docs(state): close T-SYCL-ARC-SSIMULACRA2-PARITY-2026-06-03 (fixed by 8b7ae731a)

[lusoris]

Summary

T-SYCL-ARC-SSIMULACRA2-PARITY-2026-06-03 was sitting in docs/state.md "## Open bugs" describing a defect that is not present on master. The row names uncompensated fp32 accumulation in the 3-pole IIR Charalampidis blur of core/src/feature/sycl/ssimulacra2_sycl.cpp::launch_blur; that was fixed on 2026-06-12 by 8b7ae731a ("fix: iter6 bundle", PR #865), which added Kahan compensated-summation state to all three poles — exactly the row's own closure condition (b). The ledger text was written against dd02029f9 (PR #852, 2026-06-08, the calibration-anchor partial fix) and was never refreshed, so the row read as open for three months and cost a re-investigation.

This PR moves the row into "## Recently closed" (one row per bug id, past tense, citing 8b7ae731a), and corrects the matching stale prose in scripts/ci/gpu_ulp_calibration.yaml, whose comment block still asserted "the IIR does not use Kahan compensation". Ledger and prose only — no behaviour change, no tolerance change.

Type

• docs — documentation only

Verification (why the row is closed)

All against origin/master:

1. git show origin/master:core/src/feature/sycl/ssimulacra2_sycl.cpp | grep -n -i 'kahan' → "Kahan (compensated) summation for the IIR recurrence state" and "Kahan-corrected IIR state update", both inside launch_blur<PASS>, which now carries float comp_0/comp_1/comp_2 and the full pattern per pole: const float y0 = (t0 - prev2_0) - comp_0; const float o0 = prev1_0 + y0; comp_0 = (o0 - prev1_0) - y0; (plus the _1 / _2 twins).
2. git log --oneline -S'Kahan' origin/master -- core/src/feature/sycl/ssimulacra2_sycl.cpp → single commit 8b7ae731a (PR fix: iter6 bundle — 6 fixes (wave32 comment, UBSan, TSAN, vmaf-tune CRF, AI scripts, MCP conformance) #865, 2026-06-12), whose message documents the change as bounding per-iteration error to O(eps^2) without fp64 (ADR-0220 compliant).
3. git log --oneline origin/master -- scripts/ci/gpu_ulp_calibration.yaml → last touch dd02029f9 (chore(bundle): 8 worktree branches drain — docs + sycl-parity + ci-promotion + tune-report #852), i.e. before the fix — which is why the YAML prose is stale.

Residual (recorded as prose, not as a defect)

scripts/ci/gpu_ulp_calibration.yaml keeps the sycl:0x8086:0x56a* entry at status: placeholder with ssimulacra2: 5.0e-3, because no Arc A380 re-measurement has been committed since the fix (git show origin/master:testdata/scores_sycl_a380_576.json | grep -c ssimulacra2 → 0). That is an outstanding hardware measurement, not a live code defect. Nothing was loosened: core/test/test_sycl_ssimulacra2_parity.c still has #define PARITY_TOL 5e-3 and the YAML tolerance value is unchanged.

Bug-status hygiene (ADR-0165)

• docs/state.md updated in this PR — the row moves from "## Open bugs" to "## Recently closed", closed by 8b7ae731a. Exactly one table row per bug id; the vacated slot carries the usual <!-- ... moved to Recently closed ... --> marker. The related open row T-SYCL-ARC-ADM2-PARITY-1.1E-4-2026-09-05 gained a one-clause pointer so its "same class as the SSIMULACRA2 row" cross-reference no longer sends a reader back to a closed bug.

Netflix golden-data gate (ADR-0024)

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

Deep-dive deliverables (ADR-0108)

• Research digest — no digest needed: trivial. The investigation is a ledger correction; the original analysis already lives in docs/research/0985-sycl-parity-divergence-2026-06-03.md and the closing evidence is inlined in the state.md row.
• Decision matrix — no alternatives: only-one-way fix. A ledger row describing a fixed defect either moves to the closed section or stays wrong.
• AGENTS.md invariant note — no rebase-sensitive invariants. No code, header, or build surface is touched.
• Reproducer / smoke-test command — pasted below under "Reproducer".
• CHANGELOG fragment — changelog.d/changed/1325-state-md-close-sycl-arc-ssimulacra2-parity.md; CHANGELOG.md regenerated by scripts/release/concat-changelog-fragments.sh --write.
• Rebase note — no rebase impact: ledger and CI-comment prose only, no upstream-mirrored file touched.

Reproducer

git show origin/master:core/src/feature/sycl/ssimulacra2_sycl.cpp | grep -n -i 'kahan'
git log --oneline -S'Kahan' origin/master -- core/src/feature/sycl/ssimulacra2_sycl.cpp
python3 -c "import yaml; yaml.safe_load(open('scripts/ci/gpu_ulp_calibration.yaml')); print('yaml ok')"
pre-commit run --files ./docs/state.md ./scripts/ci/gpu_ulp_calibration.yaml ./CHANGELOG.md

Known follow-ups

Re-run ./build-sycl/test/test_sycl_ssimulacra2_parity on Arc A380 hardware once a self-hosted DG2-G10 runner exists, then either promote the sycl:0x8086:0x56a* entry to status: calibrated with the measured tolerance, or open a fresh docs/state.md row with the post-Kahan number if the delta still exceeds 5e-3. The gate could not be re-run on the audit host: no SYCL build directory and no oneAPI toolchain configured there.

[lusoris]

Superseded — the row is already closed on master.

T-SYCL-ARC-SSIMULACRA2-PARITY-2026-06-03 sits under ## Recently closed in docs/state.md at origin/master (line 385), moved there by #1314 ("revert pseudo-Kahan recurrence and calibrate Arc A380"). There is no stale Open row left to move.

Rebasing would re-add a second row for the id, which scripts/ci/check-state-md-rows.sh (added by #1322) now rejects — and it also conflicts in scripts/ci/gpu_ulp_calibration.yaml against the calibrated tolerance #1314 landed. Closing rather than rebasing.