[jonit-dev]

What is this PR about?

It moves every package, template and generated game from three.js 0.185.1 to 0.187.0. It rebases the ThreeNative patch onto the new release, and ports the r186/r187 rendering changes into the owned C++ engine, so the native and Wasm engines still match three pixel for pixel.

Why does it matter?

Games get two releases of upstream fixes, such as the ACES fit, energy-conserving lighting, PMREM cube targets and watertight raycasts. The patch stops drifting from three.js.

Done

• ✅ Phase 1 (AC-1..3): every pin is on 0.187.0, the patch is rebased hunk by hunk (ledger below), and the installer specs are green.

• ✅ Native engine ported to r187: lighting (shared DFG, multi-scatter, roughness floors), TRAA, PCF shadows, MaterialX noise, the watertight ray, the lathe normal, GTAO, the bloom tap merge, the sRGB build exponent, the denoise precision and the PMREM cube mip chain. ctest -L native-engine: 527/527.

• ✅ Real bugs found and fixed on the way:

  • The patch's async PMREM bake bound a 1-layer placeholder as a cube. Dawn dropped the whole command buffer, so every environment-lit scene drew black until the bake landed.
  • The native host lacked GPURenderBundleEncoder (ReferenceError on every draw).
  • Node building before renderer.init() crashed.
  • The procedural-animals probe leaked its storage buffers.
  • The starter's TRAA used the r185 view-offset hand-off, and reported one extra reset frame on r187.
  • The engine's compute warm-up passed a dispatch size that r187 refuses (Renderer.js:3238), so ocean and fluid warm-ups compiled nothing.

• ✅ Dropped the patch's transient output-depth hunk (−1208 patch lines). r187 draws the RenderPipeline output at one sample, so the hunk never ran.

• ✅ pnpm parity vs develop 6bf21eab2: every red this PR added is fixed. The remaining reds (web temporal-aa-scaled, -lifecycle, vfx-spawn-requests, realism-velocity; desktop temporal-aa, realism-*) fail the same way on develop.

• ✅ The reference capture now fails closed on a WebGPU validation error and on a single-colour frame.

• ✅ Phases 2 and 3 (AC-4..8): typecheck, lint, unit tests and parity are green; every kit plays from spawn; the visual judge returns NEUTRAL or IMPROVEMENT for every template; perf is NEUTRAL on web and native desktop.

• ✅ AC-8 perf (presented p95, 5 paired runs per lane, RTX 2080): Midway web 16.68 / 16.68 ms, stress lab web 16.68 / 16.68 ms, Midway desktop 18.15 / 18.41 ms (+1.4 %), stress lab desktop 17.07 / 17.02 ms. All NEUTRAL. pnpm native:verify:desktop exit 0.

• ✅ CI on c38dcde74 (run 41248394799): 65 jobs pass. CI found 7 PR-only reds over three runs, all fixed:

  • exposure warm-up and cold-boot proofs reject three's first node frame at time 0; velocity's no-history control is exact on r187 (8ee2f524b)
  • 3 parity rows split by SwiftShader vs llvmpipe on r187's half-float DFG LUT now need a hardware adapter; NVIDIA native = web, 0 px (3d50a8dbc)
  • a lint waiver (8e7cc84ee), the native census (4826787f1), and sailing's wind check on its 120 s float boundary (c38dcde74)

CI

• ✅ Full board green on c3f5598f7, run 41288721287, first attempt: 89 jobs pass, 4 skipped by design (queue-watch and 3 V8 payload jobs), ci-required pass.

• ✅ The two commits after that run, 2c6954d6e and 4f0a79fad, change only the PRD and the generated Unreal port status files. pnpm check:docs passes on both, and the review lane passes on 4f0a79fad with ci-required.

• ✅ The shooter engine-cost timeout on run 41265123694 is hosted-VM variance, not r187. A local paired 1260-frame row (4 cores, SwiftShader) read 843 and 1464 s on base against 799 and 861 s on r187. The leg took 17 minutes in the green run.

• ⚠️ Local rerun on the merged tree (develop 0.3.9), one suite at a time. shooter performance peaks at 128 draw calls, down from 731. It still misses its frame limit at 17.9 ms (GPU). Develop misses it too, at 18.3 to 19.5 ms on the same machine.

• ⚠️ Head fc0a73183 has no full board of its own. The full run above is on the older head c3f5598f7. The merge queue runs the board.

Notes for review

• ⚠️ Look changes on purpose, for the AC-7 judge: the r187 ACES fit, the lighting energy rewrite, PCFSoftShadowMap → PCFShadowMap, and r187 GTAO for the 4 templates that author gtaoRadius (action-rpg, rts, runner, shooter). The other 12 templates use the engine's reachPx search, which is unchanged.
• ✅ The AO denoise pass reads a one-channel field, so it now targets R32Float (099ce70f8); RGBA32Float had cost +0.87 ms GPU on the stress lab.
• ⚠️ packages/core/src/gpu-frame-observation.ts now wraps r187's self-resolving query pool. It is no longer read-only, so please review it.
• ⚠️ PRD number collision: develop also filed docs/PRDs/Unreal/shadows/PRD-775-virtual-shadow-pages-… (another lane). This PR's PRD is the tech-debt-code-quality one.
• Pre-existing on develop 6bf21eab2, not this PR: native_engine_player_gpu_stages (same failure on the base build).
• Follow-up: core's batched-velocity scheduler may be redundant now that three r187 keeps BatchedMesh history itself (Batch.js:78).

Changed goldens (AC-5): one upstream reason each

One or more upstream causes necessarily move each frame's hash, so a group row is one reason per fixture.
diff data: 94 fixtures differ only in render.pngSha256; exactly one fixture differs in a number.

fixture(s)	what changed	upstream file:line + reason
78 render fixtures with toneMapping: "aces" — all 78 aces render fixtures, e.g. lit-render, batched-mesh, fog-*, gltf-model-*, materials-*, morph-*, nodemat-*, particles-sprite, pbr-grid-envmap, pmrem-texture, probes*, reflector-plane, screen-uv, shadows-*, skinned-*, sky-equirect, storage-attribute, texture-object, textures-data-3d, tonemap-ramp-aces, tsl-*, viewport-textures, vsm-*, fluid-particles, lights-point-spot, materialx-noise	pixel hash, every aces frame	src/nodes/display/ToneMappingFunctions.js:69 (r187): b = color*(color*0.983729 + 0.4329510) + 0.238081; r185 was color*(color + 0.4329510)*0.983729 + 0.238081. Upstream: r187 "Fix ACES Filmic fit denominator" #35110. Control: tonemap-ramp-aces moved (strip 1 227→226) while tonemap-ramp-{agx,cineon,linear,neutral,reinhard} and their fixtures did not, so the ACES curve is the only tone-curve change.
12 non-aces fixtures: gltf-model-specular, material-ao-environment, material-normal-map, material-pbr-maps, materials-bump-map, materials-force-single-pass, materials-metal-rough-maps, materials-vertex-colors, materials-vertex-node, textures-data-float, textures-data-half, textures-data-rgba8	pixel hash	src/nodes/functions/PhysicalLightingModel.js:329 (r187): direct specular is now irradiance.mul(specularBRDF).mul(multiScatteringCompensation), replacing r185's BRDF_GGX_Multiscatter(… f90: 1 …); indirect at :879-895. Upstream: r186 removed BRDF_GGX_Multiscatter and re-derived multi-scattering energy (#33983, #33985, #34046, #34055), r187 "share dielectric scattering in indirect lighting" #34656, "fix specular intensity in direct lighting" #34764. Every one of these 12 is lit; every non-aces fixture that did not change (alpha-test, alpha-transparency, blending-additive, gltf-model-webp, lines-basic, textures-anisotropy, textures-data-{canvas-mips,filters,wraps}) has no light and no shading material.
pmrem-hdr-equirect, textures-data-equirect	pixel hash	src/extras/PMREMGenerator.js (r187): the CubeUV atlas is gone, replaced by CubeCamera + cube render targets (MIN_SIZE = 256, src/extras/PMREMGenerator.js:31) — r187 "Replace the CubeUV atlas with cube render targets" #34585, plus r186 "spiral blur" #32367 and r187 "restore minimal roughness, improve roughness ramp" #34645. Both fixtures are unlit metal spheres lit only by the environment.
history-cut, traa-history	pixel hash	examples/jsm/tsl/display/TRAANode.js:237 (r187): the previous-depth node now samples this._historyRenderTarget.depthTexture instead of its own DepthTexture(1, 1); :330: the jitter wraps over all 32 offsets (% length) where r185 wrapped at length - 1. Same file also drops the post-processing view-offset sync and pre-initialises its targets on resize.
geometry-lathe	value: last path point's normal, 25 of 375 components 0.035999998450279236 → 1 (muzzle), 18 of 54 (partial)	src/geometries/LatheGeometry.js:106 (r187) adds prevNormal.normalize() before the last vertex's normal is pushed; r185 pushed it unnormalised. Upstream: r187 "Normalize the normal of the last point" #35153.

Subsystem-specific changes that ride inside the groups above and are the reason a reviewer will see those rows move:
tsl-post-smaa (r186 SMAANode fix/simplification #34031), tsl-post-bloom/tsl-post-template-high
(r186 BloomNode blur strategy #33923), tsl-post-ao/tsl-post-ao-raw (r186 GTAONode quadratic stepping,
clip-space march, banding, baked sample count #33895/#33903/#33919/#33931), shadows-*/vsm-* (r187
ShadowFilterNode PCF radius #35018 and per-shadow uniforms #35047, PointShadowNode #34899/#35186,
ShadowNode #34801/#35093), screen-uv (r186 ScreenNode DPR #34186), particles-sprite (r187
SpriteNodeMaterial centre #35064), materialx-noise (r187 RangeNode hash/random-order #34897/#34920),
reflector-plane (r187 ReflectorNode view offset #34917, bounces #34977), lit-render-scaled/
tsl-post-resized (r186 PassNode pixel ratio #34453).

