…ed view

Root cause traced to a sub-pixel threshold in applyHeightSnap (introduced
in be6f2b8): the half-dot tolerance used to detect a top-anchor resize
becomes < 1 screen pixel at low zoom (1 dot < 2 px), Konva's per-frame
scale-driven node-position FP noise easily exceeds it, every frame is
falsely flagged as a top-anchor drag, and pinBottomEdge marches the box
up out of the work area. A second false positive on deriveActiveEdges
(0.5 px default tolerance) lets the resize-time object-snap pull a
bottom-edge resize sideways.

Fixes:

- applyHeightSnap is now a no-op for shapes without a row anchor. Boxes
  and 1D barcodes don't need per-frame integer-dot snapping during the
  drag — onTransformEnd already rounds via the global snap function.
  Only stacked-2D barcodes (PDF417 / MicroPDF417 / Codablock), where a
  non-integer row count is invalid, keep the row-quantising path with
  pinBottomEdge intact.
- deriveActiveEdges default tolerance bumped from 0.5 px to 2 px so FP
  drift on a single-edge drag doesn't flip a stationary edge to active
  and trigger a sideways snap.
- New pinInactiveEdges enforces the resize invariant explicitly: edges
  the user did NOT grab (>2 px from start) are restored to their
  start-of-drag positions after every boundBoxFunc pass.
- transformStartBboxRef is captured lazily on the first boundBoxFunc
  call (using its oldBox) instead of in onTransformStart via
  getClientRect. On a rotated parent group those two coord systems
  diverge; capturing the start in the same frame Konva emits keeps the
  pin / activeEdges math consistent.
- When viewRotation is non-zero, the whole height-snap / object-snap /
  pin pipeline is skipped — Konva's bbox semantics there don't map onto
  an axis-aware model. Native resize is used; smart-align during resize
  is disabled in rotated views (drag-time smart-align unaffected).

Adds 3 regression tests for applyHeightSnap covering the noop / row-snap
/ top-anchor pin paths.

The same '(e) => onSelect(shift||ctrl||meta)' click + 'onSelect(false)'
tap pair was duplicated 14 times across the four shape renderers
(KonvaObject, LineObject, ImageObject, BarcodeObject). Extracts a small
selectionHandlers(onSelect) helper next to the shared KonvaObjectProps
type so each renderer just spreads {...selectionHandlers(onSelect)} on
its outermost selectable node.