fix(sycl): revert pseudo-Kahan recurrence and calibrate Arc A380

[lusoris]

Summary

Reverts the PR #865 pseudo-Kahan recurrence in core/src/feature/sycl/ssimulacra2_sycl.cpp which caused pole blow-up to $10^{25}$, NaNs, and score saturation at 100.0, restoring a stable float32 Charalampidis IIR recurrence matching the CUDA twin. Calibrates the Intel Arc A380 (sycl:0x8086:0x56a* and arc:dg2-g10) in scripts/ci/gpu_ulp_calibration.yaml to 5.0e-2 (places=1) per ADR-0985 Option C, accounting for fp64-less accumulation across the 6-scale pyramid on natural video clips.

Type

• feat — new feature
• fix — bug fix
• perf — performance improvement
• refactor — no behavior change
• docs — documentation only
• test — test-only
• build / ci — tooling / infra
• port — cherry-pick from upstream Netflix/vmaf
• sycl / cuda / simd — backend-specific

Checklist

• Commits follow Conventional Commits (the commit-msg hook enforces this).
• make format && make lint is green locally.
• Unit tests pass: meson test -C build.
• If I touched any SIMD/GPU code path, I ran /cross-backend-diff and the worst ULP is ≤ 2.
• If I touched a feature extractor with SIMD/GPU twins, I either updated every twin or listed the gap under "Known follow-ups" below.
• 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.
• 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 with a row in Recently closed (T-SYCL-ARC-SSIMULACRA2-PARITY-2026-06-03).

Netflix golden-data gate (ADR-0024)

• I did not modify any assertAlmostEqual(...) score in the Netflix golden Python tests.

Cross-backend numerical results

feature         cpu ↔ sycl (Arc A380)        tolerance      status
ssimulacra2     delta = 1.211e-02            5.0e-02        PASS (src01 48-frame)
ssimulacra2     delta = 0.000e+00            5.0e-02        PASS (checkerboard 1px)
ssimulacra2     delta = 0.000e+00            5.0e-02        PASS (checkerboard 10px)

Deep-dive deliverables (ADR-0108)

• Research digest — docs/research/0985-sycl-parity-divergence-2026-06-03.md updated with pseudo-Kahan pole analysis and empirical measurements.
• Decision matrix — captured in docs/adr/0985-sycl-parity-divergence-2026-06-03.md under ## Alternatives considered (Option C accepted).
• AGENTS.md invariant note — added to core/src/feature/sycl/AGENTS.md ("ssimulacra2_sycl.cpp IIR recurrence has no running accumulator — never 'Kahan' it").
• Reproducer / smoke-test command — pasted below under "Reproducer".
• CHANGELOG fragment — new file changelog.d/fixed/sycl-ssimulacra2-blur-recurrence.md added.
• Rebase note — entry added to docs/rebase-notes.md.

Reproducer

source /opt/intel/oneapi/setvars.sh >/dev/null 2>&1
meson test -C core/build test_sycl_ssimulacra2_parity -v
python3 scripts/ci/cross_backend_parity_gate.py \
  --vmaf-binary core/build/tools/vmaf \
  --reference /home/kilian/dev/vmaf/python/test/resource/yuv/src01_hrc00_576x324.yuv \
  --distorted /home/kilian/dev/vmaf/python/test/resource/yuv/src01_hrc01_576x324.yuv \
  --width 576 --height 324 \
  --features ssimulacra2 \
  --backends cpu sycl \
  --gpu-id arc:dg2-g10

Known follow-ups

Step 3 (a SYCL device CI lane so the parity test stops skipping on machines without hardware) is gated on maintainer decision D4; tracked as a TODO in ADR-0985.