No diff without an upstream reason. Nothing else in the corpus moved: all 97 non-render goldens except
geometry-lathe are byte-identical.

Re-recorded after the PMREM patch fix: pmrem-hdr-equirect, pmrem-texture, textures-data-equirect, sky-equirect, material-ao-environment and pbr-grid-envmap were first recorded as solid black frames, the dropped command buffer described above. They are now real frames (1482 to 25621 colours). pmrem-hdr-equirect is byte-identical to a capture on stock r187. test:reference --check: 205 fixtures, 0 problems.

Dropped patch hunks (AC-2): reason per hunk

The ledger has 153 rows: 82 kept, 33 ported and 38 dropped. The full per-hunk ledger, kept and ported rows included, is in the two PR comments below.

hunk	verdict	what it did	reason
NodeMaterialObserver.js #3	dropped-upstream	drop _equal from getGeometryData	r187 replaced the flag with a counter: _version: 0 in the same literal (src/materials/nodes/manager/NodeMaterialObserver.js:420 r187) plus `geomet
NodeMaterialObserver.js #4	dropped-upstream	drop _equal from getMaterialData	r187 r187:536 data = { _renderId: - 1, _version: 0 }.
NodeMaterialObserver.js #5	dropped-upstream	drop materialData._equal = false before a material diff return	r187 r187:604-686 rewrote the block around changed = true / materialData._version ++, which is this fix done properly.
NodeMaterialObserver.js #6	dropped-upstream	same, vector-valued material property branch	r187 changed = true in the value.equals branch.
NodeMaterialObserver.js #7	dropped-upstream	same, texture-valued material property branch	r187 changed = true + value.cacheVersion/value.transform tracking.
NodeMaterialObserver.js #8	dropped-upstream	same, scalar material property branch	r187 changed = true in the else branch.
NodeMaterialObserver.js #9	dropped-upstream	drop the _equal = true / else if _equal === false early-out so eve	r187 r187:684-694 replaced it with the per-render-object renderObjectData.materialVersion !== materialData._version mismatch, plus an output-target
NodeMaterialObserver.js #10	dropped-upstream	drop geometryData._equal = false, removed-attribute branch	r187 changed = true; continue;.
NodeMaterialObserver.js #11	dropped-upstream	same, changed-attribute branch	r187 changed = true.
NodeMaterialObserver.js #12	dropped-upstream	same, attribute-set changed branch	r187 changed = true.
NodeMaterialObserver.js #13	dropped-upstream	same, index changed branch	r187 changed = true.
NodeMaterialObserver.js #14	dropped-upstream	same, draw-range changed branch	r187 changed = true.
NodeMaterialObserver.js #15	dropped-upstream	drop the geometry _equal = true / else if early-out	r187 r187:805-820 if ( changed === true ) geometryData._version ++; + renderObjectData.geometryVersion !== geometryData._version.
Batch.js #1	dropped-upstream	import positionPrevious for the motion-vector batch transform	r187 src/nodes/accessors/Batch.js:2 imports { positionLocal, positionPrevious } already.
Batch.js #2	dropped-upstream	previousMatricesTexture = `batchMesh._previousMatricesTexture ?? mat	r187 r187:88-108 getPreviousMatricesTexture( batchMesh ) creates the texture from _matricesTexture on first use (and again after BatchedMesh dis
Batch.js #3	dropped-upstream	assign positionPrevious from the previous batching matrix when `need	r187 r187:146-167 does exactly this: createBatchingMatrixNode( getPreviousMatricesTexture( batchMesh ), indirectId ) → `positionPrevious.assign( pre
Batch.js #4	dropped-upstream	drop the two trailing blank lines at EOF	cosmetic only; r187 keeps them. Nothing to port.
Instance.js #3	dropped-obsolete	force the interleaved-attribute path in instance() and track the buf	r187 deleted the whole interleaved tracking: no _matrixBuffers, no interleavedMatrix/interleavedColor locals; the columns are `new InterleavedBu
Instance.js #4	dropped-obsolete	sync the interleaved buffers on OnBeforeFrameUpdate	The block this changed does not exist in r187 (same deletions as row 3); r187 uploads on attribute version change instead.
Skinning.js #1	dropped-upstream	previousBoneMatricesData() — per-skeleton history created at the cur	r187 setupPreviousBoneData( skeleton ) (r187:159-187) does the same, storing on the skeleton (previousBoneMatrices, previousBoneTexture, both de
ShadowFilterNode.js #2	dropped-upstream	PCFShadowFilter uses the shared mapSize / radius	r187:85-86 const { mapSize, radius } = getShadowUniforms( shadow );.
ShadowFilterNode.js #3	dropped-obsolete	PCFSoftShadowFilter uses the shared mapSize	PCFSoftShadowFilter was deleted upstream (r187 src/nodes/lighting/ShadowNode.js:117 keeps null /* PCFSoftShadowMap, removed */ in `_shadowFilter
PMREMNode.js #6	dropped-obsolete	Y flip only for generated PMREMs, read off the live texture (`this._te	The whole Y-flip block and textureCubeUV/_width/_height/_maxMip were removed upstream; r187 samples the generated cube texture directly (`this
RenderObject.js #1	dropped-upstream	Track the version of _sourceMaterial this render object was validate	r187 added the same field for the same reason: this.sourceMaterialVersion = this._sourceMaterial !== null ? this._sourceMaterial.version : 0; (Rende
RenderObject.js #2	dropped-upstream	getChainArray() must include the source material so override and non	r187 returns [ object, material, context, lightsNode, this._sourceMaterial ] when _sourceMaterial !== null, which is the same discriminator (ours
RenderObject.js #3	dropped-upstream	getGeometryCacheKey( useMorphTargetIds ) — a pipeline key must not d	r187 added WebGPUAttributeUtils.getVertexBufferLayoutCacheKey( renderObject ) to the pipeline cache key and removed stride/offset/index from the geo
RenderObjects.js #1	dropped-upstream	Chain key must include the source material (_chainKeys[ 4 ]) so mate	r187 introduced _overrideChainKeys (5 slots, [ null, null, null, null, null ]) used only while an override is active, with `chainKeys[ 4 ] = sourc
RenderObjects.js #2	dropped-upstream	Version gate must answer with the source material's version while draw	r187's gate is `renderObject.version !== material.version \
RenderObjects.js #3	dropped-upstream	Reset _chainKeys[ 4 ] after the lookup	r187 resets chainKeys[ 4 ] only when sourceMaterial !== null, which matches its 4-vs-5 slot scheme
Renderer.js #9	dropped-upstream	Write overrideMaterial._alphaTest so the shared override material's	r187 has the same assignment with the same rationale: `overrideMaterial._alphaTest = material.alphaTest; // bypass the setter so the version doesn't c
Renderer.js #13	dropped-upstream	Run every draw step unconditionally instead of behind `nodes.needsRefr	r187 replaced the boolean with RenderObjectRefreshType (constants.js:1711) and _renderObjectDirect gates FULL / SHARED / NONE (Renderer.js:4
Renderer.js #14	dropped-upstream	Always call nodes.updateAfter( renderObject ) after a draw	r187 calls it for every refresh type except NONE (if ( refreshType !== RenderObjectRefreshType.NONE ) this._nodes.updateAfter( renderObject )) and
WebGPUBackend.js #11	dropped-upstream	Bind-group comparison must compare the native bind group object, not t	r187 does exactly this with one call instead of two: `const bindGroupGPU = this.get( bindGroup ).group; … if ( currentBindingGroups[ i ] !== bindGroup
WebGPUBackend.js #12	dropped-upstream	Store the native bind group object in the per-pass binding cache	as row 11 — currentBindingGroups[ i ] = bindGroupGPU;
WebGPUBackend.js #16	dropped-upstream	Pipeline key uses the structural geometry key without morph-target ids	r187 replaced that entry with utils.getStripIndexFormat( object, geometry ) + this.attributeUtils.getVertexBufferLayoutCacheKey( renderObject ), w
WebGPUBackend.js #20	dropped-obsolete	(hunk touched only a blank line before copyTextureToTexture's mipmap	our hunk adds one empty line and changes no behaviour; r187's guard is `dstLevel === 0 && dstTexture.generateMipmaps === true && dstTexture.mipmapsAut
WGSLNodeBuilder.js #3	dropped-upstream	Do not name uniform buffers with a process-wide allocation id (it chan	r187 removed the very line our hunk rewrote (uniformNode.name = name ? name : 'NodeBuffer_' + uniformNode.id; is gone from getUniformFromNode), so
WebGPUTextureUtils.js #4	dropped-upstream	The canvas colour texture is only reusable when its GPU texture exists	r187 made both changes: if ( colorTextureData.texture !== undefined && colorTexture.width === width && … ) and colorTexture.image.width = width;
WebGPUBackend.js #21 (transient output depth: also Renderer.js, RenderPipeline.js, WebGPU{Attribute,Binding,Pipeline}Utils.js)	dropped-obsolete	delete _beginTransientDepth/_finishTransientDepth/_restoreTransientDepth/_isFullTransientViewport/_validateTransientDepth and the _tnOutputOwner/_tnDepthProof/_tnUniformBytes/_tnFullTriangle full-quad proof	dead on r187: RenderPipeline output is a 1-sample fullscreen pass (Renderer.js:2198, :3004)

PRD: docs/PRDs/done/PRD-775-upgrade-three-to-0.187.md

[jonit-dev]

PRD-775 ARM A — three.js patch port 0.185.1 → 0.187.0

Files: the 15 in this arm's list. Every hunk of git diff base patched -- <file> is one row, in patch order.
Marker = literal in the final file that is absent from pack/r187/package/<file> (verified with grep -F).

