Add RenderTarget depthResolveMode controlling multisampled depth resolve

[mvaligursky]

Adds control over how the samples of a multisampled depth buffer are resolved into a single depth value by the WebGPU shader-based depth resolve - used by the depth grab pass (sceneDepthMap), a depth RenderTarget#copy, and the automatic resolve into a provided depthBuffer.

const rt = new RenderTarget({ colorBuffer, depth: true, samples: 4, depthResolveMode: DEPTHRESOLVE_MAX });
rt.depthResolveMode = DEPTHRESOLVE_SAMPLE0;  // mutable at any time, used at resolve time

Behavior change: the default resolve changes from sample 0 to DEPTHRESOLVE_MIN. Sample 0 is an arbitrary sample position - at geometry silhouettes it lands on the foreground or background essentially at random per pixel, causing edge noise in depth-consuming effects such as soft particles. MIN (the nearest surface, with a standard depth buffer) is deterministic and conservative, and matches what depth consumers such as Hi-Z occlusion expect. DEPTHRESOLVE_SAMPLE0 remains available to restore the previous behavior.

Changes:

• New constants DEPTHRESOLVE_MIN (default), DEPTHRESOLVE_MAX and DEPTHRESOLVE_SAMPLE0, and a mutable RenderTarget#depthResolveMode property with a matching constructor option. WebGPU only - on WebGL2 the depth resolve sample selection is defined by the implementation and the property is ignored.
• WebgpuResolver supports all three modes, with shaders and pipelines lazily cached per mode and format. Its WGSL source is extracted into a shader chunk (shader-chunks/frag/webgpu-depth-resolve.js), with the mode selected by a DEPTH_RESOLVE_*** define handled by the shader preprocessor - the first of the internal device-level renderers to move its shader to the platform chunks.
• Fixed a crash in copyRenderTarget when copying depth from a render target with a user-provided depthBuffer and samples > 1 - the multisampled depth of such a target is stored separately from the internally-allocated case, and the copy read a null texture.

Verified on WebGPU (Metal): a known-depth scene rendered into an MSAA target and resolved with each mode - at all 100 mixed-sample edge pixels, MIN returned the near surface depth, MAX the far clear value, and SAMPLE0 a value within that range; mode switching at runtime works. The ground-fog and particles-snow examples exercise the new default through the MSAA depth grab.

Checklist

• I have read the contributing guidelines
• My code follows the project's coding standards
• This PR focuses on a single change

[github-actions]

Public API report

This PR changes the public API surface (+6 / −1), per the docs' rules (@ignore / @Private / undocumented are excluded).

Show API diff

-RenderTarget.constructor(options?: { autoResolve: boolean; colorBuffer: Texture; colorBuffers: Texture[]; depth: boolean; depthBuffer: Texture; face: number; mipLevel: number; name: string; origin: string; resolveBuffer: Texture | null; resolveBuffers: Texture | null[]; samples: number; stencil: boolean; transientColor: boolean; transientDepth: boolean })
+RenderTarget.constructor(options?: { autoResolve: boolean; colorBuffer: Texture; colorBuffers: Texture[]; depth: boolean; depthBuffer: Texture; depthResolveMode: string; face: number; mipLevel: number; name: string; origin: string; resolveBuffer: Texture | null; resolveBuffers: Texture | null[]; samples: number; stencil: boolean; transientColor: boolean; transientDepth: boolean })
+RenderTarget.get depthResolveMode(): string
+RenderTarget.set depthResolveMode(value: string)
+const DEPTHRESOLVE_MAX: "max"
+const DEPTHRESOLVE_MIN: "min"
+const DEPTHRESOLVE_SAMPLE0: "sample0"

Informational only — this never fails the build.

[github-actions]

Build size report

This PR changes the size of the minified bundles.

Bundle	Minified	Gzip	Brotli
playcanvas.min.js	2385.9 KB (+0.6 KB, +0.03%)	613.2 KB (+0.4 KB, +0.06%)	475.7 KB (+0.4 KB, +0.09%)
playcanvas.min.mjs	2383.3 KB (+0.6 KB, +0.03%)	611.9 KB (+0.4 KB, +0.07%)	475.1 KB (+0.4 KB, +0.08%)

[mvaligursky]

Automated PR review by Codex (GPT-5) at exact head 3ef58df615f34b9d79f8f40d6ffb3adffb043364.

No actionable issues found. I reviewed the public constants and RenderTarget API, the intentional default behavior change, all shader-based resolve call paths (scene depth grab, depth copy, and automatic resolve), user-provided versus internally allocated MSAA depth attachments, WGSL reduction logic, direction of min/max depth semantics, runtime mode switching, shader/pipeline cache keys and cleanup, destination formats, performance, generated types, tests, CI, and current discussion.

Verification performed:

• Full unit suite: 2,429 passing, 2 pending
• Focused render-target/WebGPU resolver tests: 40 passing
• Changed-file ESLint: clean
• Type generation and type tests: passing
• Deployed ground-fog MSAA depth-grab path: rendered successfully on WebGPU with no relevant shader, pipeline, or validation errors
• All required CI and deployment checks: green

Residual test risk: the automated resolver tests verify the preprocessed WGSL for all three modes, but do not numerically render and read back mixed-sample pixels on a real GPU. The PR description reports that numerical WebGPU coverage was performed manually; the deployed default MIN path also compiled and ran successfully during this review.