[akacross]

Unblocks the Validation Script Smoke job. It currently fails on main (e.g. c302254b) and on every PR, so all platform jobs are skipped. Fixing the first failure exposed three more behind it. This PR fixes all four.

1. Native harnesses need headers the smoke job never downloads

• android_console_input.py, added to the smoke list in 1bc4c14, required a full SDL3 tree: RuntimeError: SDL3 headers required.
• openxr_lifecycle.py compiled against subprojects/OpenXR-SDK-*/include: fatal error: 'openxr/openxr.h' file not found. On main this test was never reached.

Both now get headers through a new tools/tests/pinned_archive_headers.py, which follows the convention sdl3_clipboard_status.py uses for its sources:

• An existing provisioned tree is used as it is.
• Otherwise it takes the wrap's pinned archive from subprojects/packagecache or an HTTPS download, checks it against source_hash and a size bound, and writes only include/<dir>/*.h members (bounded in count and size) into the check's temporary directory. The archive is never unpacked wholesale.
• Its offline self-test covers extraction, tamper refusal and the provisioned-tree preference.
• --sdl-include still works.

A scan of all 393 smoke checks found no others reading an unprovisioned subproject. glew, stb_vorbis and openal-soft-prebuilt are committed to the repo.

2. validation_hardening.py: three checks were never wired in

ui_capture_multiplayer_rosters.py (1d0cbb8) and ui_capture_system_refresh.py (6a4696c) were wired nowhere, and ui_player_identity.py (c302254) was missing from openq4_validate.py. All three are offline and pass, and they're now wired like their neighbours.

3. macOS wording checks after the documentation rewrites

macos_matrix_policy.py, macos_metal_bridge.py and macos_native_backend_containment.py still expected sentences that c5f23a5 (Mac Source Ports credit) and 4054e7a (v0.13.2 notes) replaced. The stated policy hasn't changed:

• The getting-started checks now follow the guide's new wording: GitHub packages are an Apple Silicon/arm64 preview, there are no Intel or universal2 packages, and Rosetta is unsupported.
• README.html checks are unchanged.
• The v0.13.2 note gets back the canonical "Metal bridge around the OpenGL renderer, not a native Metal renderer" and "comparison-only diagnostic infrastructure, not a release backend" phrases that every other release note carries, with no change in meaning.

Verification (Linux x64)

• Full smoke list run locally (398 commands, parallel): passes except nine native harnesses that only this machine's clang 23.1 / GCC 16 headers reject (-Wunused-but-set-global under -Werror, and a missing <cstdint> in SystemDisplay.h). CI's compilers don't raise either.
• Simulated smoke runner (no provisioned subprojects, only wraps and the cached archives): 86 SDL3 and 7 OpenXR headers extracted; the android check and all 19 OpenXR lifecycle cases compile and pass.
• HTTPS archive path: not exercised locally. CI will download the archives, which is the same path sdl3_clipboard_status.py already uses in this job.

🤖 Generated with Claude Code

[themuffinator]

Thanks, this is the fix every PR has been waiting on. I've pushed two commits to the branch:

• ff3c9d7 merges main. Both conflicts were in the macOS wording, which main's 8cb58aa had already restored the same way, so they take main's side. Part 3 of the description is therefore already on main; the pinned headers and the newly wired checks remain.

• 8b8d9e4 fixes the macOS failures that your previous run exposed once the smoke job got through. android_async, android_lifecycle and android_console_input compiled with -D__ANDROID__, and Apple's libc++ then includes <android/api-level.h> from its own headers. The console harness was also missing <initializer_list>. Each harness now includes its standard headers with __ANDROID__ undefined, as the console harness already did for SDL, and defines it again for the engine code.

  Locally I used clang with LLVM libc++ plus a shim that fails any standard header read under __ANDROID__. All three failed before the change and pass after it. They also pass with g++/libstdc++ and Windows clang.

I'll merge once this run is green.

[akacross]

Thanks for the merge and the Android-header fix. The smoke job is green, and the platform jobs ran for the first time in days.

None of the six red jobs failed a test. All six were cancelled at the 90-minute PR limit:

Job	Cancelled in
Linux x64 GCC Wayland libdecor / no-libdecor	Run native Wayland validation
macOS Intel x64 OpenGL / Metal	Run selected validation profile
Windows ARM64	Run selected validation profile
Linux x64 ASan+UBSan	Run sanitizer build

In the libdecor job, for example, the 408 Python checks ran one after another from 09:02 to 10:18 (76 minutes), and the native tests that follow were cut off. That's the time #191 recovers: its own run took 15–19 minutes for the same checks with 4 jobs.

So this run can't go fully green with main's current check runner, and #191's smoke job can't pass until this lands. Two ways out of that loop:

1. Merge Unblock the script smoke job: pinned SDK headers, unwired checks, macOS wording #202 now on the evidence above: every job that completed passed, including the smoke job, Windows x64, both macOS ARM64 jobs and all four Linux ARM64 jobs. Then merge Run independent validation checks in parallel #191, after which the next runs fit the limit.
2. Or I can merge Run independent validation checks in parallel #191's branch into this one, so a single run shows the result with both.

Separately, the sanitizer job's 150 minutes from #200 didn't make it to main (it's 90 again). Its build alone took about 82 of those minutes in earlier runs, so it'll keep being cut off until it gets that time back. I can add that to this PR or to #191 if you'd like.