fix(sycl): register device images in Windows MSVC builds

[lusoris]

Summary

The native Windows MSVC+SYCL build now runs its SYCL kernels. It had never run on a GPU (the CI leg only builds), and on its first run on an Arc B580 and a UHD 770 every kernel submit failed with No kernel named ... was found, because Meson links MSVC builds with link.exe, which never registered the device images. MSVC builds now run one icpx -fsycl -fsycl-link step whose object vmaf.lib pulls into every SYCL program (ADR-1364). Running the tests natively also showed that scripts/ci/run_meson_test.py returned 0 on Windows before Meson finished, so the Windows MinGW64 and ARM64 MSVC lanes stopped gating after 17 to 19 tests; the runner now waits, and the four Windows test-harness failures it hid are fixed.

What changed:

• core/src/meson.build: on MSVC-syntax toolchains the SYCL TUs compile as relocatable device code, one device link (AOT device list, --offload-compress, -fp-model=precise, 8 parallel link jobs) wraps every image, and core/src/sycl/coff_add_anchor.py gives that object the external symbol vmaf_sycl_device_images, which core/src/sycl/common.cpp /includes. -fsycl leaves the MSVC link line (/IGNORE:4078 instead, as on the driver's own link). Linux keeps the ADR-1360 per-TU code generation.
• test_sycl_kernel_registration asks the SYCL runtime for its kernel IDs (0 before, 96 now) without a GPU; the Windows MSVC+SYCL leg runs it.
• scripts/ci/run_meson_test.py: Windows runs Meson as a child and returns its status; POSIX keeps the ADR-1333 exec.
• Windows harness fixes: test_vmaf_per_shot_input (CRT invalid-parameter fast-fail), test_device_target_header_dependencies (shell syntax in Ninja rules), test_vmaf_per_shot (/dev/zero and FIFO cases are POSIX-only), dnn tests (WSL bash.exe, backslash path spliced into Python).
• Parity gate: the cambi cell read Cambi_feature_cambi_score, which vmaf --json writes as cambi.
• Docs: new SYCL on Windows how-to; overview, backends index, oneAPI page and nav link to it.

Results on the i9-12900K host (details in Research-2125):

Check	Arc B580	UHD 770
sycl suite before / after	3/50 / 51/51	not run / 51/51
whole suite (runner fixed)	240 OK, 1 skipped (fixture)	240 OK, 1 skipped (fixture)
default model, pooled VMAF vs Windows CPU (576x324 / 4K)	7.95e-7 / 2.85e-7	same, bit-identical to B580
vmaf_v0.6.1, pooled VMAF vs Windows CPU (576x324 / 4K)	2.48e-5 / 2.05e-5	same
19 SYCL extractors vs CPU (parity gate, 576x324 and 4K)	all within tolerance	all within tolerance
Windows SYCL vs Linux container SYCL (same revision)	features bit-identical, VMAF ≤ 6.1e-12	same

Type

• fix — bug fix
• build / ci — tooling / infra
• sycl / cuda / simd — backend-specific

Checklist

• Commits follow Conventional Commits (the commit-msg hook enforces this).
• make format && make lint is green locally. Run per tool on the changed files (clang-format 23.1.1, black 26.5.1, ruff 0.16.9, shellcheck 0.11.0.1, shfmt 3.13.1, markdownlint-cli2 0.23.3), plus make docs-fragments-check, the state-row, ADR-link, research-ID and citation gates; the full make lint is POSIX-only on this Windows host.
• Unit tests pass: python3 scripts/ci/run_meson_test.py -- -C build.
• If I touched any SIMD/GPU code path, I ran /cross-backend-diff and the worst ULP is ≤ 2. No kernel changed; SYCL output is bit-identical to the Linux build of the same revision, and the parity gate results are in the table above.
• If I touched a feature extractor with SIMD/GPU twins, I either updated every twin or listed the gap under "Known follow-ups" below. No extractor changed.
• If I added a new .c / .cpp / .cu / .h / .hpp, it has the appropriate license header (see CONTRIBUTING.md).
• If this is a breaking change, the commit message uses ! or BREAKING CHANGE: and the migration path is documented below. Not breaking.
• If this PR adds an ADR, the ADR row lives in docs/adr/_index_fragments/<NNNN-slug>.md and the slug is appended to docs/adr/_index_fragments/_order.txt — do not edit docs/adr/README.md directly (regenerated by scripts/docs/concat-adr-index.sh; see ADR-0221).

