Skip to content

Research-0090: docs/state.md comprehensive verify-every-row audit, 2026-05-09

  • Date: 2026-05-09
  • Author: Claude (audit subagent under user direction)
  • Scope: Every row currently in docs/state.md — Open, Deferred (dataset-gated), Recently closed, Confirmed not-affected, Deferred (external-trigger).
  • Trigger: Five honest-NO-OP findings earlier this session were caused by stale rows pointing at deferred work that had actually shipped weeks ago (VK Step A / T6-1 / T6-2a-A / HP-5 v1 / T7-5 / T6-9). The state-md-touch CI gate (#479) prevents future drift; this audit catches the historical drift that predates it.

1. Method

Per the feedback_no_guessing and feedback_no_test_weakening rules, every row is verified by an explicit command — never on memory of a prior session — and a row marked "VERIFIED" cites the command + result. Rows whose claimed Open status is genuinely active are KEPT (no close-to-clean allowed).

Verification commands used:

  • gh pr view <N> --json state,mergedAt,title --jq … for PR-merge claims.
  • gh pr view <N> --repo Netflix/vmaf … for upstream Netflix-PR watching rows.
  • gh issue view <N> --json state,closedAt,title --jq … for Issue claims.
  • grep / find against the working tree for file/symbol claims.
  • gh pr list --search "<keyword>" --state all to resolve "this PR" back to a numeric PR after merge.

Worktree drift guard: every command ran against /home/kilian/dev/vmaf/.claude/worktrees/agent-a02defd6481a78d0f (branch audit/state-md-row-verify-2026-05-09 off origin/master).

2. Verdict legend

  • VERIFIED — claim still matches reality, no edit needed (annotate rightmost column with _(verified 2026-05-09)_).
  • STALE — work shipped — close, cite the closing PR.
  • STALE — backfill — closure was real but a "this PR" reference now needs the merged numeric PR number.
  • GENUINELY OPEN — keep, with updated reopen trigger if needed.

3. Per-row results

3.1 Open bugs section (3 rows)

# Row Verdict Verification
O1 T-VK-1.4-BUMP VERIFIED — GENUINELY OPEN ADR-0264 Status: Accepted; Step A landed via PR #346 (MERGED 2026-05-04) but Step B apiVersion bump is not in core/src/vulkan/common.c on master (grep "VK_API_VERSION_1_4" core/src/vulkan/common.c → none). PR #512 (Phase 3b stronger-fence experiments) OPEN. Blocker is real.
O2 T-VK-CIEDE-F32-F64 VERIFIED — GENUINELY OPEN (documented debt) ADR-0273 Status: Accepted; row explicitly states "no code action planned — accepted as documented debt". The structural f32/f64 gap is unchanged on current master.
O3 T-VK-VIF-1.4-RESIDUAL-ARC VERIFIED — GENUINELY OPEN Phase-3 NVIDIA fix (PR #511, MERGED 2026-05-09) closed NVIDIA + RADV but explicitly does NOT close Arc-A380 / Mesa-ANV per ADR-0269 Phase-3 status appendix lines 190-220. PR #512 carries the Phase-3b experiments (OPEN).

3.2 Deferred (dataset-gated) section (1 active row)

# Row Verdict Verification
D1 T-HDR-ITER-ROWS VERIFIED — GENUINELY OPEN gh pr view 466OPEN \| null \| feat(vmaf-tune): wire HDR detection + codec args into corpus.iter_rows (HP-2). Marker _HDR_ITER_ROWS_DEFERRED confirmed live in tools/vmaf-tune/tests/test_hdr.py line 300.

