v1.5.1 — viewer camera + screen-area LOD

Added

• Viewer camera — pan & zoom. The graph canvas gains a camera composed as
  a pure post-transform on the fit-to-content view (identity renders exactly
  the previous scene): cursor-anchored wheel zoom (up to 4×), drag-to-pan with
  a 4px threshold that keeps click/dblclick/hover semantics intact, dblclick
  on empty canvas to animate back to fit, and an elastic below-fit overscroll
  that springs home. New engine API: getCamera() / setCamera() /
  callbacks.onCameraChange.
• Screen-area LOD. Per-frame dot budgets driven by on-screen frame area
  (canvas/lod.js, ~the old 22-dot look at fit on normal frames, fewer on
  squeezed ones — the dense-project fit view finally breathes; all members
  reveal as you zoom), with a deterministic scattered reveal order, per-file
  labels past 64px dot spacing, inter-frame edges receding at high zoom, and
  sub-frame detail shedding during overscroll.
• Viewport culling + per-tick geometry cache — large graphs (activator)
  hold ~120fps at fit and deep zoom.
• Embeddable engine unit. adaptProjectData moved into canvas/adapt.js
  (re-exported from app/data.ts), and createEngine accepts optional
  isLight and storagePrefix — the canvas/ modules +
  app/{entity-store,ws-client}.js now form a dependency-free unit an
  embedder (Mesh) can vendor. Documented in
  docs/architecture/viewer-sync-engine.md.

Changed

• Frame-member adaptation no longer pre-slices to 22 nodes (MAX_FRAME_NODES
  removed); the LOD budget decides draw-time counts.