Skip to content

Upstream-backlog re-audit — 2026-04-29 (T7-4)

Quarterly upstream-backlog re-audit per .workingdir2/BACKLOG.md row T7-4. Companion to the 2026-04-18 snapshot in .workingdir2/analysis/upstream-backlog-audit.md (local-only). Doubles as the deep-dive research digest for this PR per ADR-0108. Next re-audit due 2026-07-29.

Summary

  • 12 upstream commits walked since the previous fork-side upstream port boundary (798409e3 and below), git fetch upstream master snapshot taken 2026-04-29.
  • 8 already on fork — content cherry-picked, ported, or consciously diverged-from with an Accepted ADR.
  • 4 flagged for fork action — surfaced as recommended new T-rows below.
  • 0 new bug rows for docs/state.md; no upstream commit ruled in or out a bug that affects the fork's user-visible surface.
  • No decision-matrix needed: scheduled audit, not a non-trivial decision per CLAUDE.md §12 rule 8.

Method

  1. git fetch upstream master (Netflix/vmaf).
  2. Walk git log upstream/master --oneline 798409e3^..upstream/master — 798409e3 is the lower bound because PR #181 (6eab09c0) is the most recent direct chore(upstream): port commit on fork master and ports 798409e3 together with 314db130.
  3. Classify each commit into one of:
  4. already-on-fork — landed via cherry-pick, port, or covered by an Accepted ADR;
  5. port — bug-fix / build-fix worth porting (new T-row);
  6. defer — recognised, out-of-scope right now;
  7. diverged — fork has consciously diverged with an ADR.
  8. Cross-check fork files (core/src/feature/, core/include/, .github/workflows/) for the per-commit marker symbols.

Per-commit triage

upstream-sha subject fork-status recommended-action T-row
c70debb1 libvmaf/test: port new adm/vif/speed tests port port the adm + vif test deltas; speed tests pend on T-NEW-1 T-NEW-2
314db130 libvmaf/feature: remove empty translation unit all.c already-on-fork none — landed via PR #181 (6eab09c0) —
9dac0a59 libvmaf/feature: update alias map for cambi/speed partial cambi half is on fork (PR #160 79288e8d); speed half pends on T-NEW-1 T-NEW-1
8c645ce3 feature/vif: port several feature extractor options diverged none — Research-0024 + ADR-0142/0143/0144 record deliberate divergence; fork keeps closed-enum kernelscale —
8a289703 adm: add fallback for extract_epi64 for 32-bit port small portability fix; aligns with fork's i686 build job (ADR-0151) T-NEW-3
1b6c3886 x86/cpu: remove limit of avx+ on 32-bit port partner to 8a289703; verify against fork's cpu-feature dispatch T-NEW-3
f6d6dde1 gh: add 32-bit cross build to CI already-on-fork none — fork has its own i686 build-only matrix row (PR #87 978f9583, ADR-0151) —
b949cebf feature/motion: port several feature extractor options diverged none — Research-0024 covers; fork kept its own divergence per ADR-0145 —
4dcc2f7c feature/float_adm: port several feature extractor options diverged none — Research-0024 covers; fork's float_adm divergence accepted —
2c9bb74e feature/cambi: add effective_eotf already-on-fork none — landed via PR #160 (79288e8d) —
eb88c009 build(deps): bump actions/upload-artifact 5→7 already-on-fork none — fork manages action versions via Dependabot —
798409e3 Fix null deref crash on prev_ref update in pure CUDA pipelines already-on-fork none — landed via PR #181 (6eab09c0) —

Status legend

  • already-on-fork — content is on master (cherry-pick, port, or fork-equivalent infrastructure).
  • port — recommended action: open a port-upstream-commit PR.
  • diverged — fork has consciously chosen a different design; Accepted ADR documents the rationale; no action required.

The following follow-up entries are recommended for the next .workingdir2/BACKLOG.md refresh:

  • T-NEW-1 — port feature/speed (d3647c73 speed_chroma + speed_temporal, 9dac0a59 alias-map for speed). The fork does not currently have a speed* extractor in core/src/feature/; upstream PR #1433 also adds a related speed-extractor surface. Investigate whether to port the upstream C extractor wholesale, or to absorb it into the fork's tiny-AI speed metric. Pairs with audit row 1 / 3.
  • T-NEW-2 — port adm + vif test deltas from c70debb1. The speed half of c70debb1 blocks on T-NEW-1; the adm + vif halves should drop in cleanly on top of fork's existing port chain (Research-0024). Goal: keep core/test/test_*.c test-data parity with upstream wherever the fork's algorithmic content matches.
  • T-NEW-3 — port 32-bit ADM/cpu fallbacks (8a289703 + 1b6c3886). Pairs with the existing i686 build-only CI matrix row (PR #87, ADR-0151). Currently the fork's i686 lane is build-only; these two commits add the runtime-correctness pieces (32-bit extract_epi64 fallback, lifting the AVX-on-32-bit guard) that would let the i686 lane be run as well.
  • T-NEW-4 — schedule next quarterly re-audit for 2026-07-29. Update T7-4's "next" date in BACKLOG.md.

What this audit did not cover

  • Open Netflix issues / PRs. The 2026-04-18 dossier in .workingdir2/analysis/upstream-backlog-audit.md triages every open upstream issue and PR (66 + 35 at that snapshot). This re-audit is scoped to landed upstream commits since the last fork-side port boundary, not the open queue. The next snapshot of the open queue is recommended alongside T-NEW-4 (re-run with gh issue list --repo Netflix/vmaf --state open --limit 300 per the dossier's "Re-run instructions" section).
  • Pre-798409e3 upstream history. The 12-commit window matches the fork's most recent direct port and the dossier's open-PR table. Older upstream commits (e.g. f3a628b4, 18e8f1c5, a44e5e61, dae6c1a0, f740276a) are already on fork via the Research-0024 chain or earlier ADR-tracked ports and were verified via spot-check (vif_sigma_nsq, motion_v2, VMAF_FEATURE_EXTRACTOR_PREV_REF all present in core/src/feature/).

Reproducer

# 1. Refresh the upstream tracking ref.
git fetch upstream master

# 2. Walk the window covered by this audit.
git log upstream/master --oneline 798409e3^..upstream/master

# 3. For each commit, cross-check fork master:
git log master --oneline --grep='<sha-or-subject-keyword>'

# 4. For algorithmic divergences, cross-check the ADR / research digest tree:
ls docs/research/ docs/adr/

Re-audit cadence

  • Previous: 2026-04-18 (open-queue snapshot).
  • This pass: 2026-04-29 (landed-commit re-audit).
  • Next: 2026-07-29 (per CLAUDE.md §12 rule 8 + T7-4 quarterly cadence).