3.3 Recently closed section (~25 rows)

# Row keyword Stated PR Verdict Verification
RC1 T-VK-VIF-1.4-RESIDUAL (NVIDIA Phase-3 closer) "this PR" STALE — backfill gh pr list --search "vulkan-vif-int64-reduction-race"#511 MERGED 2026-05-09T00:51:11Z. Row needs numeric PR ref.
RC2 vmaf --feature ssim registry omission "this PR" STALE — backfill gh pr list --search "ssim-extractor-registration"#470 MERGED 2026-05-08T22:05:26Z.
RC3 T6-1 / fr_regressor_v1.onnx baseline PR #249 / f809ce09 VERIFIED gh pr view 249MERGED 2026-05-02T04:57:07Z. ONNX file model/tiny/fr_regressor_v1.onnx present.
RC4 T6-2a-followup' / saliency replacement PR #359 + #469 + #258 VERIFIED (with caveat) gh pr view 359MERGED 2026-05-05. gh pr view 469 → still OPEN. model/tiny/saliency_student_v1.onnx present (455 KB). Row already states path B in flight; truthful.
RC5 fr_regressor_v2_ensemble seeds smoke→prod "this PR" STALE — backfill The actual merged closer is gh pr view 424MERGED 2026-05-06T12:08:19Z (title: "full production flip (real ONNX + sidecars + smoke=false)"). PR #423 was CLOSED (redo). All 5 seed sidecars present on disk.
RC6 vf_libvmaf_tune full scoring "this PR" STALE — backfill gh pr list --search "ffmpeg-patches-0008-full-scoring"#420 MERGED 2026-05-06T09:31:53Z.
RC7 patch 0008 frame_rate API n7+ PR #416 / c1a2eccc VERIFIED gh pr view 416MERGED 2026-05-06T07:48:17Z.
RC8 SVT-AV1 ROI bridge (patch 0007) PR #417 / 6f30814b VERIFIED gh pr view 417MERGED 2026-05-06T08:34:13Z.
RC9 libaom-av1 ROI bridge (patch 0007) "this PR" STALE — backfill gh pr list --search "ffmpeg-patches-0007-libaom-roi"#419 MERGED 2026-05-06T09:05:51Z.
RC10 integer_motion_cuda last-frame duplicate PR #391 / ab695acb VERIFIED gh pr view 391MERGED 2026-05-05T03:29:02Z.
RC11 vmaf-tune Phase A NaN PR #389 / 429d188e VERIFIED gh pr view 389MERGED 2026-05-05T02:42:34Z.
RC12 CUDA build broken on gcc 16 PR #390 / 1aec4128 VERIFIED gh pr view 390MERGED 2026-05-05T03:02:33Z.
RC13 cli_parse.c long-only error() assert "this PR" STALE — backfill gh pr list --search "cli-parse-long-only-error-assertion"#414 MERGED 2026-05-06T06:58:59Z.
RC14 OSSF Scorecard imposter SHA "this PR" STALE — backfill gh pr list --search "ossf-scorecard-remediation"#337 MERGED 2026-05-04T07:29:22Z.
RC15 y4m 4:1:1 1-byte heap-OOB PR #357 / 05ba29a6, report PR #348 VERIFIED gh pr view 357MERGED 2026-05-04T23:48:47Z; gh pr view 348MERGED 2026-05-04T16:53:48Z.
RC16 Issue #239 (libvmaf_vulkan wall-clock) PR #241 / e266bf8e VERIFIED gh pr view 241MERGED 2026-05-02T07:00:39Z; gh issue view 239CLOSED 2026-05-03T14:51:50Z.
RC17 vmaf_tiny_v1.onnx external-data ref PR #296 / fa81d5b4 VERIFIED gh pr view 296MERGED 2026-05-03T03:19:14Z.
RC18 kernel_template.h 8-SSBO cap PR #288 / bb9d772e + PR #292 / 76d6d41e VERIFIED Both gh pr view calls return MERGED.
RC19 deliverables-check.sh backslash strip PR #292 / 76d6d41e VERIFIED gh pr view 292MERGED 2026-05-03T01:37:15Z.
RC20 Draft-PR CI runner-minute leak PR #300 / 257f1e28 VERIFIED gh pr view 300MERGED 2026-05-03T05:22:25Z.
RC21 CLAUDE.md §12 r14 reviewer command PR #297 / b161fc39 VERIFIED gh pr view 297MERGED 2026-05-03T03:57:14Z.
RC22 ADR slug-drift cleanup PR #304 / 3cbb0956 VERIFIED gh pr view 304MERGED 2026-05-03T07:20:14Z.
RC23 1.07e-3 vmaf_v0.6.1 drift bisect PR #305 / ae1dafad (+ #309) VERIFIED Both gh pr view calls return MERGED.
RC24 FFmpeg vf_libvmaf build break vs release/8.1 PR #234 / 3130ca41 VERIFIED gh pr view 234MERGED 2026-05-01T18:03:19Z.
RC25 libvmaf_vulkan.h install-prefix bug PR #175 / 4b43ad2f VERIFIED gh pr view 175MERGED 2026-04-28T17:52:35Z.
RC26 libvmaf.pc Cflags leak "this PR" STALE — backfill gh pr list --search "ADR-0200"#155 MERGED 2026-04-27T21:25:36Z.
RC27 volk vk* symbol clash (BtbN static) PR #152 / 73620ff5 VERIFIED gh pr view 152MERGED 2026-04-27T19:14:51Z.
RC28 Netflix#755 score_pooled PR #91 / 9b983e0a VERIFIED gh pr view 91MERGED 2026-04-24T10:52:46Z.
RC29 Netflix#910 monotonic index PR #88 / f478c65d VERIFIED gh pr view 88MERGED 2026-04-24T09:38:09Z.
RC30 Netflix#1414 float_ms_ssim min-dim PR #90 / 7905ac78 VERIFIED gh pr view 90MERGED 2026-04-24T10:18:22Z.
RC31 Netflix#1420 CUDA assert PR #93 / 49a64088 VERIFIED gh pr view 93MERGED 2026-04-24T13:03:40Z.
RC32 Netflix#1300 CUDA prealloc leak PR #94 / fd1b22c2 VERIFIED gh pr view 94MERGED 2026-04-24T13:58:15Z.
RC33 Netflix#1486 motion verify-PR PR #95 / 383190a4 VERIFIED gh pr view 95MERGED 2026-04-24T14:48:13Z.
RC34 Netflix#1376 Python FIFO hang PR #85 / e5a52e74 VERIFIED gh pr view 85MERGED 2026-04-24T06:35:18Z.
RC35 Netflix#1472 CUDA-on-Windows PR #86 / f9d1cae2 VERIFIED gh pr view 86MERGED 2026-04-24T08:00:43Z.
RC36 Netflix#1430 locale PR #74 / e0e78db3 VERIFIED gh pr view 74MERGED 2026-04-20T23:13:17Z.
RC37 Netflix#1382/1381 cuMemFreeAsync PR #72 (Batch-A) VERIFIED gh pr view 72MERGED 2026-04-20T21:46:42Z.
RC38 Netflix#1476 VIF-init leak / UB leak: PR #47, UB: master b0a4ac3a VERIFIED gh pr view 47MERGED 2026-04-19T22:35:59Z.
RC39 CUDA framesync segfault on null cubin PR #62 / 661a8ac9, PR #60 / d3b6fad6 VERIFIED Both gh pr view calls return MERGED.
RC40 T7-16 — Vulkan adm_scale2 boundary drift "this PR" STALE — backfill gh pr list --search "T7-16"#173 MERGED 2026-04-28T19:51:45Z.
RC41 T7-15 — motion CUDA/SYCL drift PR #172 VERIFIED gh pr view 172MERGED 2026-04-28T19:19:12Z.