---

src/materials/nodes/manager/NodeMaterialObserver.js

#	intent (one line)	origin (sha / PRD)	decision	reason	marker
1	flag on the observer while the current render produces motion vectors	3630847 / f1040a1 (PRD-269, PRD-455)	kept	renderingVelocity field merged clean; r187 has no equivalent (it recomputes needsVelocity() inline and would lose the value updateSettled reads).	this.renderingVelocity = false;
2	isSettled/updateSettled: per-draw "static object and nothing it froze changed" answer, so OBJECT updates, the object UBO upload and the attribute scan can be skipped	3630847 / f1040a1 (PRD-269)	ported	Kept both methods; they still read matrixWorld, material.version, geometry.version and the scene environment, all unchanged in r187. r187 has nothing equivalent (its skip is object.static === true, which is the bug this fixes: a never-drawn static object would skip the draw that creates its bind groups).	updateSettled( renderObject ) {
3	drop _equal from getGeometryData	3630847 (PRD-269)	dropped-upstream	r187 replaced the flag with a counter: _version: 0 in the same literal (src/materials/nodes/manager/NodeMaterialObserver.js:420 r187) plus geometry.addEventListener('dispose', () => data._version ++).	-
4	drop _equal from getMaterialData	3630847 (PRD-269)	dropped-upstream	r187 r187:536 data = { _renderId: - 1, _version: 0 }.	-
5	drop materialData._equal = false before a material diff return	3630847 (PRD-269)	dropped-upstream	r187 r187:604-686 rewrote the block around changed = true / materialData._version ++, which is this fix done properly.	-
6	same, vector-valued material property branch	3630847 (PRD-269)	dropped-upstream	r187 changed = true in the value.equals branch.	-
7	same, texture-valued material property branch	3630847 (PRD-269)	dropped-upstream	r187 changed = true + value.cacheVersion/value.transform tracking.	-
8	same, scalar material property branch	3630847 (PRD-269)	dropped-upstream	r187 changed = true in the else branch.	-
9	drop the _equal = true / else if _equal === false early-out so every render object re-answers for itself	3630847 (PRD-269)	dropped-upstream	r187 r187:684-694 replaced it with the per-render-object renderObjectData.materialVersion !== materialData._version mismatch, plus an output-target version check.	-
10	drop geometryData._equal = false, removed-attribute branch	3630847 (PRD-269)	dropped-upstream	r187 changed = true; continue;.	-
11	same, changed-attribute branch	3630847 (PRD-269)	dropped-upstream	r187 changed = true.	-
12	same, attribute-set changed branch	3630847 (PRD-269)	dropped-upstream	r187 changed = true.	-
13	same, index changed branch	3630847 (PRD-269)	dropped-upstream	r187 changed = true.	-
14	same, draw-range changed branch	3630847 (PRD-269)	dropped-upstream	r187 changed = true.	-
15	drop the geometry _equal = true / else if early-out	3630847 (PRD-269)	dropped-upstream	r187 r187:805-820 if ( changed === true ) geometryData._version ++; + renderObjectData.geometryVersion !== geometryData._version.	-
16	publish the settled answer for the draw, then use it instead of needsVelocity()	3630847 (PRD-269)	ported	Same two statements, but the early-out now returns RenderObjectRefreshType.FULL (r187's needsRefresh returns an enum, not a boolean) and keeps r187's new hasDynamicInstancing() term.	this.renderingVelocity = this.needsVelocity( nodeFrame.renderer );
17	replace the object.static early-out with isSettled() so only settled objects skip	3630847 (PRD-269)	ported	if ( this.isSettled( renderObject ) === true || isBundle ) ; the body returns refreshType instead of false, so r187's target-switch / disposed-attribute / shared-UBO refresh verdicts still win. const isStatic local removed (now unused).	if ( this.isSettled( renderObject ) === true || isBundle )

---

src/nodes/accessors/Batch.js

#	intent (one line)	origin (sha / PRD)	decision	reason	marker
1	import positionPrevious for the motion-vector batch transform	c03b929	dropped-upstream	r187 src/nodes/accessors/Batch.js:2 imports { positionLocal, positionPrevious } already.	-
2	previousMatricesTexture = batchMesh._previousMatricesTexture ?? matricesTexture	c03b929	dropped-upstream	r187 r187:88-108 getPreviousMatricesTexture( batchMesh ) creates the texture from _matricesTexture on first use (and again after BatchedMesh disposed it), which is the same fallback with a lifecycle.	-
3	assign positionPrevious from the previous batching matrix when needsPreviousData()	c03b929	dropped-upstream	r187 r187:146-167 does exactly this: createBatchingMatrixNode( getPreviousMatricesTexture( batchMesh ), indirectId ) → positionPrevious.assign( previousBatchingMatrixNode.mul( positionPrevious ).xyz ), plus the OnAfterObjectUpdate copy of the matrices texture that our hunk did not have.	-
4	drop the two trailing blank lines at EOF	c03b929	dropped-upstream	cosmetic only; r187 keeps them. Nothing to port.	-

File now equals stock r187. src/objects/BatchedMesh.js was checked for the duplicated half: r187 owns _previousMatricesTexture creation/dispose (r187:294, 1342-1345, 1562-1565) and Batch.js only fills it, so there is no double owner.

---

src/nodes/accessors/Instance.js

#	intent (one line)	origin (sha / PRD)	decision	reason	marker
1	import the before-object event and drop OnFrameUpdate	3630847 (PRD-269)	ported	r187 needs OnAfterObjectUpdate (its own fallback snapshot) instead of OnObjectUpdate, plus OnBeforeObjectUpdate for the scheduled history: import { OnAfterObjectUpdate, OnBeforeObjectUpdate }.	import { OnAfterObjectUpdate, OnBeforeObjectUpdate } from '../utils/EventNode.js';
2	force the vertex-attribute path in createInstanceMatrixNode (the uniform path re-uploads every draw)	4f9638c (PRD-453-457)	kept	r187 still has the same uniformBufferSize <= builder.getUniformBufferLimit() branch (r187:44-52), so the same if ( false ) keeps the cheaper path.	if ( false ) {
3	force the interleaved-attribute path in instance() and track the buffer	4f9638c (PRD-453-457)	dropped-obsolete	r187 deleted the whole interleaved tracking: no _matrixBuffers, no interleavedMatrix/interleavedColor locals; the columns are new InterleavedBufferAttribute( instanceMatrix, 4, offset ) bound straight to the attribute, and the buffer is freed by instanceMatrix.dispose(). Nothing left to force.	-
4	sync the interleaved buffers on OnBeforeFrameUpdate	4f9638c (PRD-453-457)	dropped-obsolete	The block this changed does not exist in r187 (same deletions as row 3); r187 uploads on attribute version change instead.	-
5	scheduled previous instance matrices (threenative.velocity.previousInstanceMatrices) must land before the geometry upload, else keep Three's after-draw snapshot	f1040a1 (PRD-269/455), 2d21247	ported	r187 moved the snapshot to OnAfterObjectUpdate for both paths, so the port adds a OnBeforeObjectUpdate pre-pass and guards the after-pass for both: the instance() path (_previousInstanceMatrices) and the ReferenceAttributeNode path (getPreviousMatrix). needsUpdate = true is used instead of the removed interleaved-buffer needsUpdate bookkeeping.	threenative.velocity.previousInstanceMatrices

---

src/nodes/accessors/Skinning.js

#	intent (one line)	origin (sha / PRD)	decision	reason	marker
1	previousBoneMatricesData() — per-skeleton history created at the current pose so the first frame has zero skin motion	c93b427 (PRD-723/746/747/748/749/752)	dropped-upstream	r187 setupPreviousBoneData( skeleton ) (r187:159-187) does the same, storing on the skeleton (previousBoneMatrices, previousBoneTexture, both declared in src/objects/Skeleton.js:80,98) instead of a module WeakMap.	-
2	resolve the bone history per drawn object, and use that frame's bind inverse	c93b427 + f1040a1 (PRD-269)	ported (split)	History half: dropped-upstream — r187 resolves it per object already (referenceBuffer( 'skeleton.previousBoneMatrices' ), and the texture variant's OnObjectUpdate sets object.skeleton.previousBoneTexture). Bind-inverse half: ported — nothing in r187 reads a scheduled bind inverse, so previousBindMatrixInverse (uniform, onObjectUpdate, threenative.velocity.previousBindMatrixInverse ?? mesh.bindMatrixInverse) is added to both the skinning() and computeSkinning() previous-position paths.	const previousBindMatrixInverse = uniform( skinnedMesh.bindMatrixInverse
3	let a scheduler supply the previous bone matrices in skinning()	c93b427	ported	r187:255 writes skeleton.previousBoneMatrices.set( skeleton.boneMatrices ); now scheduledPrevious ?? skeleton.boneMatrices, reading the symbol off the drawn object inside OnObjectUpdate.	threenative.velocity.previousBoneMatrices
4	same in computeSkinning()	c93b427	ported	r187:327 same pattern; the symbol is read off skinnedMesh because that path is not per object.	threenative.velocity.previousBoneMatrices

---

src/nodes/accessors/TextureNode.js

#	intent (one line)	origin (sha / PRD)	decision	reason	marker
1	pass isRawTextureNode into generateTextureLoad so the backend can pick the raw (sample-index) load form	f1040a1 (PRD-269/455)	kept	r187's generateTextureLoad still takes 6 params, so the extra argument is inert but harmless and keeps the feature wired for the backend half of the patch. Hunk merged clean and the surrounding sampler === false branch is unchanged.	this.isRawTextureNode === true );

---

src/nodes/accessors/TextureSizeNode.js

#	intent (one line)	origin (sha / PRD)	decision	reason	marker
1	a raw pass attachment is a multisampled texture, so textureDimensions() must not be given a mip level	f1040a1 (PRD-269/455)	ported	r187 moved the snippet into builder.generateTextureSize(), which decides from getTextureSampleData(); that returns primarySamples === 1 for a render-target texture (r187 resolves MSAA into a single-sample texture, WebGPUTextureUtils isMSAA → separate -msaa texture), so it would emit the level. The raw branch is emitted before the delegation, same intent as 0.185.	if ( this.textureNode.isRawTextureNode === true ) {

