docs(adr): record eight deliberate deviations from Netflix's source, each with its size and the upstream change that ends it (ADR-1479 to ADR-1486)

[lusoris]

Summary

Eight deliberate differences from Netflix's libvmaf get an ADR each. The upstream parity audit of 2026-10-02 found them without one of their own, or covered only by ADR-1033, which names neither upstream's behaviour nor the size of the difference. Documentation only: no code, no score and no test changes.

Each ADR gives upstream's behaviour with file and line at Netflix 9e48141b, the fork's lines, the reason as found in the fork's pull request, commit or code, the measured size from the audit, the upstream pull request or issue where one exists, and what ends the deviation.

ADR	Deviation	Fork change	Measured size	Ends with
ADR-1479	ciede 4:2:2 chroma flags	#1050	0.153 on 48 of 48 frames	Netflix/vmaf#1611 (another contributor)
ADR-1480	speed_temporal buffers at speed_prescale above 1	#1643	up to 195; upstream segfaults on 2 of 5 fixtures	Netflix/vmaf#1627, issue #1626
ADR-1481	a failing extractor fails the run	#871	status only: 78 probe runs where upstream is silent, 88 where it segfaults	no upstream PR propagates the error (#1620, #1637, #1642, #1629 fix single extractors)
ADR-1482	integer adm on frames of 17 to 32 px	#1473, #1507	scale 3 up to 0.23, adm2 0.104	Netflix/vmaf#1599, #1600
ADR-1483	odd-sized chroma planes round up	4f08d32b2	psnr_cb / psnr_cr 0.684 / 0.826 dB, ciede2000 0.198	no upstream PR (#1604 covers the tools only)
ADR-1484	float_ms_ssim magnitude before pow()	#641 (ADR-1033 item 2)	NaN upstream, 0.9895 here on the 10 px checkerboard	Netflix/vmaf#1665
ADR-1485	apsnr of a plane without error	#641 (ADR-1033 item 1)	114 dB upstream, 60 dB here	Netflix/vmaf#1666
ADR-1486	float_motion scale-1 stride	#641 (ADR-1033 item 9)	up to 0.47 on the Netflix pair, 25.1 on noise	Netflix/vmaf#1667

Checked and not given a new ADR, because one records the decision: D1 (ADR-1442), D2 (ADR-0125), D6 (ADR-1402), D7 (ADR-1413), D8 (ADR-1191, ADR-1325, ADR-1472), D10 (ADR-1301). Their measured sizes against upstream go into the allowlist of the parity guard (next PR).

Two upstream forms differ from the fork's without differing in value, and the ADRs say so: #1665 takes the magnitude of the structure term only, #1666 keeps the factor 2 in the ceiling. At the sync that brings them, the fork takes upstream's lines.

Also sets a 120s timeout on test_cuda_motion_tiny_frames in core/test/meson.build so the 106-case GPU test suite does not exceed Meson's default 30s timeout under host load.

Type

• docs — documentation only

Checklist

• Commits follow Conventional Commits.
• ADR numbers claimed with scripts/adr/next-free.sh --claim; index rows live in docs/adr/_index_fragments/ and _order.txt; docs/adr/README.md is regenerated.
• Unit tests / cross-backend diff: not applicable, no code path changed.

Bug-status hygiene (ADR-0165)

• docs/state.md — no state delta: the ADRs record behaviour that is already on master; no bug is opened, closed or ruled out.

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: the measurements are the upstream parity audit's, quoted in each ADR.
• Decision matrix — each ADR's ## Alternatives considered.
• AGENTS.md invariant note — docs/development/rebase-sensitive-invariants.md, entry "Deliberate deviations from Netflix's source, by ADR".
• Reproducer / smoke-test command — below.
• CHANGELOG fragment — changelog.d/changed/adr-deliberate-upstream-deviations.md.
• Rebase note — docs/rebase-notes.md, "Eight deliberate deviations from Netflix's source have their ADR": the fork lines a sync keeps, upstream's form, and the pull request that ends each.

Reproducer

make docs-fragments-check
# upstream's lines as cited, e.g. ADR-1479 and ADR-1481:
git -C <Netflix/vmaf checkout> show 9e48141b:libvmaf/src/feature/ciede.c | sed -n '71p;73p;89p;91p'
git -C <Netflix/vmaf checkout> show 9e48141b:libvmaf/src/thread_pool.c | sed -n '191,200p'

Known follow-ups

• The upstream parity guard (next PR, stacked on this one) lists each deviation in its allowlist with these ADR numbers and fails when one no longer matches a difference.