Bug-status hygiene (ADR-0165)

• docs/state.md updated in this PR: closed T-SYCL-WINDOWS-MSVC-KERNELS-UNREGISTERED-2026-09-29, T-CI-WINDOWS-MESON-TEST-RUNNER-EXIT-0-2026-09-29, T-TEST-WINDOWS-HARNESS-MASKED-FAILURES-2026-09-29, T-CI-PARITY-GATE-CAMBI-KEY-2026-09-29; opened (RC3) T-CI-PARITY-GATE-MOTION-DEBUG-DEFAULT-2026-09-29, T-SYCL-FLOAT-SSIM-XE2-XELP-CALIBRATION-2026-09-29.

Netflix golden-data gate (ADR-0024)

• I did not modify any assertAlmostEqual(...) score in the Netflix golden Python tests.
• If I believe a golden value must change, I have explained why below AND pinged @lusoris for a CODEOWNERS exception. No golden value changes.

Cross-backend numerical results

parity gate, --backends cpu sycl, Netflix pair (576x324) / BBB 4K, Arc B580 = UHD 770
adm 1e-6/2e-6  cambi 0/0  ciede 1.2e-5/9.7e-5  float_adm 2.5e-5/1.6e-5  float_moment 0/0
float_motion 3e-6/2.7e-5  float_ms_ssim(+lcs) 1e-6/1e-6  float_psnr 0/0  float_ssim 1e-6/-(scale 8)
float_vif 2.8e-5/3e-6  motion_v2 0/0  psnr 0/0  psnr_hvs 8.3e-5/8.43e-4 (tol 3.34e-3)  ssimulacra2 0/0  vif 1e-6/1e-6
by name at --precision max: motion_sycl 1.26e-5/5.6e-6  integer_ssim_sycl 1.5e-8/7.8e-8  speed_chroma/temporal 0/0

Performance (if perf or feat)

Diagnostic only (native builds are not canonical bench rows, ADR-1102). Default model, (t(N) - t(2)) / (N - 2), minimum of five interleaved runs: 576x324 B580 2.56 ms, UHD 770 5.30 ms, CPU 2.48 ms (1 thread) / 0.51 ms (16); 4K B580 44.8 ms, UHD 770 66.9 ms, CPU 77.5 / 20.4 ms. The Linux container build of the same revision measures 2.38 / 10.56 ms and 52.3 / 76.1 ms. At 4K the B580 waits on the CPU ADM branch (no AIM device pass, T-GPU-ADM-AIM-DEVICE-PASS-MISSING-SYCL-HIP-2026-09-05); vmaf_v0.6.1, all on the device, takes 9.3 ms per 4K frame on the B580.

Deep-dive deliverables (ADR-0108)

• Research digest — docs/research/2125-windows-native-sycl-run.md.
• Decision matrix — in ADR-1364 ## Alternatives considered.
• AGENTS.md invariant note — core/src/sycl/AGENTS.md "MSVC builds register images through one explicit device link (ADR-1364)".
• Reproducer / smoke-test command — pasted below under "Reproducer".
• CHANGELOG fragment — changelog.d/fixed/windows-sycl-native-run.md.
• Rebase note — docs/rebase-notes.md "ADR-1364 — Windows MSVC SYCL device link, Windows test runner exit status".

Reproducer

# Windows, cmd, after the environment in docs/backends/sycl/windows.md
meson setup build core --buildtype release --default-library=static -Denable_float=true -Dcpp_std=c++latest -Denable_cuda=false -Denable_sycl=true
ninja -C build
set ONEAPI_DEVICE_SELECTOR=level_zero:0
python scripts\ci\run_meson_test.py -- -C build --suite sycl --print-errorlogs
build\tools\vmaf.exe -r src01_hrc00_576x324.yuv -d src01_hrc01_576x324.yuv -w 576 -h 324 -p 420 -b 8 --backend sycl --json -o out.json

Known follow-ups

• The two Windows test lanes now gate their whole suites for the first time. Failures specific to MinGW-w64 gcc or the ARM64 MSVC toolset could not be reproduced on this x64 host and will show up in this PR's checks.
• The device link was verified with oneAPI 2025.1.1; the CI leg's 2025.3.0 builds it and checks the kernel registry, but no GPU has run a 2025.3.0 Windows build.
• T-CI-PARITY-GATE-MOTION-DEBUG-DEFAULT-2026-09-29 and T-SYCL-FLOAT-SSIM-XE2-XELP-CALIBRATION-2026-09-29 (RC3).

🤖 Generated with Claude Code