---

src/nodes/accessors/VelocityNode.js

#	intent (one line)	origin (sha / PRD)	decision	reason	marker
1	prefer the scheduler's previous world matrix over Three's own snapshot	3630847 (PRD-269)	kept	Hunk merged clean inside update(); r187 still resolves the reference from the per-object frame ({ frameId, camera, object }), so object[ Symbol.for( ... ) ] ?? previousModelMatrix applies unchanged.	threenative.velocity.previousWorldMatrix

---

src/nodes/core/NodeBuilder.js

#	intent (one line)	origin (sha / PRD)	decision	reason	marker
1	needsPreviousData() must also honour object.userData.useVelocity (how ThreeNative switches velocity on per object)	c03b929	ported	r187 deleted the getDataFromObject( this.object ).useVelocity field entirely (useVelocity no longer exists in r187 src/), so only the userData clause survives: ( mrt !== null && mrt.has( 'velocity' ) ) || this.object.userData?.useVelocity === true. Dropping getDataFromObject also matches r187 dropping its import.	this.object.userData?.useVelocity === true

---

src/nodes/display/PassNode.js

#	intent (one line)	origin (sha / PRD)	decision	reason	marker
1	PassMultipleTextureNode takes a rawTexture flag	f1040a1 (PRD-269/455)	kept	Constructor merged clean onto r187's PassTextureNode base (super( passNode, null ) still correct).	this.isRawTextureNode = rawTexture;
2	mark the node raw and force setSampler( false )	f1040a1	kept	setSampler still exists in r187 TextureNode (r187:777).	this.isRawTextureNode = rawTexture;
3	unique uniform hash per raw attachment, plus setup() preconditions and the clone() passthrough	f1040a1	kept	super.getUniformHash( builder ) and builder.renderer.backend.isWebGPUBackend both still exist in r187; the r187 rewrite of PassNode (now extends Node) does not touch these members.	:raw4:
4	_rawTextureNodes / _rawTexturesDisposed fields	f1040a1	kept	Merged clean; the new fields sit beside r187's untouched _textureNodes.	this._rawTexturesDisposed = false;
5	refresh the raw node on toggleTexture()	f1040a1	kept	toggleTexture() is unchanged in r187 PassNode.	raw4:
6	getRawTextureNode( name ) accessor	f1040a1	kept	Merged clean. Moved the (previously stolen) getTextureNode JSDoc back onto getTextureNode and gave getRawTextureNode its own.	getRawTextureNode( name = 'output' ) {
7	mark the pass disposed so later raw lookups throw	f1040a1	kept	r187 added super.dispose() to PassNode.dispose(); the flag is set after it, order is irrelevant.	_rawTexturesDisposed = true;

---

src/nodes/lighting/PointShadowNode.js

#	intent (one line)	origin (sha / PRD)	decision	reason	marker
1	share the shadow uniforms instead of building them per material	afee25e (PRD-494)	ported	r187 exports nothing here; the helper that now exists upstream is getShadowUniforms( shadow ) in ShadowFilterNode.js. The port imports that name instead of our getShadowUniform. The now-unused uniform / reference / renderGroup imports were dropped.	bias } = getShadowUniforms( shadow );
2	PointShadowFilter: one radius / mapSize node per shadow (r187 builds five uniforms per material here)	afee25e (PRD-494)	ported	const { radius, mapSize, near, far, bias } = getShadowUniforms( shadow ); — r187's own body created all five as fresh nodes per builder, exactly the buffer-per-material problem the hunk fixes.	bias } = getShadowUniforms( shadow );
3	pointShadowFilter: one cameraNear / cameraFar / bias node per shadow	afee25e (PRD-494)	ported	const { near: shadowCameraNear, far: shadowCameraFar, bias } = getShadowUniforms( shadow );; the uniform( 'float' ).onRenderUpdate( () => shadow.camera.near ) bodies moved into the shared helper unchanged.	far: shadowCameraFar

---

src/nodes/lighting/ShadowFilterNode.js

#	intent (one line)	origin (sha / PRD)	decision	reason	marker
1	one uniform node per shadow (module WeakMap + getShadowUniform( shadow, name, create )) so every material shares the render-group buffer	afee25e (PRD-494)	ported	Upstream added the same idea as getShadowUniforms( shadow ) (r187:13-37) but only for mapSize/radius, and left the point-shadow uniforms per material. Kept upstream's shape, added near/far/bias to it and exported it, so there is exactly one mechanism.	export function getShadowUniforms( shadow ) {
2	PCFShadowFilter uses the shared mapSize / radius	afee25e (PRD-494)	dropped-upstream	r187:85-86 const { mapSize, radius } = getShadowUniforms( shadow );.	-
3	PCFSoftShadowFilter uses the shared mapSize	afee25e (PRD-494)	dropped-obsolete	PCFSoftShadowFilter was deleted upstream (r187 src/nodes/lighting/ShadowNode.js:117 keeps null /* PCFSoftShadowMap, removed */ in _shadowFilterLib; no references remain in src/). Nothing left to fix.	-

---

src/nodes/pmrem/PMREMNode.js

#	intent (one line)	origin (sha / PRD)	decision	reason	marker
1	module _pending / _failed maps and a _listenForDispose helper for the PMREM cache	4e8fe84 (PRD-624/573), 2d21247	ported	Kept, with r187's invalidation folded in: the cache now stores the CubeRenderTarget, so the listener sets pmrem.texture.pmremVersion = - 1 before pmrem.dispose() (r187:73-78).	const _pending = new WeakMap();
2	while a preparation is pending, never start a second bake for the same texture	4e8fe84	ported	pending.textures.has( texture ) guard kept; the cached value is now a render target, so it compares/returns prepared.texture.pmremVersion and prepared.texture.	prepared.texture.pmremVersion === texture.pmremVersion
3	on a queue-fenced backend never bake synchronously (all GGX passes in one submission hang the GPU process)	4e8fe84	ported	renderer.backend.device.queue.onSubmittedWorkDone is unchanged in r187; the fire-and-forget prepareTextureAsync + _failed guard are kept as-is.	fence.onSubmittedWorkDone
4	use _listenForDispose instead of the inline listener	4e8fe84	ported	Inline block replaced by the helper (which keeps r187's pmrem.texture.pmremVersion = - 1).	_listenForDispose( texture, cache );
5	PMREMNode.prepareTextureAsync() — serialized, per-renderer queue of background bakes	4e8fe84	ported	Rewritten for r187: cache values are CubeRenderTargets (generated.texture.pmremVersion = version, generated.dispose()), and the CubeUVReflectionMapping early-out became r187's isPMREMTexture === true (the constant is no longer imported by this file). fromCubemapAsync/fromEquirectangularAsync still exist in r187 PMREMGenerator (deprecated, await renderer.init() then _fromTexture, returning a CubeRenderTarget).	static async prepareTextureAsync( renderer, texture ) {
6	Y flip only for generated PMREMs, read off the live texture (this._texture.value.isRenderTargetTexture)	f1040a1	dropped-obsolete	The whole Y-flip block and textureCubeUV/_width/_height/_maxMip were removed upstream; r187 samples the generated cube texture directly (this._texture.sample( materialEnvRotation.mul( uvNode ) ).level( roughnessToMip( levelNode, this._maxLod ) )). The stale this._pmrem read the hunk fixed no longer exists in setup().	-

---

src/objects/BatchedMesh.js

#	intent (one line)	origin (sha / PRD)	decision	reason	marker
1	capability marker so the engine can detect the batched-velocity patch at runtime	c03b929	kept	Independent of upstream's work; _previousMatricesTexture now exists in r187 itself, but the flag is what readPatchedThree/hasBatchedVelocityPatch probes for.	threeNativeBatchedVelocityPatch

---

examples/jsm/tsl/display/SMAANode.js

#	intent (one line)	origin (sha / PRD)	decision	reason	marker
1	pin the edges texture to mip 0 so the bounded searches contain no implicit derivatives (FXC)	4e8fe84 (PRD-624/573)	kept	texture( ... ).level( 0 ) is unchanged API in r187; hunk merged clean, TextureNode.level() still exists.	texture( this._renderTargetEdges.texture ).level( 0 )

---

examples/jsm/tsl/display/SSRNode.js

#	intent (one line)	origin (sha / PRD)	decision	reason	marker
1	maxSteps uniform capping the mirror trace's raymarch steps	37fccbc (PRD-732/733/734/735)	kept	Merged clean; the field sits after r187's quality uniform.	this.maxSteps = uniform( 1e6 );
2	apply the cap to totalStep	37fccbc (PRD-732/733/734/735)	kept	Merged clean onto r187's trunc( max( abs( xLen ), abs( yLen ) ).mul( this.quality.clamp() ) ) chain.	.min( this.maxSteps )