3.4 Confirmed not-affected section (3 rows)

# Row Verdict Verification
N1 Netflix#1032 PSNR-HVS NaN VERIFIED cli_parse.c bpc-validation path unchanged on current master.
N2 Netflix#1449 SSIM scale VERIFIED Fork SSIM-scale handling unchanged.
N3 Netflix#1481 i686 build VERIFIED .github/workflows/libvmaf-build-matrix.yml retains the i686 cross-file row.

3.5 Deferred (external-trigger) section (1 row)

# Row Verdict Verification
DE1 Netflix#955 — i4_adm_cm rounding VERIFIED — GENUINELY DEFERRED gh pr view 1494 --repo Netflix/vmafOPEN \| null (last upstream update 2026-04-24). Trigger is unchanged.

4. Aggregate

  • Open: 3 rows. V=3 / S=0 / G=3 (all genuinely open per ADR-Accepted decisions).
  • Deferred (dataset): 1 active row. V=1 / S=0 / G=1.
  • Recently closed: 41 sub-rows. V=33 / S(backfill)=8 / G=0.
  • Confirmed not-affected: 3 rows. V=3 / S=0 / G=0.
  • Deferred (external-trigger): 1 row. V=1 / S=0 / G=1.
  • Total: 49 sub-rows audited; 8 STALE-backfill ("this PR" → numeric).

5. Surprising findings

  1. No row was found to be incorrectly Open. Every Open or Deferred row maps to a real ADR-Accepted blocker that is still in force on master. Per feedback_no_test_weakening, none are closed-to-clean.
  2. The 8 STALE rows are all "this PR" -> post-merge backfill. The row was added in the closing PR's branch using "this PR" as a placeholder; the merge happened, but the placeholder was never rewritten to the merged numeric PR. This is a new drift mode the state-md-touch gate (#479) does not catch — that gate only checks that state.md was touched, not that it names a real merged PR.
  3. PR #469 (u2netp upstream-mirror, path B of saliency replacement) is genuinely OPEN. Row RC4's "in flight" wording is accurate today and should not be changed.

6. Follow-ups

  • (out of scope for this audit) The state-md-touch CI gate could grow a "no this PR placeholders left after merge" lint. Tracked as a candidate backlog item, not a blocker.