[v11] Triage all 51 open PRs for validity, currency, duplication and safety

[DennisSmolek]

Problem

51 open PRs. The queue is not junk, but a meaningful share is machine-generated, duplicated, or mis-scoped, and none of it has been evaluated against v11-working. Merging blind is a real risk:

• 10 dependabot PRs, reopened 31 Aug — typescript 5.6.3 → 7.0.2 (fix: bump typescript from 5.6.3 to 7.0.2 #2796), vite 5.4 → 8.2 (fix: bump vite from 5.4.11 to 8.2.2 #2797), rollup 2.79 → 4.63 (fix: bump rollup from 2.79.2 to 4.63.1 in the rollup-packages group #2791), @babel/plugin-transform-runtime 7 → 8 (fix: bump @babel/plugin-transform-runtime from 7.25.9 to 8.0.1 #2798). They are titled fix:, so merging any of them to a release branch cuts a release.
• Switch from semantic-release to changesets #2607 (app/copilot-swe-agent) — replaces semantic-release with changesets, +644/−2900. Would undo the OIDC trusted-publishing fix that shipped in alpha.6.
• 9 PRs from one author in May 2026 (fix(Select): onChange now fires for single-click selections #2726–fix: make Preload use async compilation #2737) mapping one-to-one onto open issues, one titled [codex]. Generated against the issue list.
• fix(StatsGl): update stats-gl to ^4.2.3 and fix hidden panel #2754 — titled a StatsGl fix, carries +7443/−13521 across 3 files.
• Security: Global loading error list can grow unbounded (memory DoS risk) #2721 — a one-line change framed as a memory-DoS security fix.
• Duplicate clusters: Preload [codex] Fix Preload async compilation coverage #2733 / fix: make Preload use async compilation #2737 / fix(Preload): compileAsync, frustum-culled objects, onComplete signal #2755 · Html occlusion fix(Html): update occlusion when occlude prop changes #2703 / fix: make Html occlusion material transparent #2736 / fix(Html): recompute occlusion when the occlude prop changes #2757 · TrackballControls fix(TrackballControls): invalidate on user input with demand frameloops #2746 / fix(TrackballControls): work under frameloop="demand" #2756 · GizmoHelper Fix GizmoHelper ViewCube radius to use focus point #2729 / fix(GizmoHelper): land tween exactly on the target axis #2763.

Task

One agent pass per PR. For each, write a comment and apply a label covering four tests:

1. Validity — does the problem still reproduce on v11-working?
2. Currency — v11 moved every component into a Component-as-a-Folder layout. Most of these target master's flat src/core/*.tsx. A clean-looking diff may not apply at all.
3. Duplication — is another open PR fixing this better?
4. Safety — does the diff match its title? Does it touch release config, CI, lockfiles, or dist/package.json? Is a "security fix" actually one?

Verdict, one of: merge-as-is · rebase-onto-v11 · superseded-by-#N · close-stale · needs-human.

Hard rule

The agent closes nothing and merges nothing. It labels and comments. A human executes.

Acceptance

Every open PR carries a verdict label and a comment justifying it. The four duplicate clusters are collapsed to one recommendation each.

[DennisSmolek]

Triage pass 1 — currency, duplication, safety

Machine-checkable parts done across all 51 open PRs. No PR has been commented on or labelled yet.

Currency: 24 of 51 PRs touch only files v11 does not have

Every open PR targets master. v11 moved every component into a Component-as-a-Folder layout, so most diffs reference paths that no longer exist:

• 24 PRs touch only paths missing from v11-working
• 34 PRs touch at least one missing path

This is not a reason to close them. master merges into v11-working, not the reverse — so the right move for almost all of these is merge on master and let the next merge carry them into v11. Rebasing them onto v11 would be strictly more work for the same result.