---

Doubts, ranked by risk

1. TextureSizeNode + TextureNode only half-wire the raw pass attachment. r187's WGSL backend takes 6 args in generateTextureLoad and declares a texture's binding type from getTextureSampleData(), which reports primarySamples === 1 for a render-target texture because r187 resolves MSAA into a separate -msaa texture. On the 0.185 patch the raw binding was forced multisampled in WGSLNodeBuilder (primarySamples > 1 || textureNode.isRawTextureNode === true). If the backend arm ports that, the level-less textureDimensions() I emit is right; if it does not, isRawTextureNode is dead and both hunks are inert but harmless (only our own PassNode ever sets the flag). Must be settled against the WGSLNodeBuilder.js / WebGPUBindingUtils.js arms before trusting raw pass reads on r187.
2. Batch.js now has no patch content at all. packages/core/__tests__/packaging.spec.ts:275 asserts the packed build/three.webgpu.js contains _previousMatricesTexture ?? matricesTexture, and packaging.spec.ts:182 asserts it in three@0.185.1.patch. Both will fail when the 0.187 patch is generated; the string lives only in the old Batch.js hunk. Needs an engine-repo test update (out of this arm's scope).
3. needsPreviousData() lost getDataFromObject( this.object ).useVelocity. r187 deleted that field, and the engine only ever sets object.userData.useVelocity (packages/core/src/render/velocity.ts:259), so the port is correct — but any ThreeNative code or persisted scene data that relied on the old data-map flag would silently stop requesting previous data.
4. Instance.js version bookkeeping is now mixed. The scheduler path uses needsUpdate = true (version++), the fallback keeps r187's previous.version = matrices.version. If a scheduler stops writing history mid-run, the fallback can assign a version that is not greater than the one the scheduled write set; worst case is one redundant re-upload or one delayed upload, and no GPU was available to confirm.
5. getShadowUniforms() now eagerly creates near/far/bias for every PCF shadow (directional/spot included) even though only point shadows build them. Unused nodes are never built into a shader, so this is allocation only — but it does mean the helper is no longer a pure mapSize/radius carrier, i.e. a future upstream change to that helper will conflict.
6. updateSettled() does not know about output-target switches or disposed attributes. A settled object on a resized/retargeted pass still refreshes because r187's pre-checks set refreshType = FULL before the isSettled() early-out — but that is now an invariant on r187's needsRefresh() body, not on isSettled(). If a later upstream change moves those checks below the early-out, settling would skip a required refresh.
7. PCFSoftShadowFilter is gone upstream (ShadowNode.js:117 keeps a null slot). Any ThreeNative material that still sets shadowMap.type = PCFSoftShadowMap now silently falls back to the removed-filter slot behaviour rather than to our de-duplicated soft filter.
8. SSRNode's maxSteps default of 1e6 is unchanged, so the mirror trace is still effectively unbounded unless a caller sets it — the port preserves the stock behaviour, per the patch's own comment.
9. SMAANode: the patch keeps its // comment between the JSDoc block and the this._edgesTextureUniform assignment; harmless for tooling but not idiomatic three.js.

ARM DONE

[jonit-dev]

Ledger — ARM B (0.185.1 → 0.187.0 port of the ThreeNative patch)

Repo: ~/.cache/tmp-tn/prd775/merge, branch rebased. Stock reference tree: ~/.cache/tmp-tn/prd775/pack/r187/package/.
Origins read from the read-only engine repo (packages/core/patches/, 4e8fe84ae squashes PRD-624, so the
immediately preceding wip commit is cited where it is more specific).

Notes that apply to every row:

• git diff base patched lists the ThreeNative hunks; git diff r187 -- <file> is my result against stock.
• Where a ThreeNative hunk and an upstream r187 change are the same fix, the row says dropped-upstream
  and cites the r187 lines; where the shape of the 0.185.1 code the hunk patched no longer exists, it says
  dropped-obsolete and proves it.
• Renderer.js hunk 13/14 are dropped-upstream (see the note under that table) — upstream's
  RenderObjectRefreshType machinery answers the same question our "always run every step" answered, and the
  isSettled gate the observer arm ported keeps it honest.

---

src/renderers/common/Bindings.js

#	intent (one line)	origin (sha / PRD)	decision	reason	marker
1	Tell the binding update whether the render object is settled, so per-object uniform groups can skip their write	a602467 (PRD-458/473)	kept	updateForRender still forwards to _updateBindings; r187 kept that call site unchanged, so only the extra argument was needed	this.nodes.isSettled( renderObject )
2	Refresh a sampled-texture / sampler binding before the first GPU bind group is created, so an async warmup that yielded while the node swapped targets does not bake a stale reference	7870147 (fix(core): refresh initial texture bindings after async warmup)	ported	r187 changed both lines to binding.generation = this.textures.updateTexture( … ) / binding.samplerKey = this.textures.updateSampler( … ); kept that return-value form and added binding.update() in front of each	// Refresh its reference before creating the first GPU binding, not after.
3	Thread settled through _updateBindings / _update	a602467 (PRD-458/473)	kept	Same call graph in r187 (updateForCompute still calls _updateBindings with one argument → undefined, i.e. not settled)	@param {boolean} [settled] - Whether the render object is a settled static object.
4	Build a bind group that a node added after the object's first draw, on every draw, and force its values to be written	a602467 (PRD-458/473)	kept	r187 has no equivalent (its _createBindings only runs for uninitialised render objects); harmless when the refresh-type gate skips the call	// A group a node added after this object's first draw has no buffers and no layout yet, and
5	A settled object still has to notice that the GPU texture behind one of its bindings was re-created (re-sized render target)	4e8fe84 / f30f9b6 (PRD-624, PRD-478)	kept	Compare binding.generation with this.textures.get( texture ).generation; r187 keeps textureData.generation and the binding.generation field, so the comparison is still meaningful	if ( settled === true && binding.isSampledTexture === true && binding.texture !== undefined )

---

src/renderers/common/Geometries.js

#	intent (one line)	origin (sha / PRD)	decision	reason	marker
1	Remember the attribute versions a settled object was last drawn with	a602467 (PRD-458/473)	kept	updateForRender is still the per-draw hook in r187; the WeakMap has no upstream equivalent	this._settledAttributes = new WeakMap();
2	Skip the attribute scan for a settled object only when no attribute version moved; add attributesChanged()	a602467 (PRD-458/473)	kept	renderObject.getMonitor() (RenderObject.js:482) and monitor.isSettled() (observer, ported by another arm) exist in the merged tree. r187's monitor also compares instance/attribute versions in equals() — that is the same signal one layer up, so this stays as the second line of defence and is inert, not wrong, when the refresh type is FULL/SHARED	attributesChanged( renderObject ) {

---

src/renderers/common/RenderContexts.js

#	intent (one line)	origin (sha / PRD)	decision	reason	marker
1	The default render target (canvas) must report the renderer's current sample count, not "1", so output-pass and direct-overlay passes do not share a render state	2d21247 (PRDs 466–468) / 1abe8f3	kept	Overwrites the sampleCount line r187 still has verbatim at RenderContexts.js:89; renderer.currentSamples exists at Renderer.js:2962	const sampleCount = renderTarget !== null ? renderTarget.samples : this.renderer.currentSamples;

---

src/renderers/common/RenderObject.js

#	intent (one line)	origin (sha / PRD)	decision	reason	marker
1	Track the version of _sourceMaterial this render object was validated against (override material is a shared scratchpad)	f30f9b6 (PRD-478)	dropped-upstream	r187 added the same field for the same reason: this.sourceMaterialVersion = this._sourceMaterial !== null ? this._sourceMaterial.version : 0; (RenderObject.js constructor, and RenderObjects.get compares it). Our sourceVersion was deleted so only one name exists	-
2	getChainArray() must include the source material so override and non-override draws of the same object are different chains	f30f9b6 (PRD-478)	dropped-upstream	r187 returns [ object, material, context, lightsNode, this._sourceMaterial ] when _sourceMaterial !== null, which is the same discriminator (ours pushed _sourceMaterial ?? material, which duplicated element 1 when there was no override)	-
3	getGeometryCacheKey( useMorphTargetIds ) — a pipeline key must not depend on morph-attribute object identity, only on their layout	a09cea3 (PRD-566)	dropped-upstream	r187 added WebGPUAttributeUtils.getVertexBufferLayoutCacheKey( renderObject ) to the pipeline cache key and removed stride/offset/index from the geometry key, i.e. the pipeline key is now built from the real vertex layout (arrayStride/stepMode/shaderLocation/offset/format), which is strictly more precise. No caller passes false any more, so the parameter (and its morphTargetsRelative branch) was removed to avoid dead code	-
4	The dynamic cache key must include the render context's sample count, so a 4× and a 1× pass do not share cached render objects	32ebfa8 (PRD-400)	ported	r187 rewrote getDynamicCacheKey() to hashArray( _cacheKeyValues ) with a fixed 5-slot array; grew the array to 6 and set slot 5 to this.context.sampleCount (RenderContext.sampleCount exists, RenderContext.js:212)	_cacheKeyValues[ 5 ] = this.context.sampleCount;

---

src/renderers/common/RenderObjects.js

#	intent (one line)	origin (sha / PRD)	decision	reason	marker
1	Chain key must include the source material (_chainKeys[ 4 ]) so material groups sharing a shadow override keep separate GPU bindings	f30f9b6 (PRD-478)	dropped-upstream	r187 introduced _overrideChainKeys (5 slots, [ null, null, null, null, null ]) used only while an override is active, with chainKeys[ 4 ] = sourceMaterial; that is the same fix, and it avoids the stale 5th slot in the non-override array	-
2	Version gate must answer with the source material's version while drawing through an override material, and rebuild when the source material was replaced	f30f9b6 (PRD-478)	dropped-upstream	r187's gate is renderObject.version !== material.version || renderObject.sourceMaterialVersion !== sourceMaterialVersion || renderObject.needsUpdate || force. sourceReplaced cannot fire in either tree because the source material is part of the chain key, and force (geometry cache-key change) is upstream's extra. Dropped our sourceVersion/sourceChanged block	-
3	Reset _chainKeys[ 4 ] after the lookup	f30f9b6 (PRD-478)	dropped-upstream	r187 resets chainKeys[ 4 ] only when sourceMaterial !== null, which matches its 4-vs-5 slot scheme	-

Result: this file is byte-identical to stock r187 (diff -q clean).

