fix(studio): rotate from the angle a CSS-rotated element shows, without GSAP too

[miguel-heygen]

What

A rotate gesture starts from the angle the element actually shows, and in a composition without GSAP the live draft shows the absolute angle the commit writes. Before this, a layer rotated by an authored CSS rotate: 30deg and turned another 25° was saved at 25°, and snapped back by 30° after the drop.

Why

The rotate base was GSAP's rotation plus Studio's old rotation variable, with || 0 when the page had no GSAP. The authored CSS rotate was ignored. The commit saved the absolute angle (0 + 25) as gsap.set({ rotation }), which replaces the CSS rotation once GSAP loads, while the draft had shown 30 + 25 by composing onto the authored value. The benchmark saw a 74.7 px drop on every CSS-rotated rotate cell without GSAP.

Related work

Stacked on #4776 (then #4774, #4766). Same reader as #4774: that PR reads position, this one adds rotation. Measured with the edit accuracy benchmark from #4760.

How

• readRotationBase(element): GSAP's rotation (plus the old variable) when the page has GSAP. Without GSAP it is the rotation GSAP's CSSPlugin will parse from the CSS rotate, scale and transform: the same folded matrix the fix(studio): start every drag from one position reader and draft what it commits #4774 position reader uses, now also giving atan2(b, a). There is no || 0 fallback.
• One rotation draft writer, applyRotationDraft, used for the drag and the release: GSAP's rotation, or without GSAP a CSS rotate of the target angle less what scale and transform already turn. A transform: rotate() or a mirrored element then drafts exactly what the commit will show.
• One restore, restoreRotationDraft, used for no-change, failed-commit and cancel: the CSS snapshot plus GSAP's rotation without the old variable.
• Deleted: readGsapRotation, the old composed rotation draft (applyStudioRotationDraft), and three "GSAP, else the old CSS channel" branches in the overlay gesture handler.
• Net lines: source 0 (64 added, 64 deleted); tests +30.

Score

The edit accuracy benchmark was run on a dev machine: every rotate cell at 100% zoom, with no GSAP, a GSAP tween or a GSAP hold, px/pct/centre/xPercent placements, 0°/30°, root and nested. That is 44 cells. Base is #4776. Both runs wait 3 s before each snapshot.

Metric	#4776	This PR
tracking	44/44	44/44
drop	38/44	44/44
reload	44/44	44/44
undo	44/44	44/44
all but smoothness	38/44	44/44

No cell got worse.

Before

A layer with an authored rotate: 30deg and no GSAP, rotated 25° more. After the drop it snaps back by 30°.

before-rot.webm

After

Same rotation: the layer stays at the angle it was dropped at.

after-rot.webm

Test plan

• Unit tests added/updated: the rotate base without GSAP reads the authored CSS rotation and drafts the absolute angle, and the draft leaves the part of the angle a transform rotate or a mirror (scale: -1 1, scaleX(-1)) already draws. Each fails without the change. The old variable's snapshot/restore test now uses the committed path.
• studio src/components/editor and src/hooks: all pass. tsc, oxfmt, oxlint and the comment ratchet are clean.
• Manual testing performed: the 44 benchmark cells above, and a before/after capture.
• Documentation updated (if applicable)
• Comments follow CONTRIBUTING.md "Comments"

[miguel-heygen]

Fixed on main by #4802: a rotate starts from the angle the element shows, its CSS rotation included, with or without GSAP. Every rotate case in the edit-accuracy gate, including those with a 30° CSS rotation, passes on main with no jump on drop. Closing.