fix(sdk,studio): R5 cutover review fixes (on top of #1539) (#1545)

* fix(sdk,studio): R5 cutover review fixes — fromTo dest, timing sync, parity

Confirmed correctness findings from the R5 review of the SDK cutover stack,
applied on top of #1539:

- fromTo add via cutover dropped its destination: handleAddGsapTween read only
  `toProperties`; now falls back to `properties` like every other method.
- handleSetTiming GSAP sync: a clip with no data-start skipped the shift (now
  treats start as 0, matching the server path) and a blank/non-numeric
  data-start wrote position: NaN (now sanitized).
- handleSetTiming no longer appends an absolute position to an auto-sequenced
  (implicit-position) tween, which collapsed staggers.
- handleSetTiming keeps data-end in sync when a clip carries BOTH data-duration
  and data-end (a stale data-end inverted the clip).
- string/relative tween positions ("+=0.5", "<") documented as a known ceiling.
- opacity/autoAlpha property seed no longer falsy-zero (`|| 1`): an element at
  opacity 0 seeds 0, not 1.
- optimistic add-keyframe cache tolerance aligned to the writer's PCT_TOLERANCE
  (2%) so a near-neighbour keyframe no longer shows then vanishes on reload.
- DOM-patch finiteness validation runs before the SDK cutover path.
- attribute ops mapping to a reserved data-* name decline the cutover up front
  instead of throwing inside dispatch.

Regression tests added for each SDK-side fix.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(studio): close two gaps in the reserved-attr cutover gate

- Lowercase the mapped attribute name before the reserved check, matching the
  SDK's validateSetAttribute (which lowercases), so a case-variant reserved
  name is declined up front instead of throwing inside dispatch.
- Also gate `html-attribute` ops (raw, non-prefixed names), not just bare
  `attribute` ops. Both the emitter and the gate now derive the name via one
  shared `sdkAttrName` helper so they can't drift.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(studio): match keyframe remove-path tolerance to the writer (mirror of add)

The optimistic remove-keyframe cache filtered with `> 0.001`, dropping only a
near-exact match, while the writer removes within PCT_TOLERANCE (2). Removing
at e.g. 49% dropped a 50% keyframe on disk but left it in the cache — a phantom
that vanished on reload, the inverted twin of the add-path tolerance fix.
Now filters with `> 2` to match.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

packages/studio/src/hooks/useGsapPropertyDebounce.ts

@@ -136,7 +136,10 @@ export function useGsapPropertyDebounce(
         defaultValue = Math.round(property === "width" ? rect.width : rect.height);
       } else if (property === "opacity" || property === "autoAlpha") {
         const cs = el.ownerDocument.defaultView?.getComputedStyle(el);
-        defaultValue = cs ? Number.parseFloat(cs.opacity) || 1 : 1;
+        // Use `|| 1` only as a non-finite fallback, not a falsy fallback: an
+        // element currently at opacity 0 must seed 0, not 1.
+        const current = cs ? Number.parseFloat(cs.opacity) : Number.NaN;
+        defaultValue = Number.isFinite(current) ? current : 1;
       }
       const { sdkSession, sdkDeps, activeCompPath } = sdkRef.current ?? {};
       if (sdkSession && sdkDeps) {