fix(studio): resume drag-paused timelines instead of only re-seeking

[vanceingalls]

What

One-cause fix for "selecting/dragging an element permanently kills its animations" in Studio.

createManualOffsetDragMember (drag start) pauses every timeline in window.__timelines and records the list in data-hf-drag-paused-timelines. resumeGsapTimelines then removed the attribute and only re-seeked the player — it never unpaused anything. The main timeline survives because the player seeks it every frame (seek works on paused timelines), but play-state-driven sub-composition timelines froze permanently after any drag gesture, and deselecting couldn't recover them — nothing else ever resumes them. Symptom set reported: scrubbing still works, the container still animates in, the scene timelines are dead until a full reload.

Fix

Resume exactly the recorded ids (tl.paused(false)) before the player re-seek — timelines the drag didn't pause are never touched, so a deliberately-paused timeline stays paused.

Verified

• 2 regression tests (unpauses recorded ids only, leaves un-recorded timelines alone, clears the attribute, still re-seeks; no-op without the attribute) — 16/16 in manualOffsetDrag.test.ts.
• Live end-to-end with a rebuilt studio: real drag gesture on an animated element in a 4-scene GSAP composition → all scene timelines remain paused:false afterward (previously all read paused:true and never recovered).

🤖 Generated with Claude Code

[vanceingalls]

• fix(studio): resume drag-paused timelines instead of only re-seeking #1876 👈 (View in Graphite)
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[vanceingalls]

R1 adversarial review — verdict: LGTM_WITH_NITS (COMMENT, not APPROVE per convention).

Mechanism check — symmetric with the pause side

Pause side (manualOffsetDrag.ts:352-368) does two guarded things: tl.pause() (only if !tl.paused?.()) and records the id list. Resume side (manualOffsetDrag.ts:527-532) now inverts exactly that: tl.paused(false) for each recorded id, then the pre-existing player.seek re-anchors position. Only ids the drag actually paused get resumed — a deliberately-paused-before-drag timeline stays paused. That matches the invariant.

Lifecycle / cross-cut checks

• manualOffsetDrag.ts:512 clears the attribute before the resume loop, so the state machine is idempotent even if the loop throws (it can't — inner try/catch), and a subsequent drag can't observe stale ids.
• Multi-member drag: createManualOffsetDragMember sets the attribute per-input-element, but the !tl.paused?.() guard at L356 means only the first member records ids; subsequent members' elements get no attribute. restoreManualOffsetDragMembers/endManualOffsetDragMembers call resumeGsapTimelines per member — the members without the attribute early-return at L513. No duplicate resume, no orphan. ✓
• Order (unpause → seek) is correct: paused(false) doesn't reset time, and seek(t) then repositions from the current player time. Reversed order would produce the same visual result but risks a one-frame flash of the frozen position.
• Element identity: member.element is the ref captured at drag-start. Even if the DOM node is detached before resume, ownerDocument.defaultView still resolves and __timelines lookup is by id — the window global — not the element. No mid-drag element-identity trap on this axis.

Tests — semantics, not presence

manualOffsetDrag.test.ts:365-402 asserts the resume outcome (paused state of recorded ids flips to false, non-recorded id's state is untouched, attribute is cleared, seek was called with the player's current time). That's the right shape — mutation-escapable-presence assertions would have missed the original bug where the attribute was cleared but nothing else fired. Good.

Nits (non-blocking)

1. Type-signature asymmetry. manualOffsetDrag.ts:344 types the pause-side as pause?: () => void; paused?: () => boolean (nullary getter); L516 types the resume-side as paused?: (value?: boolean) => boolean (getter/setter overload). Both match GSAP's real API but two files diverging on the same primitive invites future drift. Would unify into a shared `GsapTimelineHandle` type near the top of the file.
2. No wire-through test. The two added tests exercise resumeGsapTimelines in isolation via a manually-set attribute — they don't drive createManualOffsetDragMember → attribute → resume in one pass. A single integration-shape test that pauses via the real drag-start path and asserts the resume unpauses exactly those ids would pin the pause/resume contract at both ends and catch a future refactor that renames the attribute on only one side. Not blocking; the isolation test is what would have caught the original bug.
3. Comma-delimited attribute value (line 365 join / 527 split) assumes no commas in timeline ids. GSAP composition ids in HF don't contain commas today, but if that ever changes the split silently drops resumes. Pre-existing, not this PR's job — flag for the followup.
4. `cross-origin guard` comment on the resume-side try/catch is inherited from the pause path — the runtime iframe is same-origin so it never fires in practice. Fine to leave.

Body claim vs shipped diff

Body says "2 regression tests" — 2 tests added ✓. "16/16 in manualOffsetDrag.test.ts" and "live end-to-end with a rebuilt studio" are un-artifactable in the diff but the mechanism-level assertions are convincing.

CI: 39 required checks green at `136b5e4`. No prior reviews.

— Via

[terencecho]

Reviewed at 136b5e4.

Fix is symmetric with the pause side (manualOffsetDrag.ts:352-368 records ids after guarded tl.pause(); the resume now loops those same ids with tl.paused(false) before the pre-existing player-seek). Only drag-paused ids get resumed — a deliberately-paused-before-drag timeline is untouched. Attribute is cleared before the loop, so state is idempotent. Two added tests (manualOffsetDrag.test.ts:365-402) pin the outcome: recorded ids unpause, non-recorded main stays paused, attribute cleared, seek fires with the player time.

Type refinement on __timelines[id] from pause?: () => void to paused?: (value?: boolean) => boolean matches GSAP's real getter/setter overload.

All required checks green. LGTM.

[miga-heygen]

Approve. Independent read. Symmetric with the pause side — unpauses exactly the timelines the drag start paused (reads from data-hf-drag-paused-timelines, iterates + paused(false)), then re-seeks the player. Cross-origin try/catch guard is correct. Test pins the selective-resume invariant (main stays paused, drag-paused timelines resume) and the no-attribute no-op. Clean.

— Miga