Raycasters.castRay / FastModelPicker returns only a subset of models in multi-model scenes (3.4 regression, still in 3.4.6)

[amirmiz]

Describe the bug 📝

When several top-level fragment models are loaded into one world, components.get(OBC.Raycasters).get(world).castRay() — the GPU id-render pick (FastModelPicker / getFullPick) — only ever resolves items belonging to a fixed subset of the loaded models. Every other model renders normally but is completely unpickable: Highlighter selection, the measurement snapping picker (GraphicVertexPicker), and anything built on castRay silently stop working on those models.

The same scenes were fully pickable on 3.2.x. FragmentsManager.raycast() (worker BVH) resolves all models correctly at the same pointer position, so geometry/models are fine — only the GPU pick path is affected. This looks like the flat-multi-model counterpart of #746 (same FastModelPicker.renderPickPass() logic) and part of the #737/#747 cluster. It is not #747 (null return) — ours returns a valid hit for the "winning" subset — and we already run 3.4.6 (which includes the #747/#736 fixes), yet it persists.

Reproduction ▶️

No response

Steps to reproduce 🔢

1. Standard Components + World + PostproductionRenderer + OrthoPerspectiveCamera.
2. fragments.init(workerUrl), load 3+ top-level models as flat roots (fragments.core.load + world.scene.three.add(model.object)); wait for streaming.
3. Sweep the pointer over the viewport; per position await raycaster.castRay({ position }) and record hit?.fragments?.modelId.
4. Also fit the camera to each model individually and pick its centre.

Expected ✅ — every visible model is pickable; fitting to a model and clicking selects that model.

Actual ❌ — only a fixed subset (in a 12-model scene, ~3) is ever returned, even when an "unpickable" model fills the screen. The subset is stable per scene.

Debug findings (instrumented FastModelPicker)

• applyIdMaterial() builds a complete byte → modelId map (1..N) — every model is registered and gets the id material.
• Yet the decoded (modelId, itemId) only surfaces the fixed subset → the defect is in the id-render/decode (renderPickPass / runIdPass), not registration. This is the same renderPickPass() per-model-visibility logic Delta models become unpickable after editor.edit() because FastModelPicker hides parent model root #746 dissects; our models are already flat independent roots, so Delta models become unpickable after editor.edit() because FastModelPicker hides parent model root #746's delta fix wouldn't cover this.
• Ruled out: missing id attributes (an unpickable model had 69 id-meshes), MAX_MODELS (254), LOD (models fully streamed/visible), coordinate magnitude (near origin).

Workaround — delegate castRay to FragmentsManager.raycast({ camera, mouse: raycaster.mouse.rawPosition, dom, snappingClasses }); its result is shape-compatible (carries point/normal/distance/localId/itemId/fragments/object/snappingClass/snappedEdgeP1P2/facePoints/faceIndices) and resolves all models. Can share a minimal standalone repro (2–3 small .frags).

Related: #746 (same method, delta variant), #737 (Highlighter, open), #747 (null return, fixed 3.4.6 — distinct), #736 (mobile castRay null, fixed 3.4.4).

System Info 💻

components 3.4.6
components-front 3.4.3
fragments 3.4.6, three 0.182.0
camera-controls 3.1.2
web-ifc 0.0.77
Windows/Chrome desktop

Used Package Manager 📦

npm

Error Trace/Logs 📃

No response

Validations ✅

• Read the docs.
• Check that there isn't already an issue that reports the same bug to avoid creating a duplicate.
• Make sure this is a repository issue and not a framework-specific issue. For example, if it's a THREE.js related bug, it should likely be reported to mrdoob/threejs instead.
• Check that this is a concrete bug. For Q&A join our Community.
• The provided reproduction is a minimal reproducible example of the bug.

[agviegas]

Thanks, this is a great report, and the instrumentation helped a lot.

The picker path you fingered (renderPickPass / applyIdMaterial) was reworked on main as part of the #746 fix, which isn't published yet. The 3.4.6 version you're on isolated each model by toggling its whole root's visibility; the new one isolates each model by its own meshes and renders them into a shared depth buffer. I set up a 6-model scene of separate flat roots on main and every one picks correctly, a distinct model byte per model, so the subset behavior doesn't reproduce there.

So the next release should sort this out. Retest it once it's out and let us know, and if it's still off we'll reopen and dig in with the repro frags you offered.