid supplied values are errors.
• RequireDi returns no geometry unless required supplied geometry is complete.
• Automatic ignores supplied coordinates and computes the entire view.

Geometry.shapes and connections retain source model values, exact bounds and waypoints alongside numeric boxes. Labels retain exact bounds and font settings. A single translate_x/translate_y and canvas extent enclose every visible shape, label and waypoint with padding. Renderers apply that translation once. Hidden subprocess contents are omitted using the core model's shared Bpmn::scope visibility. Boundary events inherit their activity's visibility. Message connections may use visible display anchors while their source flow keeps its original semantic endpoint identities.

Supplied numeric tokens are preserved verbatim. Layout calculations require finite coordinates of magnitude at most one billion and positive shape extents. svg_number formats generated numbers to three decimal places with deterministic half-away-from-zero rounding. Limits bound visible nodes, edges, work and output. Failure returns no partial geometry.

Automatic layout

Each containment scope uses iterative strongly connected component condensation and ranks the resulting acyclic graph. Source order breaks ties. Measured task labels and external captions determine column widths and row heights. Pools and nested lanes form bands; expanded subprocesses have their own scopes. Connected data/artifacts can follow flow nodes into lanes while keeping their subprocess ownership. Category membership in Ownership.groups reserves shared member bands and determines group frames. Disconnected components pack in source order.

Boundary events are placed after their activity and reserve caption space. Connections use actual shape ports, with distinct ports for parallel edges. Circles, diamonds and stores use their outlines rather than bounding-box corners. The orthogonal router tries simple channels, then a bounded coordinate-grid A* search around unrelated shapes and label boxes. Return edges and self-loops can use outer channels. Connections whose endpoints share a pool or expanded scope stay within that scope. Cross-pool messages keep their declared visible endpoints. Connection labels avoid shapes, other labels and previously routed segments; subsequent connections route around them. No external layout service is used.

Limits::new(minimum_approach=24.0) reserves straight terminal corridors before any middle route or connection label is placed. The router also reserves space for a curve transition, separates independent collinear channels and avoids existing bends. A later edge's approach cannot be occupied by an earlier route or label. Larger minima increase automatic gaps and scope padding. Boundary event labels sit to one side of their outgoing bottom corridor. Supplied DI points keep their exact coordinates; these routing reservations apply to automatic geometry.

The coordinate-grid search indexes obstacles by row and column and caches segment checks. Index construction and lookups consume the same explicit work budget as routing. It does not increase the default work limit for dense cases.

The default bounds are 10,000 visible nodes, 50,000 connections, one million work units and 16 Mi UTF-16 output units. Increase Limits explicitly for larger searches. A dense graph can exhaust search before it reaches a node/edge bound; that produces bpmn/layout-limit and no partial layout. The frozen corpus has no legibility findings in either palette. Arbitrary inputs are not promised an optimal embedding or zero line crossings.

check(geometry, palette) returns bounded, read-only findings for unintended shape, label and route intersections, text fit and minimum size, contrast, ownership containment, boundary attachment and endpoint attachment. Exhaustion is explicit. Supplied geometry is never changed. Pools, lanes, expanded scopes, groups and their contained artifacts can overlap intentionally. Data and artifacts inside lanes use visual containment because they cannot have a flowNodeRef.

Use bpmn_palette(theme.dark) or its light counterpart to obtain the shared BPMN family hues with text shades that meet 4.5:1 against background and family fill. The sequence renderer continues to use its existing palettes and layout package.

Example

///|
test "compute and check a semantic view without DI" {
  let view = @model.Bpmn::new(
    "P",
    [
      @model.Node::new("Start", StartEvent, label="Start"),
      @model.Node::new("Review", UserTask, label="Review request"),
    ],
    [@model.Flow::new("Request", "Start", "Review", label="request")],
  )
  let result = @layout.layout(view, policy=Automatic)
  let geometry = result.geometry.unwrap()
  assert_eq(result.diagnostics, [])
  assert_eq(geometry.source, Automatic)
  assert_eq(
    @layout.check(geometry, @layout.bpmn_palette(@theme.Theme::classic().dark)),
    [],
  )
}