fix(sycl): mirror the integer-ADM contrast-masking near edge, and extend large-fixture parity to SYCL

[lusoris]

Summary

Follows #1363 (now merged). Rebased onto master; supersedes the auto-closed #1366.

Extends the ADR-1206 large-fixture parity variants from CUDA to SYCL, and fixes the real bug that immediately found.

ADR-1210 — test_sycl_adm_parity was failing on the Arc A380, on its shipped fixture:

integer_adm_scale3_csf_2_...: cpu=0.58175555 sycl=0.58191226 delta=1.57e-04 tol=1.00e-04

The CPU contrast-masking rule is asymmetric — near edge mirrors to index 1, far edge clamps to the last index (integer_adm.c:1009-1012). The SYCL twin clamped both, reading row/column 0 twice and dropping the mirrored sample. It only diverges once a scale's ADM border crop (int)(dim * 0.1 - 0.5) collapses to 0 — band dimensions ≤ 14 — and scale 3 of the 256x144 fixture is a 16x9 band, exactly that regime.

CUDA (adm_cm.cu), HIP (adm_cm.hip) and Metal (iadm_clampx) all already carry this fix from ADR-1167 / PR #1224. SYCL was the only twin that never received it. Green in CI only because the test skips with no SYCL device — every hosted runner.

suite (Arc A380)	before	after
SYCL parity, shipped fixtures	17 pass / 1 fail / 2 skip	18 pass / 0 fail / 2 skip
SYCL parity, 960x540 variants	—	16 pass / 0 fail / 3 skip

Two variants are handled explicitly rather than silently:

• float_ssim_sycl is a documented v1 scale=1-only extractor; its large variant records that contract as a skip, and still fails loudly if the twin ever stops refusing and starts returning a scale=1 score at a decimating resolution.
• test_sycl_motion_add_uv_parity is not registered. It compares CPU float float_motion against fixed-point motion_sycl, so its 2e-4 tolerance is a quantisation budget calibrated for one fixture, not a bit-exactness bound; at 960x540 it lands at 2.30e-04. Raising the tolerance to make it pass would be exactly the test-weakening the fork forbids, so it is tracked as T-SYCL-MOTION-ADD-UV-TOLERANCE-RESOLUTION-2026-09-06 instead.

HIP and Metal large variants are deliberately not registered: Metal cannot be verified on this workstation at all, and shipping test registrations that have never been run is how a lane goes red for reasons nobody has looked at.

Type

• fix — bug fix
• test — test-only
• sycl / cuda / simd — backend-specific

Checklist

• Commits follow Conventional Commits.
• make format && make lint is green locally (pre-commit run --files clean).
• Unit tests pass: meson test -C build — SYCL suites on an Arc A380: 18/0/2 shipped, 16/0/3 large.
• If I touched any SIMD/GPU code path, I ran /cross-backend-diff and the worst ULP is ≤ 2 — equivalent done directly: the CPU-vs-SYCL integer_adm_scale3_csf_2 delta goes from 1.57e-04 to inside the 1e-4 gate, and every other SYCL parity feature passes at both fixtures.
• If I touched a feature extractor with SIMD/GPU twins, I either updated every twin or listed the gap — CUDA, HIP and Metal already had this fix; only SYCL needed it, verified by reading all four.
• If I added a new .c / .h — no new source files.
• If this is a breaking change — not breaking. SYCL integer_adm scores move at coarse-band sizes ≤ 14 px, towards the CPU reference; no fork-added snapshot covers that.
• If this PR adds an ADR, the row lives in docs/adr/_index_fragments/ — index in sync.

Bug-status hygiene (ADR-0165)

• docs/state.md updated — T-SYCL-INTEGER-ADM-CM-NEAR-EDGE-2026-09-06 closed, T-SYCL-MOTION-ADD-UV-TOLERANCE-RESOLUTION-2026-09-06 opened.

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 — none does. The golden gate is CPU-only; this changes a SYCL kernel.

Deep-dive deliverables (ADR-0108)

• Research digest — no digest needed: the method is already written up in docs/research/2032-gpu-parity-resolution-blind-spot.md (added by fix(gpu): clamp the ADM contrast-masking far edge and finish the ssimulacra2 FMA unification #1363) and this is a direct application of it; ADR-1210 carries the specific evidence.
• Decision matrix — ## Alternatives considered in ADR-1210 (including why loosening the 1e-4 tolerance was rejected) and the extended ADR-1206.
• AGENTS.md invariant note — added to core/test/AGENTS.md.
• Reproducer / smoke-test command — below.
• CHANGELOG fragment — changelog.d/fixed/sycl-integer-adm-cm-near-edge.md and changelog.d/fixed/sycl-parity-large-fixture-variants.md.
• Rebase note — docs/rebase-notes.md, "ADR-1206 (SYCL) — which parity tests get a large-fixture variant".

Reproducer

Needs a host with an Intel GPU — without one the tests skip and report green, which is how this survived.

CC=icx CXX=icpx meson setup build core -Denable_sycl=true -Denable_cuda=false
ninja -C build
meson test -C build --suite sycl

On the parent commit test_sycl_adm_parity fails with delta=1.57e-04 tol=1.00e-04; on this branch the suite is 18/0.

Known follow-ups

• HIP and Metal large-fixture variants (unverifiable here; Metal has no hardware on this workstation).
• T-SYCL-MOTION-ADD-UV-TOLERANCE-RESOLUTION-2026-09-06 — needs a derived, resolution-aware fixed-point-vs-float bound, not a bumped constant.
• The SYCL Parity (Arc A380) lane builds --buildtype=release; this was found in a debug build. The two configurations should be reconciled so the lane sees what a developer sees.

🤖 Generated with Claude Code