Where the stranded files went:

PR path (master)	v11 location
src/core/Preload.tsx	src/core/Loaders/Preload/Preload.tsx
src/web/Html.tsx	src/core/UI/Html/Html.tsx
src/web/Select.tsx	src/core/UI/Select/Select.tsx
src/web/KeyboardControls.tsx	src/core/Controls/KeyboardControls/KeyboardControls.tsx
src/web/pivotControls/index.tsx	src/core/Gizmos/PivotControls/index.tsx
src/core/Svg.tsx	src/core/Helpers/Svg/Svg.tsx
src/core/GizmoHelper.tsx	src/core/Gizmos/gizmo/GizmoHelper.tsx
src/core/TransformControls.tsx	src/core/Gizmos/TransformControls/TransformControls.tsx
src/core/TrackballControls.tsx	src/core/Controls/TrackballControls/TrackballControls.tsx
src/core/SpriteAnimator.tsx	src/core/Helpers/SpriteAnimator/SpriteAnimator.tsx
src/core/VideoTexture.tsx	src/core/Textures/VideoTexture/VideoTexture.tsx
src/core/Texture.tsx	src/core/Loaders/useTexture/useTexture.tsx
src/core/Gltf.tsx	src/core/Loaders/useGLTF/useGLTF.tsx
src/core/Progress.tsx	src/core/Loaders/useProgress/useProgress.tsx
src/core/Backdrop.tsx	src/core/Staging/Backdrop/Backdrop.tsx
src/core/StatsGl.tsx	src/core/Performance/StatsGl/StatsGl.tsx

⚠️ Three of them now exist in two trees, so the fix has to be applied twice when it reaches v11:

• src/core/MeshTransmissionMaterial.tsx → legacy/Materials/… and webgpu/Materials/… (fix(MeshTransmissionMaterial): support stencil buffer in render targets #2748)
• src/core/GradientTexture.tsx → legacy/Textures/… and webgpu/Textures/… (fix(GradientTexture): regenerate texture when any gradient input changes #2750)
• src/core/ContactShadows.tsx → legacy/Materials/… and webgpu/Staging/… (fix: don't break ContactShadows when moved #2758)

Duplication: 15 PRs are 7 fixes

File	Competing PRs
src/core/Preload.tsx	#2755 #2737 #2733
src/web/Html.tsx	#2757 #2736 #2703
src/core/GizmoHelper.tsx	#2763 #2729
src/core/TrackballControls.tsx	#2756 #2746
src/core/VideoTexture.tsx	#2732 #2714
docs/abstractions/text.mdx	#2738 #2715
docs/performances/preload.mdx	#2755 #2733

Safety

Real risk:

• 10 dependabot PRs (fix: bump the react-packages group with 2 updates #2790–fix: bump @vitejs/plugin-react from 4.3.3 to 6.1.1 #2799), all titled fix:. Merging any one to a release branch cuts a release. They include majors: typescript 5.6.3 → 7.0.2, vite 5.4 → 8.2, rollup 2.79 → 4.63, @babel/plugin-transform-runtime 7 → 8. These need a human and probably a batched branch, not ten individual merges.
• Switch from semantic-release to changesets #2607 (app/copilot-swe-agent) — replaces semantic-release with changesets, +644/−2900, touching release.config.js, .github/workflows, package.json, yarn.lock. It would undo the OIDC trusted-publishing fix that shipped in alpha.6. Needs an explicit decision, not a merge.

Looked risky, isn't — I checked the diffs:

• fix(SpriteAnimator): load image-only spritesheets #2734 touches .github/workflows/release.yml in what is titled a SpriteAnimator fix. The change is node-version-file: '.nvmrc' — correct, and it is scope creep rather than anything worse. v11-working already has it in all three jobs; master has it in 2 of 3.
• Security: Global loading error list can grow unbounded (memory DoS risk) #2721, framed as a "memory DoS risk", is [...state.errors, item].slice(-100). Capping an unbounded array is a legitimate small fix. The severity framing is overblown; the change is sound.
• fix(StatsGl): update stats-gl to ^4.2.3 and fix hidden panel #2754 carries +7443/−13521 because of a lockfile, not logic. Note its dependency half is already done — alpha.6 bumped stats-gl to ^4.2.3 — so only the hidden-panel fix is still relevant.

What is left for pass 2

Validity — whether each fix still reproduces — cannot be determined by script. It needs the diff read against the current source, one PR at a time. That is the remaining work here.

Recommendation

1. Handle the 10 dependabot PRs as one batch on a branch, not ten merges. Watch the fix: prefixes.
2. Decide Switch from semantic-release to changesets #2607 explicitly (recommend: close — alpha.6 just made trusted publishing work).
3. Pick one winner per duplicate cluster; close the rest as superseded-by-#N.
4. Merge the survivors on master. They reach v11 via the next merge.

[DennisSmolek]

Triage pass 2 — validity

Diffs read against current v11-working source. Still no comments or labels on any PR.