---

src/renderers/common/RenderPipeline.js

#	intent (one line)	origin (sha / PRD)	decision	reason	marker
1	The pipeline quad is the renderer's output owner; restore renderer state (xr, tone mapping, colour space, owner) even if the draw throws	f30f9b6 (PRD-478) / 1d056ce	kept	r187 kept the un-guarded sequence at RenderPipeline.js:146-160; our try/finally wraps only _quadMesh.render( renderer ), the trailing upstream restores stay (harmless duplicates on the success path, the safety net on the throw path)	const previousOwner = renderer._tnOutputOwner;

---

src/renderers/common/Renderer.js

#	intent (one line)	origin (sha / PRD)	decision	reason	marker
1	Mark which scene owns the render context's output (transient depth may only be speculated for it)	1d056ce (#486)	kept	r187 has no _tnOutputOwner; the WebGPU backend's _beginTransientDepth / _validateTransientDepth read it	renderContext._tnOutputOwner = scene === this._tnOutputOwner ? scene : null;
2	renderScene must finish or abort the pass even when an object draw throws, and always restore the render tree	f30f9b6 (PRD-478)	kept	r187 still calls backend.finishRender unconditionally after the render lists; the nested try/finally adds the abort path (_abortTransientDepth) the transient-depth feature needs	else if ( this.backend.get( renderContext )._tnDepth ) this.backend._abortTransientDepth( renderContext );
3	(same hunk as 2 — the tree restore / _callDepth -- moved inside the inner finally)	f30f9b6 (PRD-478)	kept	as row 2	renderContext._tnOutputOwner = scene === this._tnOutputOwner ? scene : null;
4	Public deleteAttribute( attribute ) so an exclusively owned attribute is released through backend + memory accounting	f1040a1 (PRDs 269/455)	kept	Attributes.delete( attribute ) exists at Attributes.js:75 in r187	deleteAttribute( attribute ) {
5	Cold, explicit async compressed-texture gate (prepareTextureAsync)	2d21247 / 4e8fe84 (PRD-624)	kept	Calls this._textures.updateTextureAsync, which row Textures#7 provides; the WebGPU backend exposes updateTextureAsync	async prepareTextureAsync( texture, options = {} ) {
6	Cache the shadow pass's alphaTestNode per material (shadow nodes cache keyed by material)	f30f9b6 (PRD-478)	kept	NodeMaterial.alphaTestNode exists (NodeMaterial.js:221) and r187's getShadowNodes cache object has no such entry	alphaTestNode: material.alphaTestNode && material.alphaTestNode.isNode ? material.alphaTestNode : null
7	Save/restore the override material's alphaTestNode around the override block	f30f9b6 (PRD-478)	kept	Same override-save pattern as the surrounding upstream lines	let materialAlphaTestNode;
8	Capture the caster's alphaTest and alphaTestNode before mutating the shared override material	f30f9b6 (PRD-478)	kept	r187 still mutates overrideMaterial in place and restores eight fields, none of them alpha test	materialAlphaTestNode = ( overrideMaterial.isNodeMaterial ) ? overrideMaterial.alphaTestNode : null;
9	Write overrideMaterial._alphaTest so the shared override material's version does not churn per caster	f30f9b6 (PRD-478)	dropped-upstream	r187 has the same assignment with the same rationale: overrideMaterial._alphaTest = material.alphaTest; // bypass the setter so the version doesn't change. Upstream's one-liner kept verbatim	-
10	Pull alphaTestNode out of _getShadowNodes() for shadow pass materials	f30f9b6 (PRD-478)	kept	r187 destructures only { colorNode, depthNode, positionNode }	const { colorNode, depthNode, positionNode, alphaTestNode } = this._getShadowNodes( material );
11	Assign the per-caster alphaTestNode to the shadow override material	f30f9b6 (PRD-478)	kept	paired with row 10	overrideMaterial.alphaTestNode = alphaTestNode;
12	Restore _alphaTest / alphaTestNode on the scene's override material after the draw	f30f9b6 (PRD-478)	kept	paired with rows 8/11	scene.overrideMaterial._alphaTest = materialAlphaTest;
13	Run every draw step unconditionally instead of behind nodes.needsRefresh(), letting each component decide what a settled object may skip	a602467 (PRD-458/473)	dropped-upstream	r187 replaced the boolean with RenderObjectRefreshType (constants.js:1711) and _renderObjectDirect gates FULL / SHARED / NONE (Renderer.js:4076-4096); the observer arm ported needsRefresh with isSettled( renderObject ) replacing object.static, and r187's equals() compares world matrix, output-target version, geometry attribute + index versions, instance matrix/colour, morph texture and material/light data — exactly the "nothing it depends on changed" set our patch computed inside the components. The upstream structure is the same fix with upstream's own plumbing, so ours was not duplicated	-
14	Always call nodes.updateAfter( renderObject ) after a draw	a602467 (PRD-458/473)	dropped-upstream	r187 calls it for every refresh type except NONE (if ( refreshType !== RenderObjectRefreshType.NONE ) this._nodes.updateAfter( renderObject )) and, when the pipeline is not ready, still invalidates the bundle version — strictly better than unconditional	-

---

src/renderers/common/Textures.js

#	intent (one line)	origin (sha / PRD)	decision	reason	marker
1	State for the single async-prepare lane	2d21247 (PRDs 466–468)	kept	r187 has no async lane; fields sit next to _htmlTextures	this._uploadTail = Promise.resolve();
2	updateTexture() becomes the guard (pending upload / remembered upload error) and _updateTexture() the body, so an in-flight preparation cannot be raced by a sync update	2d21247	ported	r187 turned updateTexture into a generation-returning early-out (return textureData.generation;) and added the render-target early path; kept both, moved them into _updateTexture, and made every non-upload exit return textureData.generation (Bindings assigns it to binding.generation)	if ( this._pendingUploads.has( texture ) ) throw new Error( 'Compressed texture preparation is pending.' );
3	Dropping a destroyed render-target texture must invalidate the cached GPU bind groups that point at it	a602467 / 4e8fe84	kept	r187 keeps textureData.bindGroups (filled by Bindings) and the groups/versions cache objects, but has no invalidation on re-create	// delete cached bind groups so they don't point to the destroyed texture
4	Upload through the async backend when the caller supplied an upload callback	2d21247	ported	r187 added texture.clearUpdateRanges() right after backend.updateTexture(...) in the same block; kept that on the sync branch only and routed the upload branch to _initializeTexture + the callback	uploadPromise = upload( texture, options );
5	texture.onUpdate belongs to the completion of an upload, not to its submission	2d21247	kept	no upstream equivalent	if ( ! upload && texture.onUpdate ) texture.onUpdate( texture );
6	Split the one-time initialisation out of the update body so the async path can initialise before uploading and still return a result	2d21247	ported	r187 rewrote that block (dispose handler only for non-render-target textures, _tracked / _registry weak refs); kept the extraction and the weak-ref registration, and re-added the generation return that r187's rewrite has at the end of updateTexture	_initializeTexture( texture, textureData ) {
7	updateTextureAsync() + _waitForUpload(): single preparation lane keyed by texture, device-loss/abort/dispose handling, revalidation before publishing	2d21247	kept	Pure additions; needs backend.updateTextureAsync (WebGPUTextureUtils) which is in this arm	updateTextureAsync( texture, options = {} ) {
8	dispose() must abort in-flight preparations before freeing GPU data	2d21247 / 4e8fe84	ported	r187 added its own dispose() (weak-ref sweep + super.dispose()); merged into one method — abort the lane first, then r187's _tracked sweep	this._uploadDisposed = true;
9	Disposing a texture in flight cancels its preparation	2d21247	kept	r187's _destroyTexture has no such hook	this._pendingUploads.get( texture )?.controller.abort( new Error( 'Texture disposed during preparation.' ) );

---

src/renderers/common/extras/PMREMGenerator.js

#	intent (one line)	origin (sha / PRD)	decision	reason	marker
1	fromEquirectangularAsync() must fence the bake on WebGPU instead of returning while the GPU is still filtering; drop the deprecation notice because the engine's warmup uses this path	596e93b + 4e8fe84 (PRD-624)	ported	r187 kept warnOnce( … is deprecated … ) + return this._fromTexture( … ); dispatch now goes to _fromTextureAsync on backend.isWebGPUBackend (WebGPUBackend.js:92)	return this._fromTextureAsync( equirectangular, renderTarget );
2	Same for fromCubemapAsync()	596e93b + 4e8fe84 (PRD-624)	ported	as row 1	return this._fromTextureAsync( cubemap, renderTarget );
3	_fromTextureAsync / _renderFacesAsync / _textureToCubeUVAsync / _applyPMREMAsync / _applyGGXFilterAsync — fence the bake in bounded slices (one cube face at most in flight) and resolve only after queue.onSubmittedWorkDone()	596e93b + 4e8fe84 (PRD-624)	ported	r187 rewrote the whole class: _init, _sizeLods, _lodMeshes, _pingPongRenderTarget, _flatCamera, _renderFaces, _applyGGXFilter, _setViewport and the CubeUV atlas are gone; the generator now renders each mip level with CubeCamera into a CubeRenderTarget. Re-expressed by intent on the new code: _setSizeFromTexture() factored out of r187's inline body (also used by the sync path), _fromTextureAsync() = r187 _fromTexture with the async passes plus a final fence and disposal of an internally allocated target when the bake fails, _renderCubeAsync() = r187 _renderCube's CubeCamera face loop unrolled one face at a time (same mipmapsAutoUpdate/xr.enabled/MRT/render-target restore and the same reversed-depth clear guard as CubeCamera.update), _textureToCubemapAsync()/_applyPMREMAsync() = the r187 bodies with await. The fence budget keeps its original meaning (size(LOD 1)² × GGX_SAMPLES = cost of one heavy face) with per-face cost derived from the mip size and the material's sample count	await queue.onSubmittedWorkDone();

