docs(vmaf): research digest + ADR-1122 for VMAF v1 model port

[Tualua]

Summary

Netflix shipped VMAF v1 in libvmaf v3.2.0 (June 2026). This PR adds a cited
research digest and ADR-1122 deciding how the fork adopts it: opt-in v1
generation, vmaf_v0.6.1 stays default. Docs-only (planning) — no code yet.

Key finding: v1 is still a libsvm ν-SVR fusion, but VIF is removed and CAMBI
(banding) + speed_chroma (chroma) become core features (ADM/motion → adm3/motion3).
Six built-in models (1080p@3H, phone@5H, 4K@1.5H, 4K@3H — last on [0, 110]) + HFR
variants; Netflix recommends 10-bit even for 8-bit SDR (aligns with ADR-1121 / PR #1081).

Type

• docs — documentation only

Checklist

• Commits follow Conventional Commits.
• ADR row lives in docs/adr/_index_fragments/1122-*.md + _order.txt (README regenerated).
• No Netflix golden assertions modified (docs-only).

Bug-status hygiene (ADR-0165)

• no state delta: planning ADR, not a bug open/close/rule-out.

Deep-dive deliverables (ADR-0108)

• Research digest — docs/research/2026-07-01-vmaf-v1-models.md.
• Decision matrix — ADR-1122 ## Alternatives considered (opt-in vs v1-default vs don't-port vs models-only).
• AGENTS.md invariant note — no rebase-sensitive invariants: docs-only planning ADR, no code touched.
• Reproducer / smoke-test command — no reproducer needed: docs-only.
• CHANGELOG fragment — no changelog needed: no user-visible surface changes (planning ADR only).
• Rebase note — no rebase impact: docs-only; the actual v1 port lands under this ADR later.

Known follow-ups

The v1 implementation (build in JSONs, fix SYCL cambi/speed_chroma, [0,110]
handling, CAMBI enc-params, v1 golden tests, upstream sync to v3.2.0) is scoped in
ADR-1122 and tracked as a follow-up port — this PR only lands the decision + research.

🤖 Generated with Claude Code

https://claude.ai/code/session_01YCP4JCoduujJgccDRGvNKv

[lusoris]

Apologies for the long silence on this — the repository had a CI deadlock that blocked every PR (including our own) from 2026-06-29 until today. Nothing to do with your work. Details in #1127 if you're curious: the Required Checks Aggregator treats a sibling job still queued at its 90-minute deadline as a failure, and a large dependency backlog kept the queue deeper than that, so the gate kept failing PRs that had nothing wrong with them.

Two things I've done here:

1. CI is now running on this PR for the first time. Fork PRs need a maintainer to approve workflow runs, and that never happened. Approved now.

2. Pushed one formatting-only commit (4256164) to get markdownlint green. No content or conclusions were changed — table delimiter rows to the repo's compact pipe style, a text language on the feature_names fence, and backticks around the viewing-condition identifiers. That last one is worth flagging: MD034 read the dot in 4K@1.5H as a TLD and called it a bare URL. Backticks are the right fix anyway since these are identifiers, not prose.

Heads-up if you ever reach for it: markdownlint-cli2 --fix is unsafe in this repo (ADR-0864). It rewrote 4K@1.5H to <4K@1.5H> and silently corrupted the sentence. I reverted its output and applied the fixes by hand.

On the ADR numbers. You claimed ADR-1121 (in #1081) and ADR-1122 here, and those claims are correct and stand — I'm not going to renumber your work. But our allocator (scripts/adr/next-free.sh) only queries git ls-remote --heads origin, so it cannot see ADR files on contributor forks. It handed me 1121 today for an unrelated change and I had to release it by hand. That's our bug, not yours; I'll get the allocator taught about open fork PRs.

Also: several of the docs/adr/README.md lint errors on this PR are pre-existing index corruption, not yours — #1126 fixes them at the source. Once that lands, rebasing here should clear those.

The v1 model research is genuinely good work — the feature-delta table and the [0,110] score-range implications are exactly the parts we'd have had to rediscover. Reviewing #1081 properly next.