chore(deps): @babylonjs 9.16.0 + manifold-3d 3.5.1

[RaananW]

🛑 BLOCKED / PARKED — do not merge, do not recapture goldens

Decision: HOLD the entire bump. Kept as a draft until upstream Babylon.js fixes the reference-engine regressions below. No goldens were recaptured and no ceilings were changed. Bundle-size is green and Lite's own engine code is unchanged — the only thing blocking is the reference engine (what Lite is diffed against) changing behavior.

Update — bumped 9.15.0 → 9.16.0 and re-tested. 9.16 (plus the harness stopgap side-effect imports committed here) resolved the original 10 render-blocked scenes (WebGPU shadow sampler scene4/18/22/66/72/140, .babylon activeCamera scene24/145, GS-depth scene127/128 — all now pass in isolation). However, a full isolated 9.16 parity run reveals the pure-split blast radius is much wider than the originally-identified 12: 24 reference-engine scenes still fail, all on the same root cause (the 9.15/9.16 pure/non-pure tree-shaking split dropping transitive side-effect registrations). Each was re-run in its own fresh browser to rule out REUSE_BROWSER cascade — 24 genuinely fail, only scene221 passes.

Blocking scenes — 24 failures under 9.16, grouped by root cause:

• glTF file-loader plugin not registered ×7 — reference glTF scene never renders (canvas never ready, 60s capture timeout): scene30, scene33, scene176, scene212, scene241, scene244, scene253
• Shadow registrations dropped ×5 — scene65, scene116 (MAD 0.29), scene141, scene214, scene215
• Frame-graph post-process blocks not registered ×3 — scene142, scene143, scene144
• Thin instances ×2 — scene17, scene165 (MAD 27.8)
• .babylon Sponza (activeCamera-null family) ×1 — scene9 (MAD 73.7)
• Character controller / moving platform ×2 — scene104, scene105
• Composite gizmos ×1 — scene222 (MAD 11.2)
• RTT override ×1 — scene110
• GS GPU picking ×1 — scene129
• PBR clearcoat (known intended 9.16 change) ×1 — scene19

Revive criteria: upstream Babylon.js restores the dropped side-effect registrations in its non-pure entry points (glTF glTFFileLoader, shadow, frame-graph post-process, gizmo) so format/feature entry points transitively trigger them again. Making 9.16 green from the Lite side alone would require adding fragile stopgap side-effect imports across all ~24 scenes — not a durable fix. Once upstream lands the fix: recapture the golden-less scenes & re-run the full parity suite under the isolated lock on LAB_TEST_PORT=5379.

---

Summary

High-risk runtime / reference dependency bumps (Group 3 of 3 parallel dep updates).

Package	From	To	Where
@babylonjs/core	9.5.0	9.16.0	lab
@babylonjs/loaders	9.5.0	9.16.0	lab
@babylonjs/materials	9.10.1	9.16.0	lab
@babylonjs/addons	9.8.0	9.16.0	lab
babylonjs-gltf2interface	9.5.0	9.16.0	lab
@babylonjs/havok	1.3.12	1.3.13	lab
manifold-3d	3.4.0	3.5.1	babylon-lite runtime + lab devDep

⚠️ This touches the parity REFERENCE engine (@babylonjs/core is what Babylon-Lite is diffed against) and a Lite runtime dep (manifold-3d, bundle-size sensitive). All @babylonjs/* kept on a single 9.16.0 version to avoid mismatches. Touches pnpm-lock.yaml — expect conflicts with the sibling tooling / TS6-Vite8 dep PRs.

Root cause — Babylon.js pure/non-pure tree-shaking split

9.15/9.16 split each module into a side-effect-free .pure.js plus a thin non-pure .js that applies the Register*() side effect. Aggregate barrels / format entry points now resolve the .pure variant, so legacy static registrations no longer arrive transitively. Example (proven by diffing 9.5.0 vs 9.16.0 loaders): loading a .gltf needs two registrations — the 2.0 parser factory (still triggered by glTF/2.0) and the SceneLoader .gltf/.glb file-loader plugin (moved into non-pure glTF/glTFFileLoader.js, no longer pulled by glTF/2.0). In 9.5.0 glTF/2.0/glTFLoader.js statically imported ../glTFFileLoader.js, which registered the plugin at module load; 9.16 decouples them.

Harness fixes included in this branch (behavior-preserving)

• acquireBjsPage() REUSE_BROWSER data-capture reuse — the 20 physics/picking/gizmo specs now reuse the worker-owned context/page under REUSE_BROWSER instead of popping a window per spec.
• 10 stopgap side-effect imports restoring the originally-blocked scenes under 9.16: scene4/18/22/66/72/140 (shadow texture), scene24/145 (parser), scene127/128 (WGSL depth shaders). Temporary until upstream restores registrations.

Bundle size — ✅ GREEN

pnpm build:bundle-scenes regenerated the per-scene manifests (committed; aggregate manifest.json stays gitignored). Counted rawKB/gzipKB unchanged; manifold 3.5.1 wasm growth lands in ignoredRawKB, so no ceiling is affected. bundle-size.spec.ts passes. No ceilings changed.

Guardrails honored

• ❌ No golden reference screenshots edited (6 committed goldens that got regenerated during runs were reverted; working tree clean).
• ❌ No bundle-size ceilings raised.
• ✅ Per-scene manifests committed; aggregate manifest.json left gitignored.
• 🚫 pnpm test:perf not run (agent-forbidden).

Decision

The failures are upstream reference-engine changes on golden-less scenes (pure-split registration drops), not Lite regressions. User decision: HOLD — wait for upstream Babylon.js to restore the dropped registrations in its non-pure entry points, then recapture goldens and re-run parity. Parked as draft.

[bjsplat]

Lab - Static Site

Open deployed site

_{Build 20260707.8 - merge @ f312de1}

[bjsplat]

📋 parity — Test Report

View full Playwright report

_{Build 20260707.8 · merge @ f312de1}

[bjsplat]

📋 parity — Test Report

View full Playwright report

_{Build 20260709.8 · merge @ 2e0de5f}

[bjsplat]

Lab - Static Site

Open deployed site

_{Build 20260709.11 - merge @ 24ff2c2}

[bjsplat]

📋 parity — Test Report

View full Playwright report

_{Build 20260709.11 · merge @ 24ff2c2}