The shape-insert hover preview renders a single optional ghost element
through `drawSlide` / `SlideRenderer.forceRender`. The drag-move
preview needs to paint one ghost per element in the selection, so the
ghost slot is generalized from `ghost?: Element` to a `readonly
Element[]` array.

Painting each ghost inside its own `save / globalAlpha = GHOST_ALPHA /
restore` band — rather than one outer band around the whole loop —
keeps the alpha guarantee local to each `drawElement` call, so a
future renderer that mutates ctx state without restoring can't leak
past the ghost boundary.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

Three small fixes from the post-Task-1 code review:

- Switch the ghost-array param's type form from `ReadonlyArray<Element>`
  to `readonly Element[]` — matches the convention used elsewhere in
  `packages/slides/src/view/`.
- Refresh two doc comments in editor.ts that still referenced the
  pre-array `forceRender(slide, doc, ghost)` signature, so they no
  longer lie about the API.
- Replace the new renderer test's save/restore counting assertion with
  a behavioral check that `globalAlpha` was actually set to
  `GHOST_ALPHA` during the ghost paint, via an instrumented setter on
  the spy context.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

`paintLive` synthesizes a slide with overridden frames, which is the
right affordance for resize / rotate / adjustment where the user
expects to see the live frame as it changes. Drag-move wants the
opposite contract: the original slide stays unchanged on the canvas,
translucent ghosts ride on top at the dragged offset, and the selection
handles anchor to the starting frame.

Adding a sibling helper keeps each paint path single-purpose rather
than threading a flag through `paintLive`. This commit only adds the
helper; `startDrag` continues to use `paintLive` and switches over in
the next commit. A short-lived `@ts-expect-error TS6133` directive
absorbs `noUnusedLocals` while the producer lives alone, and is
removed in the same commit that wires up the consumer.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

Switch `startDrag` from the synthesized-slide `paintLive` path to the
new `paintMoveGhost` helper. The mousemove loop now:

- Filters the selection to non-connector elements (`ghostSources`).
- Builds a `ghosts` array with `(x + dx, y + dy)` frames so the
  original slide is left untouched while the ghosts ride above at
  GHOST_ALPHA.
- Maintains a `committed` Map keyed by element id with the latest
  dragged frame so `onUp` can replay the snap result into one
  `store.batch` without re-running snap math.

`paintLive` itself stays untouched — it is still the correct affordance
for resize / rotate / adjustment where the user expects the live frame.
The stale `@ts-expect-error TS6133` directive above `paintMoveGhost`
from the previous commit is removed now that there is a real caller.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

The previous commit's `ghostSources` filter carved connectors out of
the drag preview, but the commit loop still walked the unfiltered
`originals`. `store.updateElementFrame` throws for connector elements
(they must be updated via endpoints, not frames), so any multi-select
drag that included a connector would error on pointerup — and the
"connectors excluded" comment in the code became a false promise.

Source the `committed` Map from `ghostSources` instead, so connectors
are excluded symmetrically: no ghost during the preview, no commit on
release, no throw. Add a regression test that drags a shape + connector
multi-selection and asserts both that the commit does not throw and
that the connector frame is unchanged.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

Show `cursor: move` while the mouse pointer is over the bounding box of
a selected shape so the drag affordance is discoverable before the
user clicks. The check runs in the existing canvas `pointermove`
listener (no new event subscription) and short-circuits whenever
something else owns the cursor:

- non-mouse pointers (touch has no cursor concept)
- insert mode (`crosshair` wins)
- text-edit mode (the embedded editor owns the box)
- pointer over a resize / rotate handle (those cursors already render)

`pointermove` fires at frame rate, so a `lastHoverCursor` cache makes
the DOM write only happen when the value actually changes. The cache
is reset on `pointerleave` and on `setInsertMode(null)` so a follow-up
hover always gets a fresh write.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

The first cut of the hover hit-test used a hand-rolled axis-aligned
bbox check. On a rotated shape that flips the cursor to `move` over
empty quadrants of the AABB and misses the rotated extent — so the
move-cursor affordance and the click-to-select hit-test (which already
uses `containsPoint`) disagreed.

Route through `containsPoint` so both paths share one rotation-aware
implementation. The existing import already pulled `combinedBoundingBox`
from `model/frame`, so no new import surface is needed. A regression
test exercises the failing axis-aligned path with a 45deg-rotated
rect.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

A stale `move` cursor persisted for one frame when the user
double-clicked a selected text box to enter edit mode: the entry path
ran before any pointermove would clear the cache. Drop the cached
cursor at the entry point so the text-edit affordance owns the canvas
cursor immediately.

Also reconcile the design doc with what shipped: move ESC mid-drag
cancel into Non-Goals (deferred to follow-up; undo is the v1
fallback), correct the affected-files table to reflect that overlay.ts
was not modified, and add the connector-in-selection and rotation-
aware-hover tests to the listed unit-test coverage.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

Three small adjustments needed to integrate the ghost-preview drag-move
with origin/main's group-aware drag that landed on top of the original
plan:

- Skip the commit batch when (liveDx, liveDy) is (0, 0). An empty
  `store.batch` still pushes an undo snapshot and clears the redo
  stack, which would silently corrupt history on a pure click-without-
  drag. Covers the original "connector-only drag" case as well as a
  mousedown-mouseup with no movement.
- Wrap the connector endpoint drag's `forceRender(..., ghostConnector)`
  in `[ghostConnector]` to match the new ghost-array signature — a call
  site that origin/main added between the original commit and the
  rebase target.
- Pull `containsPoint` into the existing `'../../model/frame'` import
  (re-establishing the rotation-aware hover hit-test that an earlier
  rebase resolution accidentally dropped from the import line).

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>