Run GPU tests on Metal on Apple silicon and in a macos-15 job

[michakraus]

Summary

The test suite now runs GPU tests on a real Apple GPU.

• test/metal/ checks the two device paths that were measured by hand and not tested before: the products of a PoissonTensor with a wrapped device array, 𝕁 * view(A, [1, 2, 3, 4], :) and 𝕁 * B', which ext/GPUArraysCoreExt.jl carries; and the tensor_mat_mul! kernel on MetalBackend(). Each result must stay on the device and match the host in Float32, with scalar indexing off.
• test/runtests.jl runs them last on every Apple-silicon Mac, so a failure on the device hides no host result. Where Metal.functional() is false (a Mac without a usable device, or a sandbox) they skip themselves.
• Pkg.test(test_args = ["metal"]) runs them after the reachability and exports checks, and stops. Asked for that way, a missing device fails the run instead of skipping.
• .github/workflows/Metal.yml runs them that way on macos-15, for Julia min and 1. GitHub's macOS runners have Apple's paravirtual GPU, which Metal.jl supports from 1.10 on macOS 15 or later (JuliaGPU/Metal.jl#789). The job is not a required check.
• Metal (1.10) joins test/Project.toml on every platform. Its precompile workload is guarded by Sys.isapple() && Sys.ARCH === :aarch64, and nothing loads it off Apple silicon.
• Comments and the GPU Support section of docs/src/index.md that said the suite or CI has no GPU test now name what test/metal/ covers.

Note for the required checks. CI's macOS-latest entry (macOS 26, arm64) is required and runs the full suite, so it now runs the Metal tests too.

The sibling change is JuliaGNI/GeometricOptimizers.jl#117. Its Metal job already passed on the macos-15 runner ("Apple Paravirtual device"): 817/817 in 2m46s, under --check-bounds=yes.

Verification

• On an Apple M4 Max with Metal 1.11.1, test/metal/metal.jl passes 5/5.
• With no Metal device, Pkg.test(test_args = ["metal"]) passes the reachability and exports checks, fails on @test Metal.functional(), and runs nothing else. Checked before and after the reorder.

Pre-PR verification

Fixed

• test/runtests.jl: a full run on Apple silicon ran Metal before the subjects, against the ordering rule in the file's header; a Metal failure would have hidden every subject result. Metal now runs last in a full run; only the test_args = ["metal"] path runs it early, and then stops.
• src/layers/positional_encoding.jl:75 ("the suite has no GPU test") and test/parameters/map_to_cpu_tests.jl:2 ("There is no GPU in CI") were made false by this change. Both now say that test/metal/ does not cover them.
• test/runtests.jl: re-wrapped an overlong comment line.
• test/metal/metal.jl: a comment that told the bug's history is now in the present tense.

Unresolved

• None.

Pre-existing

• ext/GPUArraysCoreExt.jl:10–12 tells history in the past tense (nit). Not in this PR.

Checked and clean

• Aqua (project extras, deps compat, stale deps, piracy count 3) and ExplicitImports: 10/10 pass with Metal in the test environment. The in-process ambiguity count is 1 before and after using Metal, so the == 1 gate holds.
• Test-environment resolve (Julia 1.13): Metal adds 23 packages. The only change to a shared package is GPUArraysCore 0.2.1 → 0.2.0, because GPUArrays pins GPUArraysCore = "= 0.2.0". The only difference is neutral_element, which this package does not use.
• docs/src/index.md claims no more than test/metal/ asserts. macOS-latest is a required check, as the CHANGELOG states.
• fatou lint is clean on the changed files apart from an intended duplicate-include advisory: metal/runtests.jl is included on two mutually exclusive paths. No method definition changed.

Not checked

• Metal precompiling on Linux and Windows (the CHANGELOG claim): this PR's CI will show it.
• The resolve on Julia 1.11: a local network error. The min jobs will show it.
• Type stability: no package method changed.

🤖 Generated with Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 86.07%. Comparing base (cf1fbba) to head (7c59a5d).
⚠️ Report is 3 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #307   +/-   ##
=======================================
  Coverage   86.07%   86.07%           
=======================================
  Files          76       76           
  Lines        3016     3016           
=======================================
  Hits         2596     2596           
  Misses        420      420           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.