test(metal): align storage-mode fixture with GPU policy

[besmpl]

Summary

• drive the layered Queue.WriteTexture storage-mode matrix through Device.isAppleGPU, the predicate production texture allocation actually uses
• skip the Shared-direct cases when the host does not expose an Apple GPU family
• leave Metal production behavior unchanged

Root cause

The test added in #261 simulated Private and Shared texture allocation by changing hasUnifiedMemory. After #272, texture allocation intentionally uses Apple GPU-family membership (isAppleGPU) instead, avoiding Shared texture policy on Intel/AMD devices.

On Apple Silicon, the stale fixture therefore could not force the private staging case: CreateTexture still saw isAppleGPU=true, allocated Shared storage, and failed the expectation:

texture shared mode = true; want false

This updates the test injection point to match the current policy. The Private-staging and Shared-direct data paths remain exercised on Apple GPUs; non-Apple GPUs continue to exercise the production-supported Private path.

Validation

• CGO_ENABLED=0 GOWORK=off go test -count=1 ./hal/metal -run '^TestQueueWriteTextureLayeredShapes/1D_array/private_staging$' -v
• full layered storage matrix, repeated three times
• CGO_ENABLED=0 GOWORK=off go test -count=1 ./hal/metal
• CGO_ENABLED=0 GOWORK=off go test -count=1 ./...
• CGO_ENABLED=0 GOWORK=off go test -count=1 -tags rust ./...
• default and Rust-tagged builds, plus the root WASM build

The complete local macOS default and Rust-tagged test runs now pass without excluding hal/metal. This is deliberately separate from Android PR #268.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[kolkov]

Clean fix. The test injection was stale after #272 switched texture storage policy from hasUnifiedMemory to isAppleGPU — this aligns the fixture with the actual production predicate. Verified that CreateTexture (device.go:233) uses d.isAppleGPU, so the old hasUnifiedMemory toggle was a no-op for texture allocation. LGTM.