feat: Validation Phase C — feature gates, texture usage, indirect count (#333)

[lkmavi]

Summary

• Implements Validation Phase C (feat: validation Phase C — expand to ~70% Rust wgpu-core parity #333): usage-time feature gates and texture usage validation toward ~70% Rust wgpu-core parity.
• Adds core.RequireFeature / FeatureError, a 25-feature registry (AllFeatureRequirements), format/shader/query gates, texture usage matrix, Float32Filterable bind-time checks, SPIR-V OpCapability introspection, and public MultiDrawIndirectCount / MultiDrawIndexedIndirectCount APIs with Vulkan backend support.
• Introduces tiered test infrastructure: core/testutil.ValidationEnv, registry-driven featureGateCases, smoke tests in wgpu_feature_gate_test.go, and layered indirect-count unit tests (HAL / core / public API). Docs: docs/VALIDATION-TESTING.md.

Test plan

• go test ./... -count=1
• go test ./core/... -run TestFeatureGates -v
• go test ./hal/... ./core/... . -run 'IndirectCount|ValidateIndirect|RecordIndirectCount' -count=1
• GOGPU_GRAPHICS_API=software go test . -run FeatureGate -v
• CI green on all backends

[codecov]

Codecov Report

❌ Patch coverage is 85.24871% with 86 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
query_native.go	44.18%	22 Missing and 2 partials ⚠️
core/shader_features.go	73.46%	9 Missing and 4 partials ⚠️
core/feature_format.go	80.00%	8 Missing and 2 partials ⚠️
core/testutil/assert.go	33.33%	5 Missing and 5 partials ⚠️
core/validate.go	76.19%	4 Missing and 1 partial ⚠️
device_native.go	91.80%	0 Missing and 5 partials ⚠️
renderpass_indirect_count.go	87.17%	3 Missing and 2 partials ⚠️
core/spirv_capabilities.go	87.87%	2 Missing and 2 partials ⚠️
renderpass_native.go	96.00%	3 Missing and 1 partial ⚠️
core/texture_usage_validate.go	84.21%	2 Missing and 1 partial ⚠️
... and 2 more

📢 Thoughts on this report? Let us know!

[kolkov]

Excellent work, @lkmavi! ~2,300 LOC, 15+ validation checks, ~80 new test cases, enterprise-quality test infrastructure (ValidationEnv, featureGateCases registry with completeness guard). The indirect count three-layer implementation with Vulkan native vkCmdDrawIndirectCount is particularly well done.

Two issues to fix before merge:

Blockers

B1. VAL-C16: Depth/stencil + RenderAttachment incorrectly rejected

// core/texture_usage_validate.go
if format.IsDepthStencil() {
    const forbidden = gputypes.TextureUsageRenderAttachment | gputypes.TextureUsageStorageBinding
    if usage&forbidden != 0 {
        return &CreateTextureError{Kind: CreateTextureErrorInvalidUsage}
    }
}

The W3C spec explicitly says RENDER_ATTACHMENT is used for both color AND depth/stencil attachments (spec line 4510): "The texture can be used as a color or depth/stencil attachment in a render pass." Rejecting depth textures with RenderAttachment usage breaks every application with depth testing. Rust wgpu does NOT reject this at texture creation time.

Fix: remove TextureUsageRenderAttachment from the depth/stencil forbidden mask:

if format.IsDepthStencil() {
    if usage.Contains(gputypes.TextureUsageStorageBinding) {
        return &CreateTextureError{Kind: CreateTextureErrorInvalidUsage}
    }
}

B2. VAL-C17: Storage + RenderAttachment blanket rejection

if usage.Contains(gputypes.TextureUsageStorageBinding) &&
    usage.Contains(gputypes.TextureUsageRenderAttachment) {
    return &CreateTextureError{Kind: CreateTextureErrorInvalidUsage}
}

The W3C WebGPU spec does NOT prohibit this combination at texture creation time. Usage conflicts are detected at bind-time and draw-time, not creation time. Rust wgpu does not have this check. A texture can be RGBA8Unorm with both Storage and RenderAttachment usages — valid per spec.

Fix: remove this check entirely, or narrow it to specific invalid cases (if any exist).

---

Everything else is solid:

• RequireFeature pattern matches Rust require_features() (resource.rs:429)
• Feature gates for BC/ETC2/ASTC/DepthClipControl/ShaderF16/Timestamp all correct
• Indirect count buffer validation thorough (nil, usage, alignment, bounds)
• Vulkan vkCmdDrawIndirectCount with proc address loading correct
• SPIR-V capability scanning handles malformed modules gracefully
• Test infrastructure (featureGateCases registry with completeness guard) is reusable and excellent
• CI 15/15 green

Happy to re-review after the two texture usage fixes!

[lkmavi]

Addressed both blockers from the review:

• B1 (VAL-C16): depth/stencil now only rejects StorageBinding; RenderAttachment is allowed (W3C: color or depth/stencil attachment).
• B2 (VAL-C17): removed the blanket Storage+RenderAttachment rejection at creation; conflicts stay bind/draw-time. Dropped unused dimension param.

Tests updated (StorageAndRenderAttachment expects success; added DepthRenderAttachment). Ready for re-review.

[kolkov]

Both blockers verified as fixed:

B1: Depth/stencil now only rejects StorageBinding; RenderAttachment allowed (W3C spec line 4510).
B2: Blanket Storage+RenderAttachment rejection removed. Test confirms RGBA8 with both usages succeeds.

15/15 CI green. ~2,300 LOC, 15+ validation checks, ~80 tests, enterprise test infra. Excellent work, @lkmavi!