refactor: single-owner edge geometry pipeline + layout orchestrator - #53

Integrates two refactor branches of the visualization layer (refactor/edge-geometry-pipeline, refactor/layout-orchestrator), merged and seam-tested here. One PR, both halves reviewable via their own commit series.

Why

Edge geometry had no single owner: an edge's polyline was mutated by ~6 loosely-coupled stages across three files, endpoint anchors were stored but never read (every consumer re-inferred them from geometry with three different tolerance heuristics), hit-testing measured a polyline that was not the one drawn, and PixiRenderer carried an untestable layout-orchestration state machine layering three staleness mechanisms.

What changed

Edge geometry (refactor/edge-geometry-pipeline)

• Anchors are now the durable contract: decided exactly once at layout time from the pristine route (edgeAnchorAtBoundary, exact segment-direction reading instead of tolerance ladders) and consumed by every later stage. Deleted inferEdgeAnchor/inferAnchorSide/inferApproachSide/inferAnchorsFromPolyline/inferPointSide.
• New layout/edgeRoutePipeline.ts: the draw-time stage order is one function body — anchorEndpoints -> spreadEndpointLanes -> detourAroundObstacles, each a pure stage returning new records; a stage that moves an endpoint emits its updated anchor. edgeDrawing.ts is left with layer management and stroking.
• anchorEdgePolyline reports { points, rerouted } instead of silently discarding ELK's route; surfaced as RoutedEdge.origin.
• layout/routingConstants.ts owns every routing tolerance/margin, with derived values expressed as arithmetic (OBSTACLE_QUERY_MARGIN = NODE_OBSTACLE_MARGIN + DETOUR_GUTTER + allowance) so they cannot drift. Hot-loop constants are re-bound as module-local consts in edgeGeometry.ts (V8 does not constant-fold live imported bindings; an interleaved A/B run isolated ~8% routing cost before the re-bind).
• Hit-testing (hover, tooltip, double-click drill-in) now measures the polyline that was actually drawn, falling back to the layout route before first draw. This is the one deliberate behavior fix: after lane spreading, obstacle detours or node drags, the old code tested a line that was not on screen.
• edgeRoutingBudget.ts moved renderers/ -> layout/ (it is a layout-side dependency); extractLayout split into pure ELK-free functions (layout/elkExtract.ts).

Layout orchestration (refactor/layout-orchestrator)

• New layout/layoutOrchestrator.ts owns the layout-request lifecycle previously inlined in PixiRenderer: the run-latest coalescing queue, the request-id stale guard, the pending gate on the cheap visibility redraw, and the latest-vs-applied visible-set reconciliation. All side effects injected, so the invariants are unit-tested with fakes (burst coalescing, stale-result discard, mid-flight visibility reconciliation, error path releasing the gate, edges-phase no-op before first layout).
• PixiRenderer supplies effects and keeps only rendering; the four dead re-export shims (NodeRenderer, EdgeRenderer, LabelRenderer, interactionManager) are deleted.

Integration (this branch)

• tests/vizIntegrationSeam.test.ts wires the real geometry chain in as the orchestrator's effects and pins the composed invariants neither half tests alone: every edge phase receives the newest APPLIED layout (including a previous edge phase's result) as its input, routed polylines survive edge-phase reruns through that adoption, and anchors and geometry still agree after the full draw-time pipeline (including lane-moved endpoints). Both docs' merge conflicts resolved by combining the halves.

Edge-routing benchmark is at parity with main (all scenarios within run variance; 800n/500e indexed routing 251ms vs 259ms on main). Visual output is unchanged apart from the hit-test fix and rare anchor-side differences where the old tolerance ladder guessed wrong.