Correction to pass 1: file-based clustering over-counted duplicates

Grouping by touched file was too coarse. Three of the "clusters" are PRs fixing different bugs in the same file:

Pass 1 said	Actually
src/web/Html.tsx ×3	#2703 + #2757 are the same bug (occlude prop change). #2736 is a different bug — the transparent flag, i.e. #2352.
src/core/GizmoHelper.tsx ×2	#2729 fixes the orbit radius; #2763 fixes tween landing. Different, possibly complementary.
src/core/VideoTexture.tsx ×2	#2714 is unmount cleanup; #2732 is lazy-loading hls. Unrelated.

Real duplicates: Preload ×3, Html-occlude ×2, TrackballControls ×2, docs typo ×2. Eight PRs, not fifteen.

Already fixed on v11 — close these

• fix: lazy load hls events for video textures #2732 (lazy-load hls events) — src/core/Textures/VideoTexture/VideoTexture.tsx already imports hls.js type-only and reaches Events through the lazily-loaded _HLSModule. Done.
• fix(StatsGl): update stats-gl to ^4.2.3 and fix hidden panel #2754 — its dependency half is done; alpha.6 bumped stats-gl to ^4.2.3. Only the hidden-panel fix remains, and the PR's +7443/−13521 is almost entirely lockfile.

A v11-only bug this surfaced — Preload is broken on WebGPU

Preload lives in core/ and calls gl.compile(scene, camera) synchronously, then immediately restores visibility. On WebGL that is correct — compile() returns Set<Material>.

On WebGPU it is not. In @types/three 0.185, Renderer.compile is:

get compile(): (scene: Object3D, camera: Camera, targetScene?: Scene | null) => Promise<void>

A getter aliasing compileAsync. So on the /webgpu entry Preload kicks off async compilation and then flips visible back and disposes the cube target before it finishes. The all prop silently does nothing.

This makes all three Preload PRs more necessary on v11 than on master. Worth its own issue regardless of which PR wins.

Verified still-valid (bug read in current v11 source)

PR	Bug	v11 evidence
#2733 #2737 #2755	Preload sync compile, ignores frustum-culled, no completion signal	Preload.tsx:42 still gl.compile
#2703 #2757	occlusion not recomputed when occlude changes	no prevOcclude/invalidate in Html.tsx
#2736	occlusion plane not transparent	legacy/Materials/HtmlMaterial.tsx:53 lacks it — note the WebGPU one already sets mat.transparent = true, so this is a legacy/webgpu parity gap
#2729	ViewCube radius measured to target, not focus point	gizmo/GizmoHelper.tsx:112
#2763	tween does not land exactly on axis	no applyTweenRotation in v11
#2746 #2756	TrackballControls dead under frameloop="demand"	TrackballControls.tsx:53-67 invalidates only on change, which update() alone dispatches — the chain cannot start itself
#2696	onLoad receives the unmapped array	useTexture.tsx:38 is onLoad?.(textures)
#2728	user renderOrder on <Svg> ignored	Svg.tsx:64 let renderOrder = 0, prop not accepted
#2751	Backdrop props untyped	Backdrop.tsx:24 has no annotation though BackdropProps exists at line 7

Intent valid, implementation does not port

• fix: lazy load MeshOpt decoder in useGLTF #2731 (lazy MeshOpt) — v11 imports MeshoptDecoder from three/addons/... at module scope, not from three-stdlib. The eager-import problem is real; the patch is written against the old import and would need rewriting.

Heuristic only — not individually verified

#2726 #2727 #2735 #2750 #2758 #2767 #2721. Pattern-matching says the fix is absent from v11, but I did not read each diff against the source. Treat as unconfirmed — the same heuristic produced four false "already fixed" results (#2696, #2728, #2731, #2751) that were wrong on inspection.

Recommended winners

Cluster	Winner	Why
Preload	#2755	Its author already knows compile() is a compileAsync alias on WebGPU and says so in a comment. Uses a ref so the callback cannot go stale. Needs #2737's disposed guard added — as written it mutates objects and fires the callback after unmount.
Html occlude	#2757	Shallow-compares arrays and calls invalidate(). #2703 compares by identity, so an inline occlude={[ref]} array is a new object every render and it would force a recompute every frame.
TrackballControls	#2756	Same fix as #2746, simpler, and the comment explains the bootstrap problem correctly.

Everything above is a merge onto master. They reach v11 at the next merge — except #2736, which needs applying to legacy/ on v11 specifically, and the Preload winner, which wants the WebGPU race handled.