docs(upstream): record what this fork has open against Netflix/vmaf

[lusoris]

Summary

The fork has eight pull requests open against Netflix/vmaf and nothing in the tree said so. That matters at the next sync, because two of them fix bugs a fork PR also fixes, and one fixes it a different way.

docs/development/known-upstream-bugs.md gains two sections:

• The eight open contributions, each tied to the fork's own bug id: chore(deps): Update pre-commit hooks #1588 dictionary ownership, chore(deps): Update Python (patch) #1589 percentile pooling, chore(deps): Update dependency oneapi-src/level-zero to v1.34.0 #1590 model-collection allocation, chore(deps): Update Go deps (minor + patch) #1591 thread-pool errors, fix(release): stage the native vmaf CLI with RUNPATH $ORIGIN #1599 scale-3 DWT index -1, fix(helm): give the server Deployment its own component selector #1600 the zero-shift rounding, fix: keep backslashes in CLI option values; build svm.cpp with libc++ 23 #1601 the 16-bit vertical DWT overflow, fix(release): build the native Linux bundle on the Debian 13 release track #1602 the SIMD adm_cm centre tap. All were rebased onto upstream 86da14d0 and re-validated there on 2026-09-19. No CI has ever run on any of them: every workflow upstream sits at action_required.
• Six upstream defects found while validating and not reported, so the fork does not rediscover them: the checkasm check_adm_dwt2 source over-read, direct-read YUV breaking odd dimensions, frames below 17 px crashing integer ADM, a negative shift in get_best15_from32(), a zero-length variable-length array in the tool, and the SIMD adm_cm casting its accumulator to float before dividing.

The part worth acting on. Upstream #1601 fixes the 16-bit DWT overflow by starting the sum from the normalization offset rather than widening the accumulator to int64, which is what the fork's #1477 does. Upstream measured the int64 form at 3.5 to 6 % of throughput. The offset form adds no operation and measures within noise, so it is worth bringing back here, and a sync must not resolve that conflict by keeping both. Upstream #1602 makes SIMD follow scalar as the fork does, and says plainly that float_adm is closer to the unwrapped vector value, so the wrap is an artefact of the scalar reference; if upstream removes the int16 cast instead, the fork's #1474 needs revisiting.

Three docs/state.md rows now cite the upstream pull request carrying their fix. The Netflix#1109 row records that the reporter's clips return 404 and that the frame-alignment diagnosis lived only in the fork's own docs, so there is nothing to retract upstream.

One correction is recorded against the fork's own upstream work: #1589 as first submitted appended four values to the pooling enum while upstream's report writers index a four-entry name table, so every report's pooled_metrics got garbage keys. The fork has that guard; it was lost in the adaptation, and the first round validated with unit tests but never ran the tool. Corrected upstream on 2026-09-19.

Type

• docs — documentation only

Checklist

• Commits follow Conventional Commits (the commit-msg hook enforces this).
• make format && make lint is green locally.
• I did not modify any assertAlmostEqual(...) score in the Netflix golden Python tests.
• docs/state.md updated in this PR: three rows now cite the upstream pull request that carries their fix.

Deep-dive deliverables (ADR-0108)

• Research digest — no digest needed: trivial. The evidence is the engagement ledger this summarises, kept out of tree under .workingdir/evidence/.
• Decision matrix — no alternatives: only-one-way fix. Recording what exists upstream is not a choice between designs.
• AGENTS.md invariant note — carried in docs/rebase-notes.md, the cross-package invariant index the harness imports; it names the files whose conflicts this changes.
• Reproducer / smoke-test command — below.
• CHANGELOG fragment — changelog.d/changed/upstream-pr-links.md.
• Rebase note — docs/rebase-notes.md.

Reproducer

python3 -m mkdocs build --strict
bash scripts/ci/check-state-md-rows.sh

Every upstream pull request number in the new section resolves to an open pull request on Netflix/vmaf.

Known follow-ups

• Bring upstream fix: keep backslashes in CLI option values; build svm.cpp with libc++ 23 #1601's offset-first approach back to the fork's fix(adm): form the 16-bit vertical DWT sum in int64 #1477, replacing the int64 accumulator.
• The rows on the upstream-port stack (port: reconcile with upstream Netflix/vmaf September 2026 (ADM/VIF SIMD fixes) #1473, fix(adm): make the AVX2 and AVX-512 contrast masking wrap like scalar #1474, fix(gpu): make the integer ADM twins agree with the CPU on tiny frames and full-range content #1476, fix(adm): form the 16-bit vertical DWT sum in int64 #1477, fix(gpu): form the 16-bit scale-0 vertical DWT sum in 64 bits #1478) still need their own citations; that stack is blocked on three golden lines a maintainer applies by hand, so those edits belong on its branches rather than here.