Three slides-editor bugs surfaced while testing groups; they travel
together because the same `/shared/<id>` deck exercises them in
sequence.

1. Lines drift after grouping their endpoints (the reported bug).
   `MemSlidesStore.group()` keeps a connector as a group child when
   both endpoints reference candidates, but `drawElement`'s connector
   branch skipped the per-element frame transform — `drawConnector`
   expects world-coord endpoints and `buildElementWorldLookup` lifts
   grouped targets to world, so the group transform compounded on top
   of already-world endpoints and the line jumped by the group's
   offset.

   `drawElement` now threads a cumulative `parentTransform`; on a
   connector child it inverts the parent transform onto the ctx and
   hands `drawConnector` the lookup's view of the connector — free
   endpoints are world-coord via `walkWorld`, attached endpoints
   already resolve through the lookup, so both kinds agree. The
   "v1 group() never includes connectors" NOTE in
   `element-renderer.ts` was stale since #263; #320 then bolted the
   world-coord lookup on top, which is exactly what broke this.

2. Rotate angle tooltip drifts by the pasteboard offset.
   `acquireRotateTooltip` appends to `overlay.parentElement` so
   `renderOverlay`'s innerHTML reset can't wipe it mid-drag, but
   `showTooltip` was computing coords against `overlay`'s rect.
   These containers shared an origin until #353 gave the overlay a
   non-zero `slideOffsetCssX/Y` inside `canvasWrap`. Measure against
   the tooltip's actual parent rect instead.

3. Rotate tooltip flickers at the previous drag's last position on
   re-acquire. `acquireRotateTooltip` flipped `display: block`
   immediately while `transform` still held the previous drag's
   terminal coords. Acquire now returns the cached element hidden;
   `startRotate` calls `showTooltip(clientX, clientY, 0)` once
   immediately so `transform` and `display: block` land in the same
   paint frame. Bonus: a 0° readout pops at the click position with
   no input latency.

Pre-merge code review surfaced one blocking concern and seven
non-blocking follow-ups (tracked in the task doc).

Fixed before merge: `invertGroupTransform` now returns `null`
instead of throwing on a singular matrix, and the connector branch
skips the paint. A degenerate group (frame.w = 0 with refSize > 0,
reachable via PPTX import) used to throw, escape `drawElement`'s
try/finally, and blank every later element on the slide.

Deferred follow-ups: flip blindness in `groupToTransform`;
reference-equality identity gate; ghost connector inside a group
ghost; `tooltipContainer` cached once at `startRotate`;
`overlay.parentElement` null fallback inside `acquireRotateTooltip`;
three copies of the 6-coef affine inverse math
(`element-renderer.ts` + `model/group.ts`); acquire-then-show
invariant naming.

Regression coverage:
- `slide-renderer.test.ts` — attached connector: `ctx.transform`
  runs before `moveTo` with the expected inverse coefficients; free
  connector: pre/post-group `moveTo`/`lineTo` args match (the lookup
  swap); singular parent transform: skips that connector but keeps
  painting the rest of the slide.
- `editor.test.ts` — pasteboard layout: tooltip transform encodes
  parent-frame coords, not overlay-relative; two-cycle drag: second
  pointerdown's transform reflects the new click position, not the
  previous drag's last move.