---

src/renderers/common/nodes/NodeManager.js

#	intent (one line)	origin (sha / PRD)	decision	reason	marker
1	updateGroup() skips the per-object uniform write when the object is settled	a602467 (PRD-458/473)	kept	r187's updateGroup( nodeUniformsGroup ) still returns true for NodeUpdateType.OBJECT; only used by Bindings._update	updateGroup( nodeUniformsGroup, settled ) {
2	isSettled( renderObject ) — the one place the settled question is asked	a602467 (PRD-458/473)	kept	Delegates to renderObject.getMonitor().isSettled(); the merged NodeMaterialObserver (other arm) has isSettled at line 376	isSettled( renderObject ) {
3	Settled objects skip their OBJECT-type updateBefore nodes	a602467 (PRD-458/473)	kept	r187 loops nodeBuilder.updateBeforeNodes unconditionally	if ( settled === true && node.getUpdateBeforeType() === NodeUpdateType.OBJECT ) continue;
4	Settled objects skip their OBJECT-type updateAfter nodes	a602467 (PRD-458/473)	kept	r187 only has the // update frame state for each node comment here — our comment replaced it	if ( settled === true && node.getUpdateAfterType() === NodeUpdateType.OBJECT ) continue;
5	Settled objects skip their OBJECT-type per-draw node updates	a602467 (PRD-458/473)	kept	r187 loops nodeBuilder.updateNodes unconditionally	if ( settled === true && node.getUpdateType() === NodeUpdateType.OBJECT ) continue;

---

src/renderers/webgpu/WebGPUBackend.js

#	intent (one line)	origin (sha / PRD)	decision	reason	marker
1	Hold the GPU instance and the adapter for the lifetime of the backend — a collected adapter takes its instance with it and every in-flight operation then fails with "A valid external Instance reference no longer exists"	f1040a1 (PRDs 269/455)	kept	r187 still releases the adapter (const adapter = … await navigator.gpu.requestAdapter( … ) and nothing else holds it)	this.adapter = adapter;
2	Rebuild the canvas render-pass descriptor when depth / stencil / reversedDepthBuffer change, not only the sample count	2d21247 (PRDs 466–468)	kept	r187's guard is only descriptor === undefined || canvasData.samples !== samples	canvasData.reversedDepthBuffer !== renderer.reversedDepthBuffer
3	Transient-depth speculation: helpers that swap the canvas depth for a pass-local MSAA depth and prove a full-viewport, identity-transform triangle before keeping it	1d056ce (#486), f30f9b6 (PRD-478)	kept	All new methods (_isFullTransientViewport, _beginTransientDepth, _finishTransientDepth, _restoreTransientDepth, _resumeTransientDepth, _abortTransientDepth, _releaseTransientDepth, _validateTransientDepth); names they touch all exist in r187 (renderer.getCanvasTarget(), textureUtils._depthBuffers (this arm), renderObject.getIndirect(), renderObject.getIndex(), renderObject.getBindings())	_beginTransientDepth( context, descriptor ) {
4	beginRender keeps the (possibly swapped) descriptor and the encoder on the render-context data	1d056ce	kept	r187 keeps only a local descriptor / encoder	descriptor = this._beginTransientDepth( renderContext, descriptor );
5	finishRender folds the speculation back before the occlusion query resolve and the submit	f30f9b6 (PRD-478)	kept	no upstream hook	this._finishTransientDepth( renderContext );
6	Submit inside try/finally and always release the private depth (and mark it reusable once the discard was really encoded)	1d056ce / f30f9b6	kept	r187 calls submit( this.device, renderContextData.encoder.finish() ) bare	this._releaseTransientDepth( renderContext );
7	A viewport change to anything but the full viewport must give the canonical depth back	1d056ce	kept	r187's updateViewport only writes the pass viewport	if ( ! this._isFullTransientViewport( renderContext ) ) this._restoreTransientDepth( renderContext );
8	Same for scissor	1d056ce	kept	r187's updateScissor only writes the pass scissor	this._restoreTransientDepth( renderContext );
9	Record that real (non-proven) draws happened into the speculative pass, so the discard is undone	f30f9b6 (PRD-478)	kept	r187's _draw has no such state	renderContextData._tnDepth.logical = 0;
10	Validate the speculation per draw (identity matrices, shipped projection chain, no discard / frag_depth / sample_mask / clip_distances, plain non-indexed triangle) and fall back when it does not hold	f30f9b6 (PRD-478)	kept	Reads _tnDepthProof (WebGPUPipelineUtils, this arm), _tnBindings / _tnUniformBytes (WebGPUBindingUtils, this arm), _tnFullTriangle / _tnTriangleVersion / uploadCall (WebGPUAttributeUtils, this arm)	_validateTransientDepth( renderObject, params, vertexBuffers ) {
11	Bind-group comparison must compare the native bind group object, not the binding id	f30f9b6 (superseded)	dropped-upstream	r187 does exactly this with one call instead of two: const bindGroupGPU = this.get( bindGroup ).group; … if ( currentBindingGroups[ i ] !== bindGroupGPU ) (WebGPUBackend.js, _draw bind-group loop)	-
12	Store the native bind group object in the per-pass binding cache	f30f9b6 (superseded)	dropped-upstream	as row 11 — currentBindingGroups[ i ] = bindGroupGPU;	-
13	A changed bind-group layout must force a new render pipeline	a09cea3 (PRD-566)	ported	r187's needsUpdate condition ends at data.clippingContextCacheKey; kept upstream's data.geometryVersion term and added ours after it	data.bindGroupLayoutKeys !== bindGroupLayoutKeys
14	… and the layout keys have to be stored with the rest of the pipeline state	a09cea3 (PRD-566)	kept	r187 stores geometryVersion + clippingContextCacheKey only	data.bindGroupLayoutKeys = bindGroupLayoutKeys;
15	Bind-group layout keys are part of the pipeline cache key	a09cea3 (PRD-566)	kept	r187's getRenderCacheKey has no layout term	`const bindGroupLayoutKeys = renderObject.getBindings().map( bindGroup => this.get( bindGroup ).layoutKey ).join( '
16	Pipeline key uses the structural geometry key without morph-target ids	a09cea3 (PRD-566)	dropped-upstream	r187 replaced that entry with utils.getStripIndexFormat( object, geometry ) + this.attributeUtils.getVertexBufferLayoutCacheKey( renderObject ), which keys on the actual vertex layout (see RenderObject#3)	-
17	Backend passthrough for the cold async compressed-texture path	2d21247	kept	this.textureUtils.updateTextureAsync is in this arm	updateTextureAsync( texture, options ) {
18	Keep the generated shader source with the compiled program so the transient-depth proof can compare it	f30f9b6 / 1d056ce	kept	r187 stores only { module, entryPoint }	programGPU._tnShaderCode = program.code;
19	Allocate the CPU-side mirror of a uniform buffer only when it can carry matrices	1d056ce	kept	r187 allocates no mirror	uniformBufferData._tnUniformBytes = new Float32Array( byteLength / 4 );
20	(hunk touched only a blank line before copyTextureToTexture's mipmap guard)	4e8fe84 (PRD-624)	dropped-obsolete	our hunk adds one empty line and changes no behaviour; r187's guard is dstLevel === 0 && dstTexture.generateMipmaps === true && dstTexture.mipmapsAutoUpdate === true, which is kept verbatim	-
21	A framebuffer copy resumes the canonical depth pass first	1d056ce	kept	r187's copyFramebufferToTexture starts with const renderContextData = this.get( renderContext );	this._restoreTransientDepth( renderContext );

---

src/renderers/webgpu/nodes/WGSLNodeBuilder.js

#	intent (one line)	origin (sha / PRD)	decision	reason	marker
1	textureLoad on a multisampled or raw (non-sampler) texture must take an i32 sample index — naga v25 (wgpu-native) rejects u32 there	c8546d0 (#279)	kept	r187 still emits u32( ${ levelSnippet } ) unconditionally; the added 7th parameter is already passed by TextureNode.js:635 (this.isRawTextureNode === true, other arm)	`const indexType = ( rawTexture
2	A raw texture node never gets a sampler in the bind group layout	f1040a1 (PRDs 269/455)	kept	r187's needsSampler has no raw-texture term	node.isRawTextureNode !== true &&
3	Do not name uniform buffers with a process-wide allocation id (it changes identical shader source per buffer)	4e8fe84 / #289	dropped-upstream	r187 removed the very line our hunk rewrote (uniformNode.name = name ? name : 'NodeBuffer_' + uniformNode.id; is gone from getUniformFromNode), so naming is left to the base builder	-
4	@invariant on the vertex position output for the temporal passes	f1040a1 (PRDs 269/455)	ported	r187 added the same @invariant behind this.context.invariant === true (WGSL and GLSL builders). The engine (and its starter templates) set the flag through TSL: renderer.contextNode = context( { invariantPosition: true } ), so dropping invariantPosition would silently disable the fix. Accept both flags	this.context.invariantPosition === true
5	Sample-rate interpolation policy for the temporal passes (context( { sampleInterpolation } )), keeping flat/integer varyings alone	f1040a1 (PRDs 269/455)	kept	r187's getVaryings honours only varying.interpolationType	const policy = this.context.sampleInterpolation;
6	The fragment stage must derive needsSampler from the layout frozen in getUniformFromNode (nodeData.uniformGPU), not from live texture filter state, or a rebuilt shader shifts every later binding index	f1040a1 (PRDs 269/455)	kept	r187 recomputes from texture.isCubeTexture / isUnfilterable; getDataFromNode( node, stage, globalCache ) and uniformGPU exist in r187's builder	const needsSampler = Array.isArray( uniformGPU ) && uniformGPU[ 0 ].isSampler === true && uniformGPU[ 0 ].isSampledTexture !== true;
7	A raw texture node binds as texture_multisampled_2d (it is a resolved-or-not MSAA attachment view)	f1040a1 (PRDs 269/455)	kept	r187 checks primarySamples > 1 only	`if ( primarySamples > 1

---

src/renderers/webgpu/utils/WebGPUAttributeUtils.js

#	intent (one line)	origin (sha / PRD)	decision	reason	marker
1	DynamicDrawUsage is needed to decide whether a buffer may be re-uploaded twice in one render call	f30f9b6 (PRD-478)	kept	r187 imports only Float16BufferAttribute from BufferAttribute.js	import { DynamicDrawUsage } from '../../../constants.js';
2	Record which buffer holds the engine's full-screen triangle, its version and its upload call	f30f9b6 / 1d056ce	kept	_validateTransientDepth (WebGPUBackend, this arm) verifies the bytes; r187 records nothing	bufferData._tnFullTriangle = backend.device.features?.has( 'transient-attachments' )
3	Re-upload a shared buffer whose dirty bytes belong to a fresh attribute wrapper, and never repeat a write already made in this render call	f30f9b6 / 1d056ce	kept	r187's updateAttribute re-writes unconditionally; the guard is what makes a re-built shader wrap an existing buffer safely	bufferData._tnFullTriangle = false;
4	Track the version and the render call of the last upload	f30f9b6 / 1d056ce	kept	r187 stores only _itemSize	bufferData.uploadCall = backend.renderer.info.calls;
5	Destroying an interleaved attribute must also forget the destroyed buffer on its InterleavedBuffer, as the WebGL backend does, or the next createAttribute() submits a destroyed buffer	f30f9b6 (PRD-478)	ported	r187 rewrote the same method (backend.getBufferAttribute( attribute ), guard-less data.buffer.destroy(), single backend.delete( bufferAttribute )); kept upstream's public accessor and added our guard + interleaved-entry cleanup	if ( bufferAttribute !== attribute ) backend.delete( bufferAttribute );

---

src/renderers/webgpu/utils/WebGPUBindingUtils.js

#	intent (one line)	origin (sha / PRD)	decision	reason	marker
1	Re-create the GPU bind-group layout when the entry set changed, instead of reusing the first layout this bind group ever had	32ebfa8 (PRD-400)	kept	r187 returns bindingsData.layout.layoutGPU whenever a layout exists; deleteBindGroupData( bindGroup ) exists at r187 line 622	if ( bindingsData.layoutKey === bindGroupLayoutKey ) return bindingsData.layout.layoutGPU;
2	Keep the binding description with the GPU bind group for the transient-depth proof	f30f9b6 / 1d056ce	kept	_tnGroups is filled by row 5	bindingsData._tnBindings = this._tnGroups?.get( bindGroupGPU );
3	Mirror the bytes actually written to a uniform buffer (including partial ranges)	f30f9b6 / 1d056ce	kept	r187's updateBinding only issues writeBuffer	data._tnUniformBytes = values.slice();
4	A raw pass attachment binds a fresh 2D view of the MSAA/depth texture instead of the default view	f1040a1 (PRDs 269/455)	kept	r187's sampled-texture branch starts at textureData.externalTexture !== undefined; textureData.msaaTexture still exists in r187 (WebGPUTextureUtils.js:467) alongside the new msaaTextures array used for layered targets	textureData.rawTextureView = raw.createView(
5	Capture the binding descriptions of a new GPU bind group	f30f9b6 / 1d056ce	kept	r187 returns createBindGroup( … ) without bookkeeping	this._tnGroups.set( bindGroupGPU,
6	Raw attachments count as multisampled when the sampler/bind-group state is computed	f1040a1 (PRDs 269/455)	kept	r187 checks primarySamples > 1 only	`if ( primarySamples > 1
7	A multisampled FloatType texture cannot use a filtering sampler even when float32-filterable exists	f1040a1 (PRDs 269/455)	kept	r187's guard is this.backend.hasFeature( 'float32-filterable' ) only	if ( texture.multisampled !== true && this.backend.hasFeature( 'float32-filterable' ) ) {
8	Deleting a bind group's layout data also drops the cached GPU bind groups built against it	32ebfa8 (PRD-400)	kept	r187's deleteBindGroupData clears layout/layoutKey and leaves groups/versions pointing at bind groups of the old layout	bindingsData.groups = [];

---

src/renderers/webgpu/utils/WebGPUPipelineUtils.js

#	intent (one line)	origin (sha / PRD)	decision	reason	marker
1	Capture the immutable pipeline proof (depth state, topology, cull, position format/offset/stride, compiled programs) used by the transient-depth speculation, and record the pipeline object on it	1d056ce (#486) / f30f9b6 (PRD-478)	kept	r187 builds the descriptor and the pipeline with no such record; backend.get( vertexProgram ) / _tnShaderCode are provided by this arm	pipelineData._tnDepthProof = depthStencil.depthWriteEnabled === true
2	Pop the error scope before awaiting the pipeline and handle its rejection, so an unrelated draw cannot enter the scope and no rejection escapes as unhandled	2d21247 (PRDs 466–468)	ported	r187 already pops before awaiting (const errorScopePromise = device.popErrorScope(); + comment); kept that ordering and added our .catch( … asyncError ??= err ) so a rejected scope is reported instead of thrown	asyncError ??= err;

---

src/renderers/webgpu/utils/WebGPUTextureUtils.js

#	intent (one line)	origin (sha / PRD)	decision	reason	marker
1	The default canvas owns its private depth variants (one per effective sample count) until resize/dispose	1d056ce (#486)	kept	r187 has no such cache; DepthTexture import comes with it	this._depthBuffers = new Map();
2	A 3D texture that is not rendered into must not ask for RENDER_ATTACHMENT (wgpu-native rejects 3D render attachments)	afee25e (PRD-494)	kept	r187's condition is isCompressedTexture !== true && isCompressedArrayTexture !== true && format !== RGB9E5UFloat; dimension, options.needsMipmaps and options.renderTarget all exist in r187's createTexture	const notAnAttachment3D = dimension === GPUTextureDimension.ThreeD &&
3	Destroying the canvas depth texture tears down the whole cache (private buffer, listeners, reused variants)	1d056ce / f30f9b6	kept	r187's destroyTexture only destroys that one texture (plus its msaa textures)	depthBuffers._tnPrivateDepth?.texture.destroy();
4	The canvas colour texture is only reusable when its GPU texture exists, and it is sized through image, not source	2d21247 (PRDs 466–468)	dropped-upstream	r187 made both changes: if ( colorTextureData.texture !== undefined && colorTexture.width === width && … ) and colorTexture.image.width = width;	-
5	getDepthBuffer() returns a cached, correctly formatted depth buffer per sample count instead of recreating on every mismatch	1d056ce (#486)	ported	r187 still has the 0.185.1 body (early-out on width/height/samples/depth/stencil, then createTexture( depthTexture )). Our replacement was written against that body and is kept whole; it only reads currentSamples, getSampleCount, createTexture/destroyTexture, all unchanged in r187. Upstream's depth/stencil early-out is intentionally folded into the format/type cache check	let cache = this._depthBuffers.get( depthTexture );
6	updateTextureAsync() — tile the compressed mip data under a host-time and bytes-per-write budget, scope the writes, revalidate before publishing	2d21247 (PRDs 466–468)	kept	Pure addition; writes through device.queue.writeTexture directly, so upstream's writeTextureLayer( …, dataOffset, … ) signature change and the new _getBlockData() fallback do not affect it (_getBlockData is now total, which only weakens our "unsupported allocation" guard)	async updateTextureAsync( texture, options = {} ) {

---

Doubts, ranked by risk

1. Renderer.js rows 13/14 (always-run vs RenderObjectRefreshType) — highest risk. I dropped our "every step of a draw runs" in favour of r187's FULL/SHARED/NONE gate, because r187's NodeMaterialObserver.equals() compares the same signals our settled path compared (attribute and index versions, instance matrix/colour versions, morph-texture version, output-target version, world matrix) and the observer arm replaced object.static with isSettled. If that equivalence is wrong for one case (a geometry attribute written while the refresh type stays NONE), the symptom is a stale vertex buffer. The defensive code our patch put inside Bindings._update / Geometries.attributesChanged is still present and still correct — it just no longer runs on NONE frames.
2. PMREMGenerator.js row 3 (largest rewrite). r187 replaced the CubeUV generator with a per-mip CubeCamera design, so the per-face fence had to be re-expressed. My _renderCubeAsync() duplicates CubeCamera.update()'s per-face body (coordinate-system update, mipmapsAutoUpdate off, xr off, MRT null, reversed-depth clear, render-target/cube-face/mipmap restore) and sets needsPMREMUpdate. It is the only place in the arm where I wrote rendering code that upstream does not have. Everything else in that file is r187 code with awaits; _applyPMREMAsync deliberately duplicates _applyPMREM's loop (lodBias math included) to keep the upstream function untouched — a future upstream change to that math must be mirrored.
3. invariantPosition vs upstream invariant (WGSLNodeBuilder row 4). I accept both, but the engine's templates ("renderer.contextNode = context({ invariantPosition: true })") depend on the old name; if the engine ever renames the flag, both builders must move together (WGSL here, GLSL in another arm's file — the GLSL builder in r187 only reads context.invariant, so a invariantPosition-only scene gets @invariant in WGSL and nothing in GLSL. That asymmetry already existed in 0.185.1, where the engine only patched the WGSL builder).
4. Bindings.js row 5 (settled sampled-texture generation). The settled branch sets needsBindingsUpdate but appends neither the texture id nor its version to the (cacheKey, version) pair WebGPUBindingUtils.createBindings uses, so a re-created texture could still hit the cached GPU bind group. This is our patch's behaviour as written; upstream's own non-settled path has the same hole for a texture that is re-created without a version bump, so I did not redesign it here.
5. WebGPUTextureUtils.js rows 5/6 and _getBlockData. r187 made _getBlockData() total (uncompressed fallback), so updateTextureAsync's "unsupported allocation" throw can no longer fire for a format it does not list; harmless for the compressed-only caller but the guard is now weaker than its comment.
6. WebGPUBindingUtils.js row 8 sets groups/versions to [] (our patch) while Textures._destroyTexture sets them to undefined. Both invalidate the cache; the sibling path is the one r187 owns, so I left both as they are rather than harmonise.
7. Cross-arm dependencies I could not edit (verified to exist in the merged tree at the time of writing): NodeMaterialObserver.isSettled (line 376), TextureNode.generateTextureLoad( …, isRawTextureNode ) (line 635) and PassNode.isRawTextureNode / _rawTexturesDisposed, Material._alphaTest. If any of those arms drops its half, the ported hunks that read them go inert.