[ric-oliv]

The two Unsupported AOT jobs check that Sentry turns SentryNative off on runtimes that have no native SDK, linux-arm and win-x86. They wait for the whole build matrix, because they test the macOS leg's packages. On #5788 they started 14 minutes after macOS had uploaded those packages, as soon as the slowest leg finished, and added 1.5 minutes after the matrix.

They don't need the macOS packages, though. Every leg already packs its own and runs the same aot.Tests.ps1 against them in its integration tests. The macOS artifact was simply the package that was around when #4347 added the jobs.

This PR runs the test inside the legs instead:

• linux-arm in the linux-arm64 leg, which uses the same ubuntu-24.04-arm runner.
• win-x86 in the win-x64 leg, which uses windows-latest like the old job.
• The unsupported-aot job is gone, and so is the tail after the matrix.

The linux-arm job also had a flake. 4 of the last 57 attempts failed, and the one whose log was still around failed with "Exec format error". The test app built, but the host couldn't start a 32-bit ARM binary, and the same commit had passed on its first attempt. Some arm64 hosts can't run 32-bit ARM code, so the test no longer depends on the host:

• The leg registers QEMU for 32-bit ARM with docker/setup-qemu-action, pinned like in alpine.yml. QEMU then runs the test app and its container on every host.
• A 32-bit ARM process sees aarch64 when the arm64 kernel runs it and armv7l under QEMU. The leg compiles a tiny uname program, logs what the host does natively next to its CPU part, and fails unless QEMU runs it. So every run proves the QEMU path, and the log tells us which hosts caused the flake.

On today's main, win-x86 adds about 1.3 minutes to win-x64, which is one of the slowest legs until #5782 lands.

Measured on the first run:

• The tail after the matrix is gone. The last legs finished at 23:14 and the workflow ended there.
• On linux-arm64, the four new steps took about 1 minute: 18 seconds for the dependencies, 4 for QEMU, 38 for the test. Aot passed in 14.7 seconds and Container in 16.3, both under QEMU.
• That host was a Neoverse N2, CPU part 0xd49, and it ran 32-bit ARM natively, printing aarch64. So the flaky hosts are a different CPU, and the next one will show up in this log.
• On win-x64 the test took 38 seconds. Aot passed and Container was skipped, as in the old job, because it only runs on Linux.
• The workflow still took 42.5 minutes, because linux-x64 spent 19 minutes in Setup Environment on a slow Mono download. ci: stop installing Mono #5794 removes that download.
• The run after the comment fix matched: the QEMU check passed, the test took 36 seconds, and linux-x64 hit the same slow download again, with Setup Environment at 13.5 minutes.

#skip-changelog

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 76.12%. Comparing base (b41b96c) to head (3803459).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #5791      +/-   ##
==========================================
- Coverage   76.13%   76.12%   -0.02%     
==========================================
  Files         528      528              
  Lines       19579    19579              
  Branches     3829     3829              
==========================================
- Hits        14906    14904       -2     
- Misses       3791     3792       +1     
- Partials      882      883       +1     